The Level Index · Open Benchmark (LevelBench-100 v3.0)

Can AI Close Real SMB Books?
38 AI Systems Tested on Messy Takeovers

We tested 38 AI models, CLI tools, and accounting workflows across 100 real SMB takeover, month-end close, and CFO checks plus 14 software portals. Raw AI models score 31.0% to 71.0% alone, while AI guided by Level Playbooks (2,200+ past experiences) and controller review reaches 100.0% accuracy with zero bad QuickBooks entries.

Compare All 38 Systems ↓8 Accounting & CFO Jobs Scorecard Download Open Dataset (JSON)

SMB Takeover Worlds

HVAC · Ecommerce · Healthcare MSO · PE Rollup

AI Setups Tested

27 raw models · 4 CLI tools · 7 playbook setups

100 + 84

Real Takeover Checks

100 accounting checks + 84 browser portal trials

71.0%

Best Raw AI Alone

Gemini 4 Argon (Medium Thinking) (median 55.0%)

100.0%

With Level Playbooks

2,200+ past experiences · 0.0% bad-entry rate

The 30-Second Operator Takeaway

Why raw AI scores 31% to 71% on real SMB books, and why playbooks from 2,200+ experiences matter

1. Real Client Books Are Messy

All 27 raw AI models make bad QuickBooks entries (14.3% to 57.1% risk) when a prior bookkeeper posted manual journal entries, a $14,500 check lacks an image, or a clerk guesses in Slack on tied bills.

2. 2,200+ Experiences Reach 100%

When AI is guided by Level's playbooks from 2,200+ past SMB cleanups and operator review, accuracy jumps to 100.0% with 0.0% bad entries across all 100 checks.

3. Frontier AI + Fast Worker AI

Gemini 4 Argon (71.0% raw) and Claude Opus 5.5 (69.0% raw) help diagnose messy takeovers; fast worker AI like GPT-5.6 Terra ($0.025/close) runs daily inside our playbooks at 100.0%.

Visual Comparison · 100 Real-World SMB Takeover & CFO Checks

Accuracy Across AI Setups (Higher Bar = Fewer Accounting Mistakes)

Evaluated across 4 anonymized SMB QuickBooks takeover packages & 12 close workflows

AI + Level Playbooks (2,200+ Past Experiences) + Operator Review

Accumulated playbooks from 2,200+ SMB cleanups check prior-bookkeeper JEs, match AP bills & hold missing check images ($0.008 to $0.090/close; $0.015 to $0.025 on Flash 3.8 / Terra 5.6)

100.0% Suite Accuracy · 0.0% Bad-Entry Rate

Multi-File Agentic CLI Tools · Gemini CLI, Antigravity & Codex CLI

Can cross-reference multiple exports in a workspace (76.0% to 84.0% raw; 83.0% to 89.0% with SOP prompts), but still need playbook rules on prior-bookkeeper JEs & tied bills

76.0% to 84.0% Raw · 10.7% to 17.9% Bad-Entry Risk

Gemini 4 Argon · Medium Thinking (Best Unassisted Raw AI Model)

Top raw foundation model on messy SMB takeover books (71.0% raw · 82.0% with SOP prompt · 26.09s · $0.040/close)

71.0% Raw Accuracy · 14.3% Bad-Entry Risk

Claude Opus 5.5, Claude Fable 5.1 & GPT-6.1 Sol (Frontier Raw Models)

Opus 5.5 (69.0% raw), Fable 5.1 (68.0% raw), and GPT-6.1 Sol (66.0% to 67.0% raw) do clean math well, but get tripped up by unlinked prior-bookkeeper JEs, Slack opinions & QBO debt grouping

66.0% to 69.0% Raw · 17.9% to 21.4% Bad-Entry Risk

Fast Worker Models Alone · Gemini 3.8 Flash & GPT-5.6 Terra Low

Score 57.0% to 58.0% raw on messy takeover files when run alone, but jump to 100.0% (+42.0 to +43.0 pp lift) when guided by Level Playbooks at $0.015 to $0.025/close

57.0% to 58.0% Raw · 100.0% With Level Playbooks

Open-Weight Models Alone · DeepSeek V3, Kimi K2.5, Qwen 3 & Qwen 2.5

Score 31.0% to 46.0% raw on messy SMB books (guessing paper checks & accepting bogus M&A add-backs); DeepSeek V3 reaches 100.0% (+54.0 pp lift) inside Level Playbooks

31.0% to 46.0% Raw · 46.4% to 57.1% Bad-Entry Risk

Option 1 · Manual Bookkeeper Takeover

15 to 25 hrs / mo

Chasing missing check stubs, unlinked prior-bookkeeper journal entries, and CRM/subledger tie-outs by hand in spreadsheets.

Option 2 · Raw AI Model or Chatbot Alone

31.0% to 71.0% accuracy

Fast, but carries a 14.3% to 57.1% bad-entry risk (double-books AP bills, guesses check categories, accepts bogus M&A add-backs).

Option 3 · Level Playbooks + Controller Team

100.0% · 0.0% Bad Entries

Playbooks from 2,200+ past experiences tie out subledgers in 0.04s ($0.015 to $0.040/close) with human controller sign-off in a 5-day monthly close.

Accuracy, Stage by Stage · How 2,200+ Experiences Compound

From Raw Chatbot (57% Median) to Level Playbooks (100.0%)

Kept in Production PlaybookPrompt-Only Attempt (Drifts on Takeover Traps)

Unlike prompt tuning learned and scored on the same 10 tasks, Level Playbooks are distilled from 2,242 real SMB and contractor cleanups and evaluated out-of-sample across 100 takeover checks.

Where Playbooks Helped · Takeover Workflow Lift

Raw AI Alone vs. With Level Playbooks (2,200+ Experiences)

No Playbooks (Raw AI)With Level Playbooks

M&A QoE bogus broker add-backs (WF11)

38% → 100%

Prior-bookkeeper JE & missing check stubs (WF01)

50% → 100%

13-week cash bridge & QBO debt grouping (WF09)

50% → 100%

Contractor WIP & Mon-Fri sub cutoffs (WF06)

56% → 100%

Supply-house AP bills & Slack guess (WF02)

63% → 100%

Prior-bookkeeper forced bank plug (WF04)

63% → 100%

MSO-PC intercompany & owner sweep (WF07)

63% → 100%

See model-by-model findings & answers to 4 common QuickBooks takeover questionsDetailed head-to-head results for Gemini 4 Argon, Claude Opus 5.5, Claude Fable 5.1, GPT-6.1 Sol, GPT-5.6 Terra, and DeepSeek V3.

On messy SMB client takeover exports, unassisted raw AI models score 31.0% to 71.0% (median 55.0%). Across the 28-check safety and subledger pool, all 27 raw models exhibit a 14.3% to 57.1% bad book-entry rate when operating without takeover playbooks: reclassifying Uncategorized Expense rows that a prior bookkeeper already neutralized with a manual Journal Entry, following a junior clerk's Slack guess on tied vendor bills, or guessing categories on paper checks missing check stub images. When guided by Level Playbooks (built from 2,200+ past SMB experiences) and operator review, all 7 evaluated setups achieve 100.0% measured accuracy and a 0.0% bad-entry rate.

1. Gemini 4 Argon (Medium Thinking) leads all 27 unassisted foundation models at 71.0% raw takeover accuracy: Evaluated on Vertex AI, Gemini 4 Argon (Medium) passes 71 of 100 checks unassisted (82.0% with accounting SOP prompts, 14.3% bad book-entry rate) in 26.09 seconds ($0.04 per close), outperforming Argon High (65.0% in 194.42s) because bounded reasoning avoids over-complicating subcontractor cutoff proration.
2. Claude Opus 5.5 (69.0% raw) and Claude Fable 5.1 (68.0% raw) lead on holding missing check images, while multi-file CLI harnesses reach 76.0% to 84.0%: Claude Opus 5.5 (69.0% raw, 80.0% with SOP prompts) and Claude Fable 5.1 (68.0% raw, 79.0% with SOP prompts) hold missing check stubs reliably, while multi-file CLI agents (Gemini CLI at 84.0% raw / 89.0% SOP-guided, Antigravity CLI at 83.0% / 88.0%, Codex CLI · Sol 6.1 at 81.0% / 87.0%, and Codex CLI · Terra 5.6 at 76.0% / 83.0%) cross-reference multiple exports in a workspace to resolve QBO debt reclassifications and M&A add-back schedules.
3. GPT-6.1 Sol scores 66.0% to 67.0% raw (77.0% to 78.0% with SOP prompts) and 100.0% when paired with Level Playbooks: Unassisted GPT-6.1 Sol High (67.0%) and Sol Medium (66.0%) handle multi-entity eliminations and FIFO landed-cost math cleanly, and jump by +33.0 to +34.0 percentage points to 100.0% accuracy inside Level Playbooks.
4. Fast worker models (GPT-5.6 Terra Low at 57.0% raw, Gemini 3.8 Flash at 58.0% raw, Claude Haiku 5.5 at 55.0% raw) jump by +42 to +45 percentage points inside Level Playbooks: At $0.015 to $0.025 per close, fast worker models struggle on raw takeover traps alone, but reach 100.0% accuracy and 0.0% bad book entries when governed by Level's playbooks from 2,200+ past SMB cleanups.
5. Open-weight models (DeepSeek V3 at 46.0%, Kimi K2.5 at 41.0%, Qwen 3 80B at 35.0%, Qwen 2.5 72B at 31.0%) cannot be run unassisted on client books: Without playbooks, open-weight models guess categories on paper checks without images, flip working-capital cash signs, and accept bogus seller EBITDA add-backs (46.4% to 57.1% bad book-entry rate). Inside Level Playbooks, DeepSeek V3 reaches 100.0% accuracy (+54.0 percentage-point lift) at $0.008 per close.

“Why does my QuickBooks profit margin not match my actual bank cash?”

Tested in WF06 & WF09: Cash-basis QBO reports miss completed-but-unbilled field jobs ($28,500 WIP), count upfront customer deposits for future-month work as current income ($19,000), and sweep debt reclassifications into operating cash flow.

“How do I stop QuickBooks bank rules from double-counting paid vendor bills or prior-bookkeeper adjustments?”

Tested in WF01 & WF02: Prior bookkeepers often leave feed rows in Uncategorized Expense and post manual Journal Entries to fix them, or pay open AP bills by wire without linking the bill. Raw AI reclassifies and expenses those outflows a second time.

“Why don't my field software numbers (ServiceTitan, BuildOps, Jobber) match QuickBooks or Sage Intacct?”

Tested in WF01, WF04 & WF06: Field CRMs export unsigned optional estimates alongside completed jobs, while QuickBooks tracks bank settlements, two Intuit merchant fee outflows ($52.40 and $182.60 = $235.00), uncleared check float ($1,300.00), and Monday-Friday subcontractor cutoffs.

“When do I transition from a part-time bookkeeper to Level's accounting & fractional CFO team?”

Tested in WF07, WF09, WF10 & WF11: Once a business passes $2M to $15M+ in revenue, multi-entity intercompany sweeps, 13-week bank covenant forecasting ($100,000.00 floor), true CAC stripping out retention events, and M&A Quality of Earnings require playbooks built from thousands of closes.

Scorecard · 8 Real Accounting & CFO Jobs

What AI Does Well vs. Where It Breaks on Messy SMB Books

Most AI tests only ask whether a chatbot can categorize a clean coffee receipt. Here is how the best raw AI model alone, fast worker AI alone (GPT-5.6 Terra), open-weight AI, and AI paired with Level Playbooks (2,200+ past experiences) perform across all 8 jobs required to clean up and close a company's books:

Accounting / CFO Job	Best Raw AI	Terra 5.6 Raw	Open-Weight AI	With Level Playbooks
1. Expense Categorization & Prior-Bookkeeper CleanupCleaning up a messy September bank and card feed during a client takeover: catching a prior bookkeeper's manual Journal Entry (JE-0899) that already reclassed an Uncategorized Expense row, isolating two separate Intuit merchant fee outflows ($52.40 and $182.60 = $235.00), excluding August cutoff rows and voided checks, and holding $20,200.00 of paper checks missing check stub images.	80.0%	60.0%	40.0%	100.0%
2. Supply-House AP, Wholesale Remittance & Bank ReconciliationMatching $56,700.00 in supply-house bank outflows ($56,500.00 of AP bill principal plus a $200.00 wire fee) to open vendor bills while keeping walk-in counter debit-card purchases in COGS, applying wholesale ACH remittances net of promo/early-pay discounts while isolating a Shopify wholesale order paid by paper check in Undeposited Funds, and separating duplicate bank-feed deposits ($6,400.00) from uncleared check float ($1,300.00).	75.0%	66.7%	29.2% to 45.8%	100.0%
3. Cash-to-Accrual Cutoffs, Prepaids & Field WIP SchedulesAmortizing annual general liability insurance ($18,000) and fleet telematics ($9,600) while leaving month-to-month SaaS in operating expense, accruing completed-but-unbilled field jobs ($28,500.00 WIP) while ignoring unsigned CRM estimate options, deferring upfront customer deposits for November work ($19,000.00), and prorating cross-month subcontractor bills using Monday-Friday business days (4/10 working days = $3,600.00).	64.7%	35.3%	29.4%	100.0%
4. Multi-Entity MSO Eliminations & Landed-Cost InventoryEliminating $20,700.00 of intercompany management fees and shared payroll across MSO and Clinical PC books, restoring $7,600.00 in Owner Sweep liabilities without double-counting $2,800.00 direct transfers, capitalizing $4,800.00 of inbound ocean freight/duty into FIFO inventory ($40 to $44/unit) while keeping $1,800.00 of outbound customer parcel shipping in fulfillment expense, and isolating 30 units ($1,320.00) of warehouse shrinkage.	82.4%	64.7%	23.5% to 47.1%	100.0%
5. 13-Week Cash Flow Modeling & Debt Covenant Stress TestingBuilding an indirect working-capital cash bridge (-$42k AR, -$18k Inventory, -$6k Prepaid, +$31k AP, +$52k Deferred Revenue = +$17,000.00 NWC cash impact) when QuickBooks groups Current Portion of Term Debt (+$15k) under Other Current Liability, and stress-testing a $135,000.00 commercial AR collection delay against a $100,000.00 bank line-of-credit covenant floor (-$6,200.00 Week 6 shortfall).	50.0%	50.0%	12.5% to 50.0%	100.0%
6. Unit Economics (CAC/LTV) & M&A Quality of Earnings (QoE)Computing fully loaded CAC ($120.00 across 400 new customers) after removing an existing-client VIP retention dinner ($4,500) and a CEO personal ski trip ($3,200) miscoded into Marketing, deducting Stripe fees ($3.78/order) to reach $65.00 contribution margin and 5.42x LTV:CAC, and auditing a seller's M&A EBITDA bridge by rejecting bogus add-backs for ongoing QuickBooks/CRM software ($14,000) and rainy-day technician drive time ($22,000).	56.2%	43.8%	37.5% to 43.8%	100.0%
7. Formula-Driven Excel Workbooks (.xlsx) & Executive Board Decks (.pptx)Auditing a companion monthly financial Excel workbook and executive board deck: catching hardcoded prior-month total cell E14 ('231400' instead of '=SUM(E5:E13)'), blocking Slide 2 for citing an unpinned $84,200 figure instead of ledger-pinned $54,925.00, checking 16:9 slide visual bounds, and enforcing plain operator tone.	87.5%	87.5%	50.0% to 87.5%	100.0%
8. Browser-Use AI: Accounting, Payroll & Field-Software Portal ExtractionExtracting complete, reconciled ledgers, AP bills, payroll journals, and job-cost tables across 84 portal-gate trials (14 production portals x 6 reliability gates: QuickBooks Online, ServiceTitan, BuildOps, Housecall Pro, LMN, Roofr, MarketSharp, Rippling, Gusto, Ramp, Brex, Settle, Finaloop, and Shopify) without session lockouts or silent row truncation.	71.4%	64.3%	21.4% to 41.7% (18/84 to 35/84)	100.0%

See where raw AI breaks on each of the 8 jobs (and how Level Playbooks from 2,200+ experiences fix it)Detailed failure modes for prior-bookkeeper manual JEs, missing check images, tied AP bills, WIP cutoffs, FIFO landed cost, 13-week cash flow, M&A QoE add-backs, and browser portals.

1. Expense Categorization & Prior-Bookkeeper Cleanup

WF01 (10 binary checks)

What Raw AI Does Well: Frontier reasoning models accurately classify standard SaaS, fuel, and supply-house card charges and separate merchant-processing fee outflows from customer settlements.

Where Raw AI Breaks: When a prior bookkeeper left a row in Uncategorized Expense and neutralized it with a manual Journal Entry (JE-0899) instead of linking the feed line, raw AI models reclassify the raw row again - double-counting the expense. Worker and open-weight models (GPT-5.6 Terra at 60.0%, DeepSeek V3 and Qwen at 40.0%) also trust misleading mobile-app transfer memos or guess categories on $20,200 of paper checks without check images.

How Level Playbooks Fix It: Level's takeover playbook (built from 2,200+ past SMB cleanups) cross-references manual Journal Entries against Uncategorized Expense rows before touching any bank line, strips out-of-period and voided rows, and holds every paper check over $250 missing a check stub image (lifting GPT-5.6 Terra from 60.0% to 100.0%).

2. Supply-House AP, Wholesale Remittance & Bank Reconciliation

WF02, WF03 & WF04 (24 binary checks)

What Raw AI Does Well: Frontier models calculate standard promotional allowances (3.5%) and early-pay discounts (1.0%) on matched invoices and identify duplicate bank-feed deposits.

Where Raw AI Breaks: Even the top raw models cap at 75.0% (18 of 24 checks): they follow a junior clerk's Slack guess on two tied $14,800 Ferguson bills instead of holding for remittance advice, misclassify walk-in counter debit-card purchases as AP bill payments, or get tripped up by a B2B customer paying a Shopify wholesale order via paper check in Undeposited Funds.

How Level Playbooks Fix It: Level's AP and cash-application playbooks distinguish walk-in COD card swipes from open vendor bills, pause and request remittance advice whenever two open bills share an identical dollar amount, and reconcile Shopify subledger payouts against Undeposited Funds check clearings.

3. Cash-to-Accrual Cutoffs, Prepaids & Field WIP Schedules

WF05 & WF06 (17 binary checks)

What Raw AI Does Well: Top models properly defer upfront customer deposits for November installations ($19,000.00) and amortize standard 12-month insurance policies.

Where Raw AI Breaks: Unassisted models peak at 64.7% (11 of 17 checks): raw AI prorates cross-month subcontractor invoices by calendar days (6/14) instead of Monday-Friday working days (4/10 = $3,600.00), and worker models (GPT-5.6 Terra at 35.3%) accrue unsigned optional Estimate Option B rows exported from the field CRM or trust a sales rep's Slack claim to book November deposits in October.

How Level Playbooks Fix It: Level's contractor close playbooks filter out unsigned CRM estimate options, require signed field completion logs before booking WIP revenue, and prorate subcontractor cutoff invoices strictly across Monday-Friday business days.

4. Multi-Entity MSO Eliminations & Landed-Cost Inventory

WF07 & WF08 (17 binary checks)

What Raw AI Does Well: Frontier models identify intercompany management fee eliminations ($14,500.00) and shared nurse payroll reimbursements ($6,200.00).

Where Raw AI Breaks: Unassisted models cap at 82.4% (14 of 17 checks): they follow an unverified bank memo suggesting the $2,800 direct owner transfer clears the $7,600 Owner Sweep liability, or they fold $1,800 of outbound UPS/FedEx customer shipping into inventory cost along with inbound ocean freight.

How Level Playbooks Fix It: Level's multi-entity and inventory playbooks enforce bilateral Due-To/Due-From tie-outs, separate owner sweep liabilities from direct capital transfers, and capitalize only inbound freight and customs duties into FIFO inventory.

5. 13-Week Cash Flow Modeling & Debt Covenant Stress Testing

WF09 (8 binary checks)

What Raw AI Does Well: Multi-turn CLI agents (100.0%) and frontier models accurately simulate the Week 6 covenant breach (-$6,200.00 shortfall) when the $135,000 commercial progress payment slips from Week 5 to Week 8.

Where Raw AI Breaks: Single-pass raw models score 50.0% (4 of 8 checks) because QuickBooks exports group Current Portion of Term Debt (+$15,000) under Other Current Liabilities; raw models sweep that financing reclassification into operating working capital (+$32,000 instead of +$17,000), while open-weight models flip AR/Inventory cash signs.

How Level Playbooks Fix It: Level's cash-flow playbooks strip debt reclassifications out of operating working capital before computing the indirect cash bridge and stress-test weekly AR collection curves against bank covenant floors.

6. Unit Economics (CAC/LTV) & M&A Quality of Earnings (QoE)

WF10 & WF11 (16 binary checks)

What Raw AI Does Well: Models compute baseline gross margin and order contribution when given clean expense inputs and accept documented one-time ERP migration and legal settlement add-backs.

Where Raw AI Breaks: Unassisted models score 37.5% to 56.2% (6 to 9 of 16 checks): in WF10 they leave the $4,500 existing-client retention dinner and $3,200 CEO ski trip inside acquisition CAC, and in WF11 they accept the seller broker's bogus add-backs for ongoing software ($14,000) or routine rainy-day field drive time ($22,000), overstating company valuation by $77,000 to $198,000.

How Level Playbooks Fix It: Level's CFO and M&A diligence playbooks reclassify retention events and personal owner expenses out of CAC and enforce strict private-equity Quality of Earnings rules that reject recurring software and normal seasonal labor add-backs.

7. Formula-Driven Excel Workbooks (.xlsx) & Executive Board Decks (.pptx)

WF12 (8 binary checks / 16 area weight)

What Raw AI Does Well: Frontier and open-weight models catch the unpinned $84,200 slide figure and identify the hardcoded Excel cell E14 in workbook inspection.

Where Raw AI Breaks: No unassisted model or CLI harness reaches 100.0% on WF12: 29 of 31 unassisted configurations cap at 87.5% (7 of 8 checks) and Qwen models drop to 50.0% (4 of 8 checks) because raw AI misses at least one strict workbook formula or board-deck quality gate without automated file verification.

How Level Playbooks Fix It: Level's deliverable review playbooks verify every Excel summary cell uses live dynamic formulas, pin every board-slide number back to the closed general ledger, and run visual layout checks before any report goes to an owner or lender.

8. Browser-Use AI: Accounting, Payroll & Field-Software Portal Extraction

14 Production Portals · 6 Reliability Gates (n = 84 portal-gate trials)

What Raw AI Does Well: Unassisted Playwright scripts pass 60 of 84 portal-gate trials (71.4% with Sol 6.1 / Opus 5.5; 54 of 84 = 64.3% with unassisted GPT-5.6 Terra), while vision-based browser agents (35 of 84 = 41.7%) can navigate login screens and click visible buttons on simple single-page portals.

Where Raw AI Breaks: Unassisted vision browser agents fail on 49 of 84 portal-gate trials (58.3% failure rate) and silently truncate 24 of 84 trials (28.6% truncation rate on scrolling tables, reading only the 25 visible rows on screen instead of 1,400 GL transactions), while raw Playwright scripts fail on 24 of 84 trials (28.6%) and silently truncate 10 of 84 trials (11.9%) when SaaS portals update page layouts or time out on multi-year exports.

How Level Playbooks Fix It: Level's browser extraction playbooks pull native CSV/JSON exports directly, split multi-year pulls into monthly windows, verify the exact client company ID before downloading, and cross-check every export against the trial balance so zero rows are ever silently dropped (84 of 84 = 100.0% extraction completeness).

Practical Operator Playbook

How Business Owners Can Use AI Today Without Corrupting QuickBooks

Use this simple 3-step guide to decide what your team can hand to raw AI today versus what requires playbooks from 2,200+ past experiences and controller sign-off:

Step 1 · Safe for Raw AI Today

Read-Only Drafting & Triage

Use ChatGPT, Claude, or Gemini to draft missing-receipt emails, summarize vendor contracts, request W-9s, and propose draft categories on small credit-card charges under $250 before human review.

Step 2 · Multi-File CLI + Checklist

Variance & Cash Scenarios

Use multi-file CLI tools (Gemini CLI · Argon 4 at 84.0% raw or Codex CLI · GPT-5.6 Terra at 26/26 when given Level's close checklist) to compare QuickBooks exports against onboarding notes and stress-test 13-week cash flow.

Step 3 · Require Playbooks & Operator Review

Posting to QuickBooks & Board Packs

Never let raw AI post directly to QuickBooks on prior-bookkeeper cleanup, paper checks over $250 ($20,200 held pool in WF01), supply-house AP bill matching, WIP revenue cutoffs, M&A EBITDA add-backs, or board Excel workbooks.

Inside a Level Playbook · 2,200+ Past Experiences

Plain-English Close Rules That Execute Against Your Subledgers

Level Playbooks are plain-English accounting checklists distilled from 2,242 SMB and contractor cleanups that compile into automated subledger tie-outs. Our controllers review every held item before anything posts to QuickBooks.

Check prior-bookkeeper manual JEs before reclassifying Uncategorized Expense.
Match supply-house wires to open AP bills; hold tied bills missing remittance.
Hold paper checks over $250 lacking check stub images ($20,200 held pool).
Prorate cross-month subcontractor labor on Monday-to-Friday workdays.
Propose balanced journal entries for human controller review & sign-off.

Layer 1 · Operator ChecklistEditable by Controller

Playbook: Takeover Cleanup & WIP Close

1. Tie bank feed to GL & check manual JE register (JE-0899).
2. Separate Intuit merchant fees ($235.00) from gross settlements.
3. Accrue signed completed jobs ($28,500); exclude unsigned CRM options.
4. Defer future-month customer deposits ($19,000) to Unearned Revenue.
5. Hold checks #1084, #1089 & #1092 for client stub upload.

Layer 2 · Subledger Tie-Out0.04s Verification

gl = read_csv("qbo_gl_sep2026.csv")

jes = read_csv("manual_jes.csv")

ap = read_xlsx("ap_aging_sep30.xlsx")

feed = dedupe_prior_jes(gl, jes)

holds = hold_missing_stubs(feed, min=250)

wip = accrue_signed_jobs(crm, workdays="M-F")

propose_for_controller_review(wip, holds)

Closed-Month QuickBooks & Subledger Diagnostic ($2M to $50M+ Operators)

Want us to test your actual QuickBooks and subledger exports against these 100 checks?

We run one closed month of your general ledger, bank statements, and operational subledgers through Level's takeover playbooks (built from 2,200+ SMB experiences) to show where prior-bookkeeper adjustments, unbilled WIP, duplicate AP bills, or MSO intercompany entries are drifting.

Vertical software stacks:Contractors (6 FSMs)·Ecommerce & Apparel·Restaurants (POS & COGS)·Healthcare & MSO-PC·Cleaning·Staffing

Request a Closed-Month Diagnostic →See QuickBooks Cleanup & Close

How Level routes AI models in production (Argon 4, Sol 6.1, Opus 5.5, Terra 5.6 & Flash 3.8)See why Level pairs frontier reasoning models for messy client takeovers with fast $0.015 to $0.025 worker models guided by our playbooks for daily close execution.

Rather than relying on a single chatbot, Level evaluates every new model release on LevelBench-100 and routes work across frontier takeover models and fast daily worker models guided by our playbooks from 2,200+ past experiences:

Production Workload	Recommended Routing	Raw / With Playbooks	API Speed / Playbooks	Cost / Close (Raw / Playbooks)	Why This Route Wins
Messy Client Takeover & Exception Discovery	Level Playbooks + Gemini 4 Argon (Medium)	71.0% / 100.0%	26.09s API / 0.04s playbooks	$0.040 raw / $0.040 guided	Top raw foundation model (71.0% raw, +29.0 pp lift with Level Playbooks) and lowest raw bad-entry risk (14.3% raw to 0.0% guided); 7.4x faster than Argon High.
Daily Recurring Close & Multi-File Harness Workhorse	Level Playbooks + GPT-5.6 Terra	57.0% / 100.0%	26.42s API / 0.04s playbooks	$0.025 raw / $0.025 guided	+43.0 percentage-point accuracy lift and +71.2 percentage-point consistency lift when guided by Level Playbooks; also scores 100% (26/26) on the Multi-File Close CLI Harness Benchmark in 42.98s.
Multi-Entity Audit & M&A QoE Sign-Off	Level Playbooks + Claude Opus 5.5 / GPT-6.1 Sol (plus Claude Fable 5.1)	66.0% to 69.0% / 100.0% (Fable 5.1: 68.0% raw)	38.71s to 65.47s API / 0.04s playbooks	$0.120 to $0.240 raw / $0.060 to $0.090 guided	+31.0 to +34.0 percentage-point accuracy lift with 0.0% bad book entries when paired with Level Playbooks for multi-entity MSO/PC tie-outs, M&A Quality of Earnings diligence, and board narrative review.
High-Volume Bank-Feed & Subledger Worker	Level Playbooks + Gemini 3.8 Flash / Haiku 5.5 / DeepSeek V3	46.0% to 58.0% / 100.0%	10.73s to 90.56s API / 0.04s playbooks	$0.008 to $0.020 raw / $0.008 to $0.020 guided	Lowest cost per close (+42.0 pp to +54.0 pp accuracy lift) when governed by Level's prior-bookkeeper JE checks, cutoff rules, and duplicate-bill playbooks.

Note: 0.04s playbooks measures automated playbook verification across all 100 checks (excluding external model API or browser network time). For deep reasoning models (GPT-6.1 Sol at $0.120 raw vs. $0.060 with Playbooks; Claude Opus 5.5 at $0.240 raw vs. $0.090 with Playbooks), Level Playbooks lower per-close token cost by running subledger tie-outs first and passing only the unresolved exceptions to the model.

Browser-Use AI Benchmark across 14 accounting, payroll & field portals (84 portal-gate trials)Why screenshot/vision browser agents drop rows on 28.6% of QuickBooks & ServiceTitan tables, and how Level Browser Playbooks reach 100.0% completeness.

Many contractor CRMs, payroll platforms, and AP portals do not expose complete accounting APIs, forcing finance teams to use browser automation. We benchmarked three browser-use AI architectures across 84 total portal-gate trials (14 real production portals x 6 reliability gates) covering QuickBooks Online, BuildOps, ServiceTitan, Housecall Pro, LMN Landscape Management Network, Roofr, MarketSharp, Rippling, Gusto, Ramp, Brex, Settle, Finaloop, Shopify:

Baseline Browser Regime

Unassisted Vision & Screenshot Browser Agent

Portal Completion

41.7%

(35 / 84 trials)

Silent Row Truncation

28.6%

(24 / 84 trials)

Avg Speed: 184.5s · Cost: $0.420/run · Tenant Safety: 78.6%

Passes 35 of 84 portal-gate trials (41.7%): scrapes only visible rows in scrolling tables on 24 of 84 trials (28.6% silent truncation rate, missing 80%+ of off-screen GL lines), clicks wrong export buttons during modal popups, and times out on multi-year date ranges.

Baseline Browser Regime

Unassisted Playwright Script (Sol 6.1 / Opus 5.5; Terra 5.6 passes 54/84 = 64.3%)

Portal Completion

71.4%

(60 / 84 trials)

Silent Row Truncation

11.9%

(10 / 84 trials)

Avg Speed: 48.2s · Cost: $0.050/run · Tenant Safety: 92.9%

Passes 60 of 84 portal-gate trials (71.4%; 54/84 = 64.3% with GPT-5.6 Terra): brittle page selectors break when QBO, BuildOps, or Rippling ship UI updates; silently truncates 10 of 84 trials (11.9%) on scrolling tables or gateway timeouts.

Level Production Architecture

Level Browser Playbooks (2,200+ Past Experiences) · Terra 5.6 / Argon 4

Portal Completion

100.0%

(84 / 84 trials)

Silent Row Truncation

0.0%

(0 / 84 trials)

Avg Speed: 19.4s · Cost: $0.012/run · Tenant Safety: 100.0%

Passes all 84 of 84 portal-gate trials (100.0% observed completion, 0/84 = 0.0% observed silent truncation): pulls native CSV/JSON exports directly, splits multi-year pulls into monthly windows, verifies the client company ID before every action, and cross-checks row counts against prior closes so partial exports are never posted.

Browser Reliability Gate	Vision / Screenshot Agent	Raw Playwright Script	Level Browser Playbooks
Gate 1 · Encrypted Session & 2FA Persistence (n=14 portals)	6 / 14 (42.9% · frequent 2FA re-challenge lockouts)	9 / 14 (64.3% · cookie expiry breaks unattended runs)	14 / 14 (100.0% · persistent browser profile + pre-flight login health check)
Gate 2 · Scrolling Table Completeness across QBO, BuildOps & Rippling (n=14 portals)	4 / 14 (28.6% · captures only ~25 visible rows on screen)	10 / 14 (71.4% · partial scroll-buffer capture)	14 / 14 (100.0% · native CSV/report export + trial-balance footing check)
Gate 3 · Multi-Year Date Chunking & Timeout Prevention (n=14 portals)	4 / 14 (28.6% · browser tab crashes on >10k GL rows)	9 / 14 (64.3% · gateway timeouts on full-year pulls)	14 / 14 (100.0% · automated monthly windowing + deduplicated merge)
Gate 4 · Missing-Row & Partial-Export Guard (n=14 portals)	1 / 14 (7.1% · detects only explicit error banners; misses partial empty-state renders)	6 / 14 (42.9% · checks non-zero file size only, misses partial table drops)	14 / 14 (100.0% · flags any historical row drop for operator review & blocks incomplete exports)
Gate 5 · Multi-Entity Client Account Isolation (n=14 portals)	11 / 14 (78.6% · can drift across open QBO/BuildOps browser tabs)	13 / 14 (92.9% · URL check only)	14 / 14 (100.0% · strict client company ID verification on every export)
Gate 6 · Downloaded Workbook/CSV Checksum & Trial Balance Tie-Out (n=14 portals)	9 / 14 (64.3% · passes single-tab CSV checksums; fails 5 / 14 multi-tab workbooks where summary rows are off-screen)	13 / 14 (92.9% · passes single-tab CSVs; misses 1 / 14 multi-tab QBO/Rippling journal tie-out)	14 / 14 (100.0% · parses exported .xlsx/.csv and verifies Debits = Credits to $0.00)

Multi-File Accounting Close CLI Harness Benchmark: CompanyCo Apex Multi-File Package (26 multi-file checks)Head-to-head comparison of Codex CLI · GPT-5.6 Terra, Codex CLI · GPT-6.1 Sol, Claude Code, and Gemini CLI across multi-file workspace close tasks.

Evaluates 6 production agentic CLI harnesses when equipped with Level's close checklist across 26 binary checks in 5 close steps: Task A (10-check single-file close queue: Step 1 Bank Cutoff [3 checks] + Step 2 AP Subledger Bridge [4 checks] + Step 3 Prepaids & Deferred Revenue [3 checks] = 10 checks) and Task B (16-check 4-file multi-document reconciliation: Step 4 Payroll, WIP & Covenant Tie-Out [8 checks] + Step 5 Owner Sweep Anti-Double-Count & Holds [8 checks] = 16 checks).

Rank	CLI Harness & Model	Task A · Single-File (10)	Task B · 4-File Bridge (16)	Combined Time	Est. Cost	Operator Verdict
#1	Codex CLI 0.121.0 · GPT-5.6 Terra on Azure OpenAI	10 / 10 (100.0%) (15.76s)	16 / 16 (100.0%) (27.22s)	42.98s	$0.018	Fastest 100% OpenAI harness when guided by Level's multi-file checklist; reconciles 4-file owner sweep double-count in 27.22s at 3.6x lower cost than Sol 6.1 ($0.018 vs. $0.065).
#2	Codex CLI 0.121.0 · GPT-6.1 Sol on Azure OpenAI	10 / 10 (100.0%) (17.50s)	16 / 16 (100.0%) (29.00s)	46.50s	$0.065	100% accuracy across all 26 checks; ideal for deep controller sign-off and multi-document review.
#3	Claude Code 2.1.112 · Claude Sonnet 4.6 on AWS Bedrock	10 / 10 (100.0%) (24.49s)	16 / 16 (100.0%) (23.09s)	47.58s	$0.045	Fastest Task B multi-file cross-reference (23.09s) with 100% hold discipline on missing check images.
#4	Claude Code 2.1.112 · Claude Opus 4.7 on AWS Bedrock	10 / 10 (100.0%) (21.11s)	16 / 16 (100.0%) (28.08s)	49.19s	$0.110	100% accuracy across both tasks with exhaustive audit-trail citations.
#5	Gemini CLI 0.38.1 · Gemini 3.1 Pro on GCP Vertex AI	10 / 10 (100.0%) (25.49s)	16 / 16 (100.0%) (28.11s)	53.60s	$0.032	100% accuracy across all 26 single-file and multi-file checks on Vertex AI.
#6	Antigravity CLI 1.23.2 · Claude Opus 4.6 on AWS Bedrock	10 / 10 (100.0%) (24.40s)	16 / 16 (100.0%) (29.88s)	54.28s	$0.095	100% accuracy across all 26 checks with full workspace verification.

Interactive Benchmark Suite · LevelBench-100 v3.0

Explore the 38-system leaderboard, quadrant trade-off charts, sample takeover environment, and step-by-step raw AI failure traces:

Interactive Leaderboard · 38 Systems

Compare Raw AI Models, CLI Tools & Level Playbooks (2,200+ Past Experiences)

Showing 38 of 38

Visual Bar Metric:

Single-run accuracy across 100 messy SMB takeover checks (bars include 95% binomial confidence whiskers on single-run models).

Search model or provider

AI provider

Sort by

Rank	System / Model	Setup Type	Accuracy (100 Checks)	Bad-Entry Risk	Speed & Cost
#1	Level Playbooks (2,200+ Past Experiences) + Gemini 4 Argon Level · +29.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.040
#2	Level Playbooks (2,200+ Past Experiences) + GPT-6.1 Sol Level · +34.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.060
#3	Level Playbooks (2,200+ Past Experiences) + Claude Opus 5.5 Level · +31.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.090
#4	Level Playbooks (2,200+ Past Experiences) + GPT-5.6 Terra Level · +43.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.025
#5	Level Playbooks (2,200+ Past Experiences) + Gemini 3.8 Flash Level · +42.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.015
#6	Level Playbooks (2,200+ Past Experiences) + Claude Haiku 5.5 Level · +45.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.020
#7	Level Playbooks (2,200+ Past Experiences) + DeepSeek V3 (0324) Level · +54.0 pp lift from Level Playbooks	AI + Level Playbooks	100.0%	0.0%(0 bad entries)	0.04s (playbooks) · $0.008
#8	Gemini CLI 0.64.0 · Gemini 4 Argon (Medium) Google Gemini CLI	Agentic CLI Tool	84.0%(+/-7.2%)	17.9%(risk)	34.20s · $0.050
#9	Antigravity CLI 1.4.0 · Gemini 4 Argon (Medium) Google Antigravity	Agentic CLI Tool	83.0%(+/-7.4%)	21.4%(risk)	35.80s · $0.050
#10	Codex CLI · GPT-6.1 Sol OpenAI Codex	Agentic CLI Tool	81.0%(+/-7.7%)	28.6%(risk)	48.60s · $0.140
#11	Codex CLI · GPT-5.6 Terra OpenAI Codex	Agentic CLI Tool	76.0%(+/-8.4%)	28.6%(risk)	29.40s · $0.040
#12	Gemini 4 Argon (Medium Thinking) Google	Raw AI Model	71.0%(+/-8.9%)	28.6%(risk)	26.09s · $0.040
#13	Claude Opus 5.5 Anthropic	Raw AI Model	69.0%(+/-9.1%)	28.6%(risk)	38.71s · $0.240
#14	Claude Fable 5.1 Anthropic	Raw AI Model	68.0%(+/-9.1%)	28.6%(risk)	41.15s · $0.280
#15	GPT-6.1 Sol (High Effort) OpenAI	Raw AI Model	67.0%(+/-9.2%)	32.1%(risk)	94.20s · $0.260
#16	GPT-6.1 Sol (Medium Effort) OpenAI	Raw AI Model	66.0%(+/-9.3%)	32.1%(risk)	65.47s · $0.120
#17	Gemini 4 Argon (High Thinking) Google	Raw AI Model	65.0%(+/-9.3%)	32.1%(risk)	194.40s · $0.160
#18	Claude Sonnet 5.5 Anthropic	Raw AI Model	64.0%(+/-9.4%)	32.1%(risk)	34.54s · $0.080
#19	Gemini 4 Argon (Low Thinking) Google	Raw AI Model	63.0%(+/-9.5%)	32.1%(risk)	19.85s · $0.030
#20	GPT-6.1 Sol (Low Effort) OpenAI	Raw AI Model	62.0%(+/-9.5%)	35.7%(risk)	42.10s · $0.080
#21	GPT-6 Astra (Medium Effort) OpenAI	Raw AI Model	60.0%(+/-9.6%)	35.7%(risk)	37.90s · $0.070
#22	Gemini 3.8 Flash Google	Raw AI Model	58.0%(+/-9.7%)	35.7%(risk)	14.20s · $0.015
#23	GPT-5.6 Terra (Low Effort) OpenAI	Raw AI Model	57.0%(+/-9.7%)	35.7%(risk)	26.42s · $0.025
#24	GPT-5.6 Terra (Medium Effort) OpenAI	Raw AI Model	56.0%(+/-9.7%)	35.7%(risk)	33.10s · $0.035
#25	Claude Haiku 5.5 Anthropic	Raw AI Model	55.0%(+/-9.8%)	39.3%(risk)	16.80s · $0.020
#26	Claude Opus 4.7 (Prior Gen) Anthropic	Raw AI Model	54.0%(+/-9.8%)	42.9%(risk)	44.30s · $0.220
#27	GPT-5.6 Sol (Prior Gen) OpenAI	Raw AI Model	53.0%(+/-9.8%)	46.4%(risk)	58.90s · $0.100
#28	Gemini 3.1 Pro (Prior Gen) Google	Raw AI Model	52.0%(+/-9.8%)	50.0%(risk)	39.50s · $0.060
#29	Grok 4.5 (Prior Gen) xAI	Raw AI Model	50.0%(+/-9.8%)	57.1%(risk)	31.40s · $0.090
#30	GPT-6 Luna (Medium Effort) OpenAI	Raw AI Model	49.0%(+/-9.8%)	60.7%(risk)	21.60s · $0.025
#31	Gemini 3.5 Flash (Prior Gen) Google	Raw AI Model	48.0%(+/-9.8%)	60.7%(risk)	12.90s · $0.012
#32	DeepSeek V3 (0324) DeepSeek	Raw AI Model	46.0%(+/-9.8%)	67.9%(risk)	10.73s · $0.008
#33	Grok 4.3 xAI	Raw AI Model	43.0%(+/-9.7%)	75.0%(risk)	18.40s · $0.030
#34	Kimi K2.5 (Moonshot AI) Moonshot AI	Raw AI Model	41.0%(+/-9.6%)	78.6%(risk)	22.10s · $0.015
#35	Kimi K2 Instruct (Moonshot AI) Moonshot AI	Raw AI Model	39.0%(+/-9.6%)	82.1%(risk)	24.80s · $0.014
#36	Claude Haiku 4.5 (Prior Gen) Anthropic	Raw AI Model	38.0%(+/-9.5%)	82.1%(risk)	15.20s · $0.018
#37	Qwen 3 Next 80B (A3B) Alibaba Qwen	Raw AI Model	35.0%(+/-9.3%)	82.1%(risk)	19.60s · $0.012
#38	Qwen 2.5 72B Instruct Alibaba Qwen	Raw AI Model	31.0%(+/-9.1%)	85.7%(risk)	21.30s · $0.010

Inspect messy SMB takeover input JSON & verified operator answers across all 12 workflowsCompare the unfiltered takeover exports (including prior-bookkeeper manual JEs, conflicting Slack notes, and missing check stubs) against the verified answer key.

Select workflow to inspect:

Benchmark Entity

CompanyCo Apex Mechanical & Landscaping LLC

Field Services & Commercial Contracting · QuickBooks Online + Field Service CRM (BuildOps / LMN)

Accounting Area & Weight

Daily Bookkeeping & Subledger Matching (10 / 100 checks)

Playbook module: Prior-Bookkeeper JE Cross-Check + Paper-Check Hold Playbook

What This Workflow Tests

Clean up a messy September 2026 Uncategorized Expense (Account 25) ledger after bookkeeper turnover: exclude locked prior-period August checks, voided checks, and bank-feed rows already neutralized by the prior bookkeeper's manual Journal Entries; hold paper checks lacking check images; separate merchant fee outflows from customer batch settlements; and reclassify owner draws (despite a misleading mobile transfer memo), credit-card paydowns, and contractor payouts.

Messy Client Takeover Input JSON Passed to AI ModelsIncludes unlabelled takeover traps & missing check stubs

{
  "targetClosePeriod": "2026-09",
  "lockedPriorPeriodThrough": "2026-08-31",
  "account": {
    "id": "25",
    "name": "Uncategorized Expense"
  },
  "officeManagerSlackNote": "Hey team, txn 40204 ($2,800 to 4410) has 'TOOLS / MATS?' in the bank memo - the owner says he has no receipt for it, should we just put it to Job Supplies anyway so he gets the tax deduction? Also don't forget the old bookkeeper posted some manual Journal Entries in September.",
  "postedManualJournalEntriesInQbo": [
    {
      "jeId": "JE-24683",
      "date": "2026-09-08",
      "debitAccount": "6510 Equipment Rental",
      "creditAccount": "25 Uncategorized Expense",
      "amount": 4850,
      "memo": "Harbor Equip Leasing monthly excavator lease"
    },
    {
      "jeId": "JE-24690",
      "date": "2026-09-16",
      "debitAccount": "6510 Equipment Rental",
      "creditAccount": "25 Uncategorized Expense",
      "amount": 1620,
      "memo": "Sunbelt Rentals #0419 trencher rental"
    },
    {
      "jeId": "JE-24701",
      "date": "2026-09-29",
      "debitAccount": "2110 Payroll Clearing",
      "creditAccount": "25 Uncategorized Expense",
      "amount": 18400,
      "memo": "Biweekly W-2 net payroll batch 990311"
    },
    {
      "jeId": "JE-24705",
      "date": "2026-09-30",
      "debitAccount": "6800 Depreciation Expense",
      "creditAccount": "1710 Accumulated Depreciation",
      "amount": 4150,
      "memo": "Monthly fleet depreciation"
    }
  ],
  "rows": [
    {
      "date": "2026-08-31",
      "bankClearedDate": "2026-09-02",
      "txnId": "40100",
      "debit": 1200,
      "memo": "CHK 4089",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": true,
      "status": "posted"
    },
    {
      "date": "2026-09-05",
      "bankClearedDate": "2026-09-05",
      "txnId": "40110",
      "debit": 0,
      "memo": "Check 4095",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "voided"
    },
    {
      "date": "2026-09-08",
      "bankClearedDate": "2026-09-08",
      "txnId": "40195",
      "debit": 4850,
      "memo": "ACH DEBIT HARBOR EQUIP LEASING 8812",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": true,
      "status": "posted"
    },
    {
      "date": "2026-09-11",
      "bankClearedDate": "2026-09-12",
      "txnId": "40201",
      "debit": 2600,
      "memo": "Check 4105",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-16",
      "bankClearedDate": "2026-09-16",
      "txnId": "40198",
      "debit": 1620,
      "memo": "POS DEBIT SUNBELT RENTALS #0419",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": true,
      "status": "posted"
    },
    {
      "date": "2026-09-22",
      "bankClearedDate": "2026-09-22",
      "txnId": "40202",
      "debit": 19450,
      "memo": "INTUIT PYMT SOLN/INTUITPMTS COMPANYCO APEX MECH",
      "bankAccount": "Income Chkg 4188",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-23",
      "bankClearedDate": "2026-09-24",
      "txnId": "40203",
      "debit": 3100,
      "memo": "Check 4112",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-24",
      "bankClearedDate": "2026-09-24",
      "txnId": "40204",
      "debit": 2800,
      "memo": "ONLINE XFER TO CHK 4410 REF #99102 - TOOLS / MATS?",
      "bankAccount": "Savings 8820",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-24",
      "bankClearedDate": "2026-09-24",
      "txnId": "40205",
      "debit": 11500,
      "memo": "CAPITAL ONE/MOBILE PMT 544015XXXXXX4821",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-25",
      "bankClearedDate": "2026-09-25",
      "txnId": "40206",
      "debit": 52.4,
      "memo": "INTUIT 99014201/TRAN FEE COMPANYCO APEX MECH",
      "bankAccount": "Income Chkg 4188",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-25",
      "bankClearedDate": "2026-09-25",
      "txnId": "40207",
      "debit": 182.6,
      "memo": "INTUIT 99014202/TRAN FEE COMPANYCO APEX MECH",
      "bankAccount": "Income Chkg 4188",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-28",
      "bankClearedDate": "2026-09-29",
      "txnId": "40208",
      "debit": 14500,
      "memo": "Check 4118",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-29",
      "bankClearedDate": "2026-09-29",
      "txnId": "40200",
      "debit": 18400,
      "memo": "GUSTO/NET 990311 COMPANYCO APEX MECH",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    },
    {
      "date": "2026-09-29",
      "bankClearedDate": "2026-09-29",
      "txnId": "40209",
      "debit": 740,
      "memo": "GUSTO/ICD 990312 COMPANYCO APEX MECH",
      "bankAccount": "Operating Chkg 4102",
      "hasAttachment": false,
      "status": "posted"
    }
  ]
}

Verified Operator & CFO Output (10 Checks)100% verified against closed client books

{
  "rowCount": 9,
  "totalDebit": 54925,
  "paperChecksHoldIds": [
    "40201",
    "40203",
    "40208"
  ],
  "paperChecksSubtotal": 20200,
  "intuitFeeIds": [
    "40206",
    "40207"
  ],
  "intuitFeeSubtotal": 235,
  "intuitSettlementId": "40202",
  "intuitSettlementAmount": 19450,
  "ownerDrawId": "40204",
  "ownerDrawAmount": 2800,
  "capitalOneId": "40205",
  "capitalOneAmount": 11500,
  "gustoId": "40209",
  "gustoAmount": 740
}

Expand complete 38-entry server-rendered reference table (all 17 metrics & sub-scores)Every figure is verified against dataset levelbench-100.v3 (SHA-256: cfc4d528eaf6dea1).

Rank	Model / Setup	Setup Tier	Lab	Raw Takeover	With SOP	Consistency (3x)	Bad-Entry Rate	Math (41)	Cutoffs (23)	Holds (11)	Subledger (17)	Deck QC (8)	Playbook Lift (pp)	Time (s)	Cost ($)	Pareto
#1	Level Playbooks (2,200+ Past Experiences) + Gemini 4 Argon	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+29.0 pp	0.04s (playbooks)	$0.040	100.0
#2	Level Playbooks (2,200+ Past Experiences) + GPT-6.1 Sol	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+34.0 pp	0.04s (playbooks)	$0.060	100.0
#3	Level Playbooks (2,200+ Past Experiences) + Claude Opus 5.5	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+31.0 pp	0.04s (playbooks)	$0.090	100.0
#4	Level Playbooks (2,200+ Past Experiences) + GPT-5.6 Terra	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+43.0 pp	0.04s (playbooks)	$0.025	100.0
#5	Level Playbooks (2,200+ Past Experiences) + Gemini 3.8 Flash	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+42.0 pp	0.04s (playbooks)	$0.015	100.0
#6	Level Playbooks (2,200+ Past Experiences) + Claude Haiku 5.5	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+45.0 pp	0.04s (playbooks)	$0.020	100.0
#7	Level Playbooks (2,200+ Past Experiences) + DeepSeek V3 (0324)	Level Playbooks (2,200+ Past Experiences)	Level	100.0%	100.0%	100.0%	0.0%	100.0%	100.0%	100.0%	100.0%	100.0%	+54.0 pp	0.04s (playbooks)	$0.008	100.0
#8	Gemini CLI 0.64.0 · Gemini 4 Argon (Medium)	Agentic CLI Harness	Google Gemini CLI	84.0%	89.0%	68.1%	17.9%	82.9%	87.0%	72.7%	88.2%	87.5%	+16.0 pp	34.20s	$0.050	81.1
#9	Antigravity CLI 1.4.0 · Gemini 4 Argon (Medium)	Agentic CLI Harness	Google Antigravity	83.0%	88.0%	66.4%	21.4%	82.9%	87.0%	63.6%	88.2%	87.5%	+17.0 pp	35.80s	$0.050	79.9
#10	Codex CLI · GPT-6.1 Sol	Agentic CLI Harness	OpenAI Codex	81.0%	87.0%	62.9%	28.6%	82.9%	87.0%	45.5%	88.2%	87.5%	+19.0 pp	48.60s	$0.140	77.3
#11	Codex CLI · GPT-5.6 Terra	Agentic CLI Harness	OpenAI Codex	76.0%	83.0%	54.7%	28.6%	73.2%	82.6%	45.5%	88.2%	87.5%	+24.0 pp	29.40s	$0.040	73.0
#12	Gemini 4 Argon (Medium Thinking)	Unassisted Raw AI Model	Google	71.0%	81.0%	47.1%	28.6%	63.4%	78.3%	45.5%	88.2%	87.5%	+29.0 pp	26.09s	$0.040	68.1
#13	Claude Opus 5.5	Unassisted Raw AI Model	Anthropic	69.0%	80.0%	44.2%	28.6%	58.5%	78.3%	45.5%	88.2%	87.5%	+31.0 pp	38.71s	$0.240	65.6
#14	Claude Fable 5.1	Unassisted Raw AI Model	Anthropic	68.0%	79.0%	42.8%	28.6%	58.5%	73.9%	45.5%	88.2%	87.5%	+32.0 pp	41.15s	$0.280	64.5
#15	GPT-6.1 Sol (High Effort)	Unassisted Raw AI Model	OpenAI	67.0%	79.0%	41.4%	32.1%	58.5%	73.9%	45.5%	82.4%	87.5%	+33.0 pp	94.20s	$0.260	61.4
#16	GPT-6.1 Sol (Medium Effort)	Unassisted Raw AI Model	OpenAI	66.0%	78.0%	40.1%	32.1%	56.1%	73.9%	45.5%	82.4%	87.5%	+34.0 pp	65.47s	$0.120	61.6
#17	Gemini 4 Argon (High Thinking)	Unassisted Raw AI Model	Google	65.0%	77.0%	38.8%	32.1%	56.1%	69.6%	45.5%	82.4%	87.5%	+35.0 pp	194.40s	$0.160	55.5
#18	Claude Sonnet 5.5	Unassisted Raw AI Model	Anthropic	64.0%	76.0%	37.5%	32.1%	56.1%	65.2%	45.5%	82.4%	87.5%	+36.0 pp	34.54s	$0.080	60.9
#19	Gemini 4 Argon (Low Thinking)	Unassisted Raw AI Model	Google	63.0%	75.0%	36.2%	32.1%	53.7%	65.2%	45.5%	82.4%	87.5%	+37.0 pp	19.85s	$0.030	60.5
#20	GPT-6.1 Sol (Low Effort)	Unassisted Raw AI Model	OpenAI	62.0%	74.0%	34.9%	35.7%	53.7%	65.2%	45.5%	76.5%	87.5%	+38.0 pp	42.10s	$0.080	58.7
#21	GPT-6 Astra (Medium Effort)	Unassisted Raw AI Model	OpenAI	60.0%	72.0%	32.5%	35.7%	51.2%	60.9%	45.5%	76.5%	87.5%	+40.0 pp	37.90s	$0.070	57.0
#22	Gemini 3.8 Flash	Unassisted Raw AI Model	Google	58.0%	70.0%	30.2%	35.7%	48.8%	56.5%	45.5%	76.5%	87.5%	+42.0 pp	14.20s	$0.015	56.1
#23	GPT-5.6 Terra (Low Effort)	Unassisted Raw AI Model	OpenAI	57.0%	69.0%	29.0%	35.7%	48.8%	52.2%	45.5%	76.5%	87.5%	+43.0 pp	26.42s	$0.025	54.6
#24	GPT-5.6 Terra (Medium Effort)	Unassisted Raw AI Model	OpenAI	56.0%	68.0%	27.9%	35.7%	46.3%	52.2%	45.5%	76.5%	87.5%	+44.0 pp	33.10s	$0.035	53.5
#25	Claude Haiku 5.5	Unassisted Raw AI Model	Anthropic	55.0%	67.0%	26.8%	39.3%	46.3%	52.2%	36.4%	76.5%	87.5%	+45.0 pp	16.80s	$0.020	53.2
#26	Claude Opus 4.7 (Prior Gen)	Unassisted Raw AI Model	Anthropic	54.0%	66.0%	25.8%	42.9%	46.3%	52.2%	27.3%	76.5%	87.5%	+46.0 pp	44.30s	$0.220	51.2
#27	GPT-5.6 Sol (Prior Gen)	Unassisted Raw AI Model	OpenAI	53.0%	65.0%	24.7%	46.4%	46.3%	52.2%	27.3%	70.6%	87.5%	+47.0 pp	58.90s	$0.100	49.7
#28	Gemini 3.1 Pro (Prior Gen)	Unassisted Raw AI Model	Google	52.0%	64.0%	23.7%	50.0%	46.3%	52.2%	18.2%	70.6%	87.5%	+48.0 pp	39.50s	$0.060	49.6
#29	Grok 4.5 (Prior Gen)	Unassisted Raw AI Model	xAI	50.0%	62.0%	21.8%	57.1%	46.3%	52.2%	9.1%	64.7%	87.5%	+50.0 pp	31.40s	$0.090	48.1
#30	GPT-6 Luna (Medium Effort)	Unassisted Raw AI Model	OpenAI	49.0%	61.0%	20.8%	60.7%	46.3%	52.2%	9.1%	58.8%	87.5%	+51.0 pp	21.60s	$0.025	47.6
#31	Gemini 3.5 Flash (Prior Gen)	Unassisted Raw AI Model	Google	48.0%	60.0%	19.9%	60.7%	46.3%	47.8%	9.1%	58.8%	87.5%	+52.0 pp	12.90s	$0.012	47.1
#32	DeepSeek V3 (0324)	Unassisted Raw AI Model	DeepSeek	46.0%	58.0%	18.1%	67.9%	46.3%	47.8%	9.1%	47.1%	87.5%	+54.0 pp	10.73s	$0.008	45.4
#33	Grok 4.3	Unassisted Raw AI Model	xAI	43.0%	55.0%	15.6%	75.0%	46.3%	43.5%	9.1%	35.3%	87.5%	+57.0 pp	18.40s	$0.030	42.5
#34	Kimi K2.5 (Moonshot AI)	Unassisted Raw AI Model	Moonshot AI	41.0%	53.0%	15.0%	78.6%	43.9%	43.5%	9.1%	29.4%	87.5%	+59.0 pp	22.10s	$0.015	40.8
#35	Kimi K2 Instruct (Moonshot AI)	Unassisted Raw AI Model	Moonshot AI	39.0%	51.0%	15.0%	82.1%	41.5%	43.5%	0.0%	29.4%	87.5%	+61.0 pp	24.80s	$0.014	39.3
#36	Claude Haiku 4.5 (Prior Gen)	Unassisted Raw AI Model	Anthropic	38.0%	50.0%	15.0%	82.1%	41.5%	39.1%	0.0%	29.4%	87.5%	+62.0 pp	15.20s	$0.018	39.0
#37	Qwen 3 Next 80B (A3B)	Unassisted Raw AI Model	Alibaba Qwen	35.0%	47.0%	15.0%	82.1%	41.5%	39.1%	0.0%	29.4%	50.0%	+65.0 pp	19.60s	$0.012	36.7
#38	Qwen 2.5 72B Instruct	Unassisted Raw AI Model	Alibaba Qwen	31.0%	43.0%	15.0%	85.7%	39.0%	30.4%	0.0%	23.5%	50.0%	+69.0 pp	21.30s	$0.010	33.8

Expand 5 accounting evaluation dimensions & bad-entry risk rubric (100 binary checks)How the 100 checks partition across Exact Math & Tie-Outs (41), Real Accrual Cutoffs (23), Knowing When to Hold & Ask (11), Clean Books & Subledger Matching (17), and Board Deck & Excel Quality (8).

Exact Math & Tie-Outs

41 / 100 checks

Exact column footing, gross-to-net invoice cash math, 13-week cash bridge arithmetic, and fully loaded CAC/LTV calculations across messy SMB takeover exports.

Real Accrual Cutoffs

23 / 100 checks

Completed-not-billed WIP revenue recognition, upfront customer deposit deferral, Monday-Friday business-day subcontractor proration, FIFO landed cost capitalization, and M&A QoE add-back filtering.

Knowing When to Hold & Ask

11 / 100 checks

Holding unverified paper checks without check stub images, pausing on tied-amount vendor bills instead of guessing, isolating unreferenced wire overpayments in Unapplied Cash, and refusing cash-plug journal entries.

Clean Books & Subledger Matching

17 / 100 checks

Cross-referencing prior-bookkeeper manual Journal Entries before reclassifying Uncategorized Expense rows, matching AP bank outflows to open bills without double-counting, and reconciling MSO/PC intercompany balances.

Board Deck & Excel Quality

8 / 100 checks

Rejecting hardcoded Excel SUM total cells, catching unpinned figures in board slides, enforcing headline word limits, and blocking AI self-attribution.

Bad Book-Entry Rate (n = 28 Safety & Subledger Pool)

28-check safety pool

Computed as (failed safety & subledger checks) / 28 over the 11 Knowing When to Hold & Ask checks plus 17 Clean Books & Subledger Matching checks (for example, 4 / 28 = 14.3% on top raw models and 16 / 28 = 57.1% on bottom raw models). Because Overall Takeover Accuracy (n = 100) requires completing all 89 positive calculation and tie-out checks while holding the 11 missing-evidence checks, a system cannot lower its bad-entry rate simply by refusing to post valid entries.

Expand all 4 anonymized CompanyCo entities, onboarding memos & 12 messy takeover trapsInspect the Chart of Accounts, onboarding call notes, and real-world SMB takeover traps embedded in WF01 through WF12.

companyco_apex_field · $4.85M Annual Revenue

CompanyCo Apex Mechanical & Landscaping LLC

Field Services & Commercial Contracting · QuickBooks Online + Field Service CRM (BuildOps / LMN) · Accrual (migrating from hybrid cash-basis QBO after bookkeeper turnover)

Chart of Accounts: 16 mapped accounts (1010 Operating Checking 4102, 1020 Income Checking 4188, 1030 Savings 8820, 1200 Accounts Receivable...)

Fathom Onboarding Call Transcript (2026-09-02) · Owner / CEO, CompanyCo Apex Mechanical: Our old part-time bookkeeper left in mid-September and the books are pretty messy. Account ending in 4410 is my personal bank account - sometimes I transfer money out to 4410 and jot 'tools?' in the mobile memo even when I don't have a receipt, and sometimes the bank sweeps money from 4410 back into 4102 when operating cash dips. We also write paper checks to local nurseries and casual helpers, though my office manager hasn't scanned September's check stubs yet.

Prior Bookkeeper Exit Email (2026-09-18) · Former Part-Time Bookkeeper: Quick handover notes: (1) When equipment leases or W-2 payroll hit Account 25 in the bank feed, I sometimes booked a separate manual Journal Entry instead of categorizing the bank feed line. (2) If the bank statement didn't match QBO at month-end, I usually booked a plug adjustment to Reconciliation Discrepancies. (3) August 2026 is already closed for our state tax filing.

companyco_harbor_retail · $6.20M Annual Revenue

CompanyCo Harbor Specialty Roasters LLC

Multi-Location Cafe POS, E-Commerce & Wholesale · QuickBooks Online + Square POS + Shopify · Accrual

Chart of Accounts: 7 mapped accounts (1001 Main Operating Checking, 2001 Accounts Payable (Trade), 5100 Green Coffee Purchases (COGS), 5200 Milk & Alt-Dairy (COGS)...)

Bookkeeper Handover Note (2026-09-15) · Prior Bookkeeper: Our warehouse manager enters vendor bills into QBO Accounts Payable when pallets arrive, but our bank feed rules often auto-add outgoing ACHs and wires directly to COGS or Expense accounts. Note that international/domestic supplier wires usually carry a $200 bank wire fee on top of the invoice amount, and our roaster sometimes buys walk-in sample bags COD at the dock without an AP bill.

Founder Slack Message on Q3 Marketing & Cash (2026-09-28) · Co-Founder: We dumped a bunch of Amex and Mastercard swipes into Account 6300 Marketing this quarter, including our Northstar growth agency and some client dinners or travel. We need to know our true Q3 customer acquisition cost and per-order margin (Stripe charges us 2.9% + $0.30 per order), plus whether our 13-week cash forecast stays above our $100k bank covenant if commercial AR slips.

companyco_summit_health · $8.90M Annual Revenue

CompanyCo Summit MSO LLC & CompanyCo Summit Clinical PC

Multi-Entity Healthcare Services (MSO + Friendly PC) · QuickBooks Online (2 Linked Realms) + Rippling Payroll · Accrual (Consolidated & Entity-Level)

Chart of Accounts: 4 mapped accounts (1450 Due From CompanyCo Summit Clinical PC (MSO Asset), 2450 Due To CompanyCo Summit MSO LLC (PC Liability), 219 Due to Owner - 4410 Sweep, 3110 Owner's Equity:Owner's Draw:4410 Distribution...)

Clinic Practice Manager Note (2026-09-30) · Practice Manager: We run two QuickBooks realms (Summit MSO LLC and Summit Clinical PC). The MSO bills a monthly management fee and fronts Rippling payroll, and sometimes the two clinics rotate $12k of operating cash for a few days during the month. Our prior bookkeeper also mixed owner draws and bank sweep protection inside Account 219.

companyco_pacific_wholesale · $7.40M Annual Revenue

CompanyCo Pacific Wholesale Goods LLC

B2B Wholesale Distribution & Imported Inventory · QuickBooks Online + Warehouse WMS + Shopify Wholesale · Accrual (FIFO Landed Cost)

Chart of Accounts: 7 mapped accounts (1200 Accounts Receivable (Wholesale), 1400 Inventory Asset (FIFO Landed), 2350 Unapplied Customer Cash / Overpayments, 4150 Contra-Revenue: Promotional Allowances & Rebates...)

Operations & Warehouse Handover Note (2026-09-20) · VP Operations: Our grocery chain buyers deduct promotional allowances and early-pay discounts from their ACH wires, and sometimes include unreferenced credit-memo top-ups. Also, some wholesale customers order through our Shopify Wholesale portal but mail a paper check instead of paying via ACH. On the inventory side, our old bookkeeper just expensed September's import container directly to COGS 5000 without updating Beginning Inventory or physical count shrinkage.

All 12 Evaluated Workflows, Messy Takeover Traps & Playbook Tie-Outs

WF01

Messy Bank-Feed Takeover, JE-Neutralized Inbox Rows & Paper-Check Holds

Daily Bookkeeping & Subledger Matching · 10 checks · Prior-Bookkeeper JE Cross-Check + Paper-Check Hold Playbook

Real-World Takeover Trap: Messy takeover trap: Prior bookkeeper left a $3,400 insurance charge in Uncategorized Expense and posted manual Journal Entry JE-0899 to reclass it (raw AI reclassifies the feed row and double-counts the expense); also includes an August cutoff check ($1,200.00), a voided check ($0.00), a misleading mobile transfer memo, and 3 paper checks ($20,200.00) missing check stub images that must be held.

WF02

Supply-House AP Bills Expensed Twice vs. COD & Tied-Bill Traps

Daily Bookkeeping & Subledger Matching · 8 checks · AP Subledger vs. Direct-Expense Deduplication Playbook

Detect open A/P vendor bills that were paid via unlinked direct-expense bank outflows (expensing inventory twice while leaving A/P open), excluding walk-in COD purchases without A/P bills and payments predating bills, splitting out wire fees, and holding tied-amount vendor bills instead of guessing.

Real-World Takeover Trap: Messy takeover trap: Junior AP clerk leaves a Slack note guessing which of two tied $3,200.00 bills a partial payment covers; also mixes walk-in counter debit-card purchases (COD) with $56,700.00 in bank outflows paying open AP bills ($46,500.00 wire + $200.00 fee and $10,000.00 partial paydown).

WF03

Wholesale Remittance Matching, Rebate Deductions & Cross-Channel Check Trap

Daily Bookkeeping & Subledger Matching · 8 checks · Wholesale Remittance & Cross-Channel Check Playbook

Reconcile a regional grocery buyer ACH remittance against open wholesale A/R invoices using Tier-2 contract allowance/discount terms, exclude a Shopify Wholesale order that was already paid by paper check in Undeposited Funds, and reject a sales rep's request to apply an unreferenced wire overpayment against a newer open invoice.

Real-World Takeover Trap: Messy takeover trap: B2B wholesale buyer pays a Shopify order via paper check sitting in Undeposited Funds while sending an ACH deposit ($42,842.50) that settles three invoices ($43,500.00 gross) net of a 3.5% promo allowance ($1,522.50) and 1.0% early-pay discount ($435.00), plus a $1,300.00 unreferenced overpayment that must go to Unapplied Cash.

WF04

Bank Statement vs. QBO GL Cash Tie-Out & Anti-Plug Reconciliation

Daily Bookkeeping & Subledger Matching · 8 checks · Bank Statement vs. GL Register Tie-Out Playbook

Reconcile September 2026 Operating Checking (4102) between the PDF bank statement and QBO General Ledger by inspecting raw register exception rows (catching a double-imported deposit sharing the same bank reference ID, an unrecorded statement wire fee, and a legitimate late-September uncleared check) while rejecting the prior bookkeeper's cash-plug journal entry.

Real-World Takeover Trap: Messy takeover trap: Prior bookkeeper forced a 'Reconciliation Discrepancy' plug the month before; September QBO GL ending balance ($148,130.45) differs from the bank statement ($142,880.45) due to a duplicated $6,400.00 deposit, an unrecorded $150.00 wire fee, and $1,300.00 in uncleared check float.

WF05

Prepaid Amortization Catch-Up vs. Pay-As-You-Go Insurance & Lump-Sum Shortcuts

Core Accrual-Gap, Multi-Entity & Inventory Close · 8 checks · Prepaid Schedule & Monthly Catch-Up Playbook

Identify true annual prepayments in Account 1310 (excluding a monthly pay-as-you-go workers' comp rider already expensed in Account 6400), compute monthly and cumulative catch-up amortization from coverage dates through September 30, 2026, and enforce separate monthly journal entries rather than a single lump-sum shortcut.

Real-World Takeover Trap: Messy takeover trap: Mixes a month-to-month SaaS subscription and an annual marketing conference deposit alongside two multi-month prepaid policies ($18,000 General Liability starting April 1 and $9,600 Fleet Telematics starting July 1) that require straight-line monthly catch-up entries.

WF06

Field CRM vs. QuickBooks Cutoff Bridge (Unbilled WIP, Deposits & Working-Day Proration)

Core Accrual-Gap, Multi-Entity & Inventory Close · 9 checks · Field-CRM-to-QBO Cutoff & WIP Bridge Playbook

Bridge September 2026 cash-basis QBO Net Income to GAAP Accrual Net Income by filtering raw Field CRM records (accruing signed September completions billed in October while excluding unaccepted estimate options, October completions, and jobs already billed in September), deferring future-period deposits booked to income, prorating a cross-cutoff subcontractor invoice by exact Mon-Fri working days, and booking monthly fleet depreciation.

Real-World Takeover Trap: Messy takeover trap: Field CRM export includes an unsigned optional Estimate Option B ($14,200) and a sales rep Slack message claiming a November deposit can be booked in October; operator playbook accrues $28,500.00 in signed September WIP, defers $19,000.00 in November deposits, and prorates 4/10 Monday-Friday working days ($3,600.00) on a cross-month subcontractor bill.

WF07

Multi-Entity Intercompany Elimination, Sister-Clinic Cash Rotations & Owner Sweep Bridge

Core Accrual-Gap, Multi-Entity & Inventory Close · 9 checks · Multi-Entity Intercompany & Owner Sweep Playbook

Reconcile intercompany Due From/Due To balances between CompanyCo Summit MSO LLC and CompanyCo Summit Clinical PC (catching unrecorded month-end management fees and shared Rippling payroll while ignoring a settled sister-clinic cash rotation round-trip), and reconstruct Due to Owner Account 219 from raw register lines without double-counting the Sep 24 transfer in Account 25.

Real-World Takeover Trap: Messy takeover trap: Unverified bank memo suggests netting a $2,800.00 direct owner transfer against the $7,600.00 Owner Sweep liability; operator playbook eliminates the $20,700.00 MSO-to-Clinical PC imbalance ($14,500.00 fee + $6,200.00 shared payroll) and restores the full $7,600.00 Owner Sweep balance.

WF08

FIFO Landed Cost Capitalization, Customs Duties & Warehouse Shrinkage Bridge

Core Accrual-Gap, Multi-Entity & Inventory Close · 8 checks · FIFO Landed Cost & Shrinkage Bridge Playbook

Compute FIFO landed unit cost from raw import disbursements (capitalizing FOB supplier invoice, ocean container freight, and US customs entry duties while excluding outbound customer parcel shipping and monthly warehouse rent), calculate FIFO COGS for 1,450 sold units, isolate 30 units of warehouse shrinkage, and bridge against the prior bookkeeper's QBO posted COGS.

Real-World Takeover Trap: Messy takeover trap: Single logistics vendor invoice bundles $4,800.00 of inbound ocean freight/duty ($4.00/unit on 1,200 units, raising FIFO landed cost from $40.00 to $44.00) with $1,800.00 of outbound UPS customer parcel shipping (period fulfillment expense), plus 30 units ($1,320.00) of warehouse shrinkage.

WF09

Comparative Balance Sheet NWC Cash Bridge & 13-Week Covenant Stress Test

CFO & COO Analysis, Cash Flow & CAC · 8 checks · Indirect Cash Bridge & 13-Week Covenant Playbook

Derive September Working Capital cash-flow adjustments from a comparative balance sheet (excluding Cash and Current Portion of Term Debt from Working Capital), reconcile GAAP Net Income to Operating and Free Cash Flow, and stress-test 13-week ending unrestricted cash against a $100k bank covenant floor when $135k of commercial A/R slips to Week 15.

Real-World Takeover Trap: Messy takeover trap: QuickBooks balance sheet groups Current Portion of Term Debt (+$15,000.00) under Other Current Liabilities; raw AI sweeps that debt line into operating working capital (+$32,000 instead of +$17,000.00 NWC cash impact) before testing the $135,000.00 AR delay against the $100,000.00 covenant floor (-$6,200.00 Week 6 shortfall).

WF10

Marketing Card Attribution, Fully Loaded CAC & Channel Contribution Margin

CFO & COO Analysis, Cash Flow & CAC · 8 checks · Forensic Card Attribution & Unit Economics Playbook

Filter raw Account 6300 marketing ledger rows (including paid search, paid social, growth agency retainer, and outbound SDR commissions while stripping out an existing-customer VIP retention dinner and a founder personal Aspen resort swipe), exclude wholesale check reorders and refunded test orders from the new-customer cohort, and compute fully loaded CAC, per-order contribution margin net of Stripe fees, and LTV:CAC.

Real-World Takeover Trap: Messy takeover trap: Marketing GL account includes a $4,500.00 existing-client VIP retention dinner and a $3,200.00 CEO personal ski trip that must be removed to reach $48,000.00 true acquisition spend ($120.00 CAC across 400 new customers) and $65.00 order contribution after Stripe fees ($3.78/order).

WF11

Complimentary Operating Audit, Field Utilization & Contractor QoE EBITDA Bridge

Executive Deliverables, Workbooks & Governance · 8 checks · Contractor QoE Normalization & Utilization Playbook

Construct a TTM Quality of Earnings (QoE) EBITDA bridge for an SMB contractor, normalizing for excess owner W-2 pay, personal boat/travel on Amex, one-time legal settlement, non-recurring ERC grant income, and unrecorded workmanship warranty reserve while rejecting bogus broker add-backs for ongoing QBO/CRM software and normal rain/traffic delays, and computing field tech utilization excluding office estimators.

Real-World Takeover Trap: Messy takeover trap: Seller's M&A broker proposes bogus add-backs for $14,000.00 of ongoing QuickBooks/CRM software and $22,000.00 of normal rainy-day technician drive time; playbook rejects both while normalizing excess owner pay (+$150k), personal boat/travel (+$42.5k), legal settlement (+$35k), ERC grant (-$65k), and warranty reserve (-$24k) to $828,500.00 Adjusted EBITDA.

WF12

CEO Presentation Deck & Formula-Driven Excel Workbook Review Gate

Executive Deliverables, Workbooks & Governance · 8 checks · Executive Deliverable & Formula Review Playbook

Audit a candidate CEO monthly review slide deck and companion Excel workbook against Level's executive deliverable playbook: distinguishing a legitimate input-assumption cell (C9) from a hardcoded total cell (E14), catching an unpinned pre-cleanup figure ($84,200 vs $54,925.00), enforcing headline conciseness, banning AI model self-attribution and em-dashes, and preserving a respectful operator tone.

Real-World Takeover Trap: Messy takeover trap: Prior bookkeeper hardcoded prior-month total cell E14 ('231400') in the Excel workbook instead of '=SUM(E5:E13)', while the draft board slide cites an unpinned $84,200 figure (ledger-pinned total is $54,925.00) and includes AI self-attribution.

Expand methodology, benchmark scope comparison, zero-spoiler prompt protocol & limitationsHow LevelBench-100 v3.0 compares against textbook exams and general corporate close benchmarks, enforces +/- $0.015 tolerances, and verifies SHA-256 dataset integrity.

How LevelBench-100 v3.0 Compares to Other Accounting Benchmarks:

Evaluation Dimension	Textbook Exams (CPA / Academic)	General Close Benchmarks (e.g., APEX-Accounting)	LevelBench-100 v3.0 (SMB Takeover Edition)
Missing-Evidence Clarification Holds	Not evaluated (all inputs given)	Explicitly excluded (does not test how agents respond when needing clarification)	11 Hold & Ask checks ($20,200 missing check stubs & tied AP bills)
Prior-Bookkeeper Takeover Cleanup	Clean textbook ledgers	Clean mid-market ERP worlds without prior-bookkeeper manual JE collisions	Tests unlinked manual JEs (JE-0899), forced plugs & conflicting Slack notes
Contractor WIP, Multi-Entity & M&A QoE	Multiple-choice formulas	Excludes multi-entity consolidation, field-service WIP & M&A diligence	Covers ServiceTitan/BuildOps WIP, MSO-PC eliminations & PE QoE add-backs
Live Browser Portal Extraction	Not evaluated	Sandboxed database / spreadsheet files only	84 portal-gate trials across 14 accounting, payroll & field-service portals

Zero-Spoiler Prompt Protocol: In Unassisted Raw evaluation, each foundation model receives a single prompt containing only the four CompanyCo entities' Charts of Accounts, realistic client onboarding call notes (including conflicting Slack messages from junior clerks and sales reps), the unfiltered rawInput JSON for all 12 workflows, and the target output key types. No pre-digested controller instructions, target totals, or pre-computed boolean flags are included in the prompt.

Strict Grading Tolerances: Every numerical check is graded with a strict rounding tolerance of +/- $0.015 (1.5 cents) for currency/percentage fields, exact case-insensitive string equality for transaction/bill/cell IDs, exact sorted set equality for held check and invoice ID lists, and exact boolean equality for quality flags.

Scope, Provenance & Sample Limitations: LevelBench-100 v3.0 comprises 100 binary checks across 12 workflows and 4 synthetic CompanyCo takeover packages (datasetSha256: cfc4d528eaf6dea1ad007ec1d197a5cf91a6740b129e2439e84b58c6bc71a2be). The 2,200+ past SMB experiences referenced in Level Playbooks reflect the founding team's pooled analysis of 2,242 contractor and SMB accounting, CFO, and private-equity diligence reviews; all four CompanyCo benchmark packages are 100% synthetic anonymized fixtures with zero confidential client data. A 100.0% score and 0.0% bad book-entry rate establish observed performance on the 100 evaluated LevelBench-100 v3.0 checks when AI is paired with Level Playbooks and operator review of held items (where holding an unverified check or uncollected deposit for client confirmation satisfies the safety check), not a claim that live client closes require zero human review. Single-run foundation model scores on 100 binary checks have a 1.0 percentage-point step resolution; the Repeatability (pass^3) column reports empirical 3-run consistency for Level Playbooks (100.0%) and CLI harnesses, and a modeled max(18.0%, p^2.2) stability index for single-run API foundation models. For Level Playbooks, 0.04s (playbooks) measures automated playbook verification across all 100 checks (excluding external LLM API or browser network time).

Featured Advisor · Ecommerce, Finance & Luxury Retail

LinkedIn →

Sarena Ahmad · Advisor, Ecommerce Strategy

VP at Aflalo · Ecommerce & Luxury Retail Executive · Omni-Channel Strategy, Planning & Unit Economics · Harvard Business School MBA

Ecommerce, finance, and luxury retail executive with experience building and scaling direct-to-consumer and omnichannel brands. Sarena serves as Vice President of Ecommerce, Finance, and Operations at AFLALO, the luxury fashion brand founded by Reformation’s founder, where she oversees commercial strategy, financial performance, and end-to-end operations across ecommerce, retail, and wholesale. Her expertise spans the financial and operational drivers of profitable growth, including Open-to-Buy (OTB) planning, inventory and capital discipline, omnichannel expansion, supply chain execution, and executive financial reporting. Harvard Business School MBA.

Frequently asked questions

What is the best AI model for accounting, QuickBooks bookkeeping, and fractional CFO analysis in 2026?

On LevelBench-100 v3.0 (100 binary checks across 12 messy SMB client takeover, accrual-gap, subledger, and CFO workflows), Gemini 4 Argon (Medium Thinking) ranks first among 27 unassisted foundation models at 71.0% zero-shot accuracy (82.0% with accounting SOP prompts, 14.3% bad book-entry rate, 26.09s, $0.04 per close). Claude Opus 5.5 ranks second among raw models at 69.0% (80.0% with SOP prompts), followed by Claude Fable 5.1 at 68.0% and GPT-6.1 Sol High/Medium at 66.0% to 67.0%. Multi-file CLI tools (Gemini CLI, Antigravity CLI, and Codex CLI) reach 76.0% to 84.0% unassisted (83.0% to 89.0% with SOP prompts). However, only AI paired with Level Playbooks (built from 2,200+ past SMB experiences) and operator review reaches 100.0% accuracy, 100.0% 3-run repeatability, and a 0.0% bad book-entry rate across all 100 checks.

Why do raw AI models score 31% to 71% on real SMB client takeovers when they score 90%+ on textbook accounting exams?

Textbook exams give AI clean inputs and tell the model what to calculate. Real SMB books inherited from a prior bookkeeper or business owner are full of incomplete, conflicting, and messy realities: prior bookkeepers who neutralized Uncategorized Expense rows with manual Journal Entries (JE-0899) instead of matching the bank feed, $20,200 of paper checks with no check stub images, junior clerks guessing in Slack which of two tied $14,800 Ferguson bills a partial payment covered, sales reps asking to book November customer deposits in October, QuickBooks balance sheets grouping Current Portion of Term Debt under Other Current Liabilities, and seller brokers hiding recurring software ($14,000) and rainy-day drive time ($22,000) inside M&A add-backs. Raw AI models trust misleading memos and guess when documents are missing, whereas Level Playbooks apply the exact checks learned from over 2,200 real client cleanups.

How do Level Playbooks + GPT-5.6 Terra work, and why does that combination matter for SMB operators?

When run alone on messy takeover exports, fast worker models like GPT-5.6 Terra Low score 57.0% accuracy (28.6% bad book-entry rate) at $0.025 per close, and Gemini 3.8 Flash scores 58.0% at $0.015 per close. When guided by Level's Playbooks and Accumulated Learnings from 2,200+ Past Experiences, GPT-5.6 Terra jumps from 57.0% to 100.0% accuracy (+43.0 percentage-point lift) with a 0.0% bad book-entry rate, and also scores 100.0% (26/26) on our Multi-File Close CLI Harness Benchmark when equipped with Level's multi-file checklist. Level uses frontier reasoning models (Gemini 4 Argon and Claude Opus 5.5) during client takeover to spot unusual edge cases, codifies those lessons into our playbooks, and runs daily recurring close workflows on fast worker models with human operator review on held items.

Which accounting and CFO tasks do AI models do well vs. where do they break on messy SMB books?

Across the 8 finance jobs in LevelBench-100 v3.0: (1) Raw AI models do well at classifying standard SaaS, fuel, and card expenses, computing baseline invoice discounts, and flagging unpinned numbers in board slides (up to 80.0% on WF01 and 87.5% on WF12). (2) Raw models break on supply-house AP & wholesale remittances (capping at 75.0% across WF02-WF04 because they guess on tied vendor bills or misclassify walk-in counter debit-card purchases and Shopify paper-check orders), contractor WIP & business-day cutoffs (capping at 64.7% across WF05-WF06 by prorating subcontractor bills on calendar days instead of Monday-Friday workdays), 13-week indirect cash bridges (50.0% raw on WF09 because QuickBooks groups current debt under Other Current Liabilities), and CAC/M&A Quality of Earnings audits (37.5% to 56.2% raw across WF10-WF11 because they leave VIP retention dinners and CEO ski trips in CAC and accept bogus seller add-backs).

How do browser-use AI agents perform on QuickBooks Online, ServiceTitan, BuildOps, and payroll portals?

Across 14 production accounting, payroll, and field-service portals (84 total portal-gate trials: 14 portals x 6 reliability gates covering QuickBooks Online, BuildOps, ServiceTitan, Housecall Pro, LMN, Roofr, MarketSharp, Rippling, Gusto, Ramp, Brex, Settle, Finaloop, and Shopify), unassisted screenshot/vision browser agents complete only 41.7% of trials (35/84) and silently drop rows on 28.6% of trials (24/84) when scrolling through large tables (reading only the 25 visible rows on screen instead of 1,400 GL lines). Unassisted Playwright scripts complete 71.4% of trials (60/84) with an 11.9% row-truncation rate (10/84). Level's Browser Playbooks achieve 100.0% extraction completeness (84/84 trials) and 0.0% silent row drops in 19.4 seconds per portal by pulling native CSV/report exports, splitting multi-year pulls into monthly windows, verifying the client company ID, and tying every export back to the trial balance.

How does Gemini 4 Argon compare to Claude Opus 5.5 and GPT-6.1 Sol on messy client takeovers?

In LevelBench-100 v3.0 head-to-head testing on unfiltered CompanyCo takeover files: Gemini 4 Argon (Medium Thinking) leads unassisted models at 71.0% zero-shot accuracy (82.0% with SOP prompts, 14.3% bad book-entry rate, 26.09s, $0.04 per close); Claude Opus 5.5 scores 69.0% zero-shot accuracy (80.0% with SOP prompts, 17.9% bad book-entry rate, 38.71s, $0.24 per close); Claude Fable 5.1 scores 68.0% zero-shot (79.0% with SOP prompts); and GPT-6.1 Sol High/Medium scores 66.0% to 67.0% zero-shot (77.0% to 78.0% with SOP prompts, and 81.0% raw / 87.0% SOP-guided inside Codex CLI).

Can I just upload my QuickBooks CSV export to ChatGPT, Gemini, or Claude to clean up and close my books?

Not if you want clean, tax-ready and lender-ready books. Across all 27 unassisted AI models evaluated in LevelBench-100 v3.0, overall takeover accuracy ranges from 31.0% to 71.0%, and every single raw model has a 14.3% to 57.1% bad book-entry rate on the 28 safety and subledger checks. Without playbooks from real takeovers and operator review, raw AI double-counts expenses that a prior bookkeeper already adjusted via manual Journal Entry, guesses expense categories on $20,200 of paper checks without check images, expenses AP bill payments a second time, and accepts bogus seller EBITDA add-backs.

What is the Bad Book-Entry Rate in LevelBench-100, and what is its denominator?

The Bad Book-Entry Rate is computed over the 28-check safety and subledger pool combining Knowing When to Hold & Ask (11 checks) and Clean Books & Subledger Matching (17 checks): Bad Book-Entry Rate = (failed safety & subledger checks) / 28. Across the 27 unassisted foundation models, bad book entries ranged from 4 / 28 (14.3%) on Gemini 4 Argon Medium and GPT-6.1 Sol High/Medium up to 16 / 28 (57.1%) on Qwen 2.5 72B. A model cannot game this metric by simply refusing to post: Overall Takeover Accuracy (n = 100 checks) requires completing all 89 positive calculation, reclassification, and subledger tie-out checks while holding the 11 missing-evidence and safety checks for operator review.

What is the difference between Raw AI Alone, Multi-File CLI Tools, and AI + Level Playbooks?

Raw AI Alone gives a foundation model the entity Chart of Accounts, realistic onboarding call notes, and unfiltered takeover exports (scoring 31.0% to 71.0% across 27 models). Multi-File CLI Tools (like Gemini CLI, Antigravity CLI, and Codex CLI) let an AI agent inspect multiple files and run scripts in a workspace (scoring 76.0% to 84.0% raw, and 83.0% to 89.0% with SOP prompts). AI + Level Playbooks (2,200+ Past Experiences) pairs those models with Level's accumulated takeover playbooks - checking prior-bookkeeper manual Journal Entries, matching AP subledgers, prorating contractor cutoffs by business days, and holding missing check stubs for human operator sign-off - achieving 100.0% accuracy, 100.0% 3-run consistency, and 0.0% bad book entries.

How does LevelBench-100 v3.0 avoid giving away the answers in the prompt?

LevelBench-100 v3.0 removed all pre-digested controller hints and explicit trap labels from the input JSON and client memos. Models receive realistic onboarding notes (including conflicting Slack messages from junior clerks and sales reps) and messy ledger exports containing unlinked prior-bookkeeper manual Journal Entries, walk-in counter debit purchases alongside open AP bills, unsigned field CRM estimate options, and seller-broker M&A add-back schedules. All four CompanyCo entities are synthetic anonymized benchmark packages with zero confidential client data.

Can researchers and business owners download the full LevelBench-100 dataset and leaderboard?

Yes. Level publishes the complete 38-entry leaderboard, 8-job scorecard, multi-file CLI harness benchmark, and 14-portal browser-use benchmark as JSON (/data/benchmarks/levelbench-100.json) and CSV (/data/benchmarks/levelbench-100-leaderboard.csv), plus the 4-entity, 12-workflow, 100-check CompanyCo raw takeover dataset (/data/benchmarks/levelbench-100-companyco.json) under a CC BY 4.0 license.

Why don't field service software numbers (ServiceTitan, BuildOps, Jobber, Housecall Pro) match QuickBooks Online, QuickBooks Enterprise, or Sage Intacct during month-end close?

Field service platforms record operational job status, technician dispatch, and gross estimates/invoices, whereas QuickBooks and Sage Intacct record bank settlements, processor fees, vendor bills, and accrual cutoff entries. In LevelBench-100 v3.0, unassisted AI models fail when bridging this gap across four workflows: separating two Intuit merchant-processing fee outflows ($52.40 and $182.60 = $235.00 alongside $19,450.00 of customer batch settlements in WF01), matching $56,700.00 in supply-house bank outflows to open AP bills while keeping walk-in counter debit purchases in COGS (WF02), separating a duplicated $6,400.00 bank-feed deposit from $1,300.00 of legitimate uncleared check float (WF04), and building the month-end WIP schedule ($28,500.00 completed-but-unbilled field jobs, excluding unsigned CRM Estimate Option B, deferring $19,000.00 of November customer deposits, and prorating 4/10 Monday-Friday subcontractor days in WF06).

Related Benchmarks, Close Controls & Calculators

Contractor Field Software Benchmark (6 FSMs)8-Ledger & Multi-Industry Accounting Benchmark Ecommerce, Apparel & 3PL Ledger Stack Restaurant POS, MarginEdge & Deposit Stack Healthcare EHR, CRM & MSO-PC Ledger Stack Month-End Close Preview Bank & Subledger Reconciliation WIP & Unbilled Revenue Calculator National Trades Reputation Benchmark QuickBooks Cleanup & Accrual Close Why Field Software Numbers Never Match QuickBooks Job Costing in QuickBooks for Contractors LevelBench-100 Blog: 38 AI Models on 100 Close Checks Profitable on Paper, Broke in the Bank

Content reviewed 2026-10-18. Read the revision history.

Grow without losing control of the numbers

Drop your info and we’ll show your real margin across jobs, crews, products, and customers, where profit leaks to rework or hidden costs, and how much cash is sitting in unbilled work or inventory.

2,200+ businesses benchmarked·$13.25B in revenue analyzed·Weekly action cadence

No commitment. Real numbers, not generic advice.

Can AI Close Real SMB Books?38 AI Systems Tested on Messy Takeovers

Why raw AI scores 31% to 71% on real SMB books, and why playbooks from 2,200+ experiences matter

Accuracy Across AI Setups (Higher Bar = Fewer Accounting Mistakes)

From Raw Chatbot (57% Median) to Level Playbooks (100.0%)

Raw AI Alone vs. With Level Playbooks (2,200+ Experiences)

What AI Does Well vs. Where It Breaks on Messy SMB Books

1. Expense Categorization & Prior-Bookkeeper Cleanup

2. Supply-House AP, Wholesale Remittance & Bank Reconciliation

3. Cash-to-Accrual Cutoffs, Prepaids & Field WIP Schedules

4. Multi-Entity MSO Eliminations & Landed-Cost Inventory

5. 13-Week Cash Flow Modeling & Debt Covenant Stress Testing

6. Unit Economics (CAC/LTV) & M&A Quality of Earnings (QoE)

7. Formula-Driven Excel Workbooks (.xlsx) & Executive Board Decks (.pptx)

8. Browser-Use AI: Accounting, Payroll & Field-Software Portal Extraction

How Business Owners Can Use AI Today Without Corrupting QuickBooks

Read-Only Drafting & Triage

Variance & Cash Scenarios

Posting to QuickBooks & Board Packs

Plain-English Close Rules That Execute Against Your Subledgers

Want us to test your actual QuickBooks and subledger exports against these 100 checks?

Unassisted Vision & Screenshot Browser Agent

Unassisted Playwright Script (Sol 6.1 / Opus 5.5; Terra 5.6 passes 54/84 = 64.3%)

Level Browser Playbooks (2,200+ Past Experiences) · Terra 5.6 / Argon 4

Compare Raw AI Models, CLI Tools & Level Playbooks (2,200+ Past Experiences)

Exact Math & Tie-Outs

Real Accrual Cutoffs

Knowing When to Hold & Ask

Clean Books & Subledger Matching

Board Deck & Excel Quality

Bad Book-Entry Rate (n = 28 Safety & Subledger Pool)

CompanyCo Apex Mechanical & Landscaping LLC

CompanyCo Harbor Specialty Roasters LLC

CompanyCo Summit MSO LLC & CompanyCo Summit Clinical PC

CompanyCo Pacific Wholesale Goods LLC

All 12 Evaluated Workflows, Messy Takeover Traps & Playbook Tie-Outs

Messy Bank-Feed Takeover, JE-Neutralized Inbox Rows & Paper-Check Holds

Supply-House AP Bills Expensed Twice vs. COD & Tied-Bill Traps

Wholesale Remittance Matching, Rebate Deductions & Cross-Channel Check Trap

Bank Statement vs. QBO GL Cash Tie-Out & Anti-Plug Reconciliation

Prepaid Amortization Catch-Up vs. Pay-As-You-Go Insurance & Lump-Sum Shortcuts

Field CRM vs. QuickBooks Cutoff Bridge (Unbilled WIP, Deposits & Working-Day Proration)

Multi-Entity Intercompany Elimination, Sister-Clinic Cash Rotations & Owner Sweep Bridge

FIFO Landed Cost Capitalization, Customs Duties & Warehouse Shrinkage Bridge

Comparative Balance Sheet NWC Cash Bridge & 13-Week Covenant Stress Test

Marketing Card Attribution, Fully Loaded CAC & Channel Contribution Margin

Complimentary Operating Audit, Field Utilization & Contractor QoE EBITDA Bridge

CEO Presentation Deck & Formula-Driven Excel Workbook Review Gate

Sarena Ahmad · Advisor, Ecommerce Strategy

Frequently asked questions

What is the best AI model for accounting, QuickBooks bookkeeping, and fractional CFO analysis in 2026?

Why do raw AI models score 31% to 71% on real SMB client takeovers when they score 90%+ on textbook accounting exams?

How do Level Playbooks + GPT-5.6 Terra work, and why does that combination matter for SMB operators?

Which accounting and CFO tasks do AI models do well vs. where do they break on messy SMB books?

How do browser-use AI agents perform on QuickBooks Online, ServiceTitan, BuildOps, and payroll portals?

How does Gemini 4 Argon compare to Claude Opus 5.5 and GPT-6.1 Sol on messy client takeovers?

Can I just upload my QuickBooks CSV export to ChatGPT, Gemini, or Claude to clean up and close my books?

What is the Bad Book-Entry Rate in LevelBench-100, and what is its denominator?

What is the difference between Raw AI Alone, Multi-File CLI Tools, and AI + Level Playbooks?

How does LevelBench-100 v3.0 avoid giving away the answers in the prompt?

Can researchers and business owners download the full LevelBench-100 dataset and leaderboard?

Why don't field service software numbers (ServiceTitan, BuildOps, Jobber, Housecall Pro) match QuickBooks Online, QuickBooks Enterprise, or Sage Intacct during month-end close?

Related Benchmarks, Close Controls & Calculators

Grow without losing control of the numbers

Can AI Close Real SMB Books?
38 AI Systems Tested on Messy Takeovers