Short answer: AI quality of earnings software automates the data preparation underneath a QoE engagement — reconstructing cash, matching transactions, mapping a trial balance, extracting agreements. It does not automate the judgment, and any vendor implying otherwise is selling you a number rather than a conclusion.
For fee ranges, see quality of earnings report cost. For the review items themselves, see our quality of earnings checklist.
That distinction is the whole subject, so it is worth being precise about which parts are which.
The two halves of a quality of earnings engagement
A QoE has a mechanical half and an analytical half, and they behave completely differently.
The mechanical half is getting the data into a state where analysis is possible: pulling bank statements, reconciling every deposit and withdrawal to the general ledger, mapping a trial balance of several thousand accounts to a standard chart, reading leases and vendor contracts into a schedule. This work scales with record quality — how many accounts, how bad the scans, how many months — rather than with deal size. On a lower middle market engagement it is the large majority of the hours.
The analytical half is judgment. Is this add-back defensible. Does this customer relationship survive a change of control. Is the working capital peg being set on a number that means anything. Does management’s explanation of the Q3 margin swing hold up against the data.
Historically both were priced together because both were done by the same people at the same hourly rate. That is the only reason a small deal ever cost what it cost.
What automates reliably today
Extraction from real source documents
Not clean exports — the actual documents. Photocopied statements, image PDFs, tables with no ruling lines that break across pages without repeating headers. This is a computer vision and layout problem more than a text problem: knowing that a given figure is a withdrawal rather than a deposit, in a specific dated row, belonging to an account identified once at the top of page one.
Transaction matching
Pairing bank activity to ledger entries across the full period rather than sampled months, and — more importantly — correctly handling the cases that break naive matching: transfers between accounts the same owner controls, owner and related-party activity, merchant processor settlement arriving net and late. Those are covered in detail in four things that break a proof of cash.
Trial balance mapping
Classifying thousands of accounts to a standard chart with a confidence score attached to each mapping, so a reviewer knows where to look rather than checking everything.
Statement production
Generating GAAP-formatted monthly and twelve-month classified income statements and balance sheets with reported EBITDA, plus a cash flow statement, from the reconciled base.
Agreement extraction
Reading leases, vendor contracts and supply agreements into a structured schedule with the terms that matter for the transaction.
What does not automate, and should not
Normalization judgment. Whether owner compensation should be adjusted to a market replacement cost, and what that cost is, is an opinion informed by the industry and the role. A model can flag that owner compensation is unusual. It cannot decide what is fair.
Quality of revenue. Concentration, contract continuity, whether a customer stays after the founder leaves. Some of this is visible in the data; the conclusion is not.
Anything requiring a conversation. Management explanations are diligence evidence. Evaluating them is a human act.
The decision to stop. Knowing when the data cannot answer the question, and saying so, is the most valuable judgment in the whole engagement and the least automatable.
How to tell whether the output is right
The test is not the headline match rate. Match rate is the easiest number to optimize — match aggressively and you will pair two similar payroll runs or a card settlement with the wrong batch, and report a beautiful figure while quietly moving money to the wrong place.
The tests that matter are traceability and exception quality. Can you get from any conclusion back to the bank statement line and ledger entry behind it? Is the unresolved queue full of items a person genuinely has to decide about, or noise the system gave up on? Are low-confidence matches surfaced rather than silently absorbed?
We wrote the full version of that as seven tests to run before you buy, and mapped the tool landscape honestly in AI due diligence tools: the honest map.
What this looks like in practice
On a recent trailing-twelve-month engagement the mechanical half meant 431 manually scanned bank statements across 36 bank accounts, 29,823 transactions totaling $100.7 million in deposits and $99.0 million in withdrawals, against a general ledger of 1,365,571 journal entries across 5,196 accounts and 44 sites. Revenue and expenses each reconciled to within one percent of bank statement cash movements at a 99.4% average match rate.
One item was left flagged at the end: a $32 difference at a single location. That is the output working correctly. A system that resolves everything has not found anything.
By hand, that preparation is about four weeks. Automated, proof of cash comes back within 24 hours of ingestion and GAAP-formatted statements within 48 — which does not make the analytical half faster, but does mean the practitioner spends their time on it instead of on data entry.
Why the distinction matters commercially
Because it changes who can afford verification. A $30,000 engagement on a $4 million acquisition is a real share of the equity check, which is why small deals have historically been sampled or skipped. Automating the mechanical half moves the floor: our published price is $10,000 flat for a full proof of cash plus GAAP-formatted financials.
And where verification is no longer optional — SBA acquisitions at $3 million or more from October 1, 2026 — the arithmetic stops being a preference and starts being a requirement.
Talk to us about running one against a deal you already know the answer to.