The audit that never ends

Thesis

Financial audits still rely on sampling: a team looks at a slice of transactions and infers the rest. That was a sensible compromise when people were the only way to read a document. It isn’t anymore. Models that can read a bank statement, an invoice and a ledger entry and reconcile all three now make it possible to check 100% of transactions, every day instead of a sample once a year.

Why now

• Document understanding went from brittle templates to general-purpose reading in about two years. I’ve seen this firsthand building FarmerCPA’s statement parser and classifier.
• The accounting profession is short of people: fewer CPA candidates and an ageing partner base TODO: cite AICPA pipeline data.
• Inference costs keep falling, so continuous checking has gone from unaffordable to cheap.

Why it’s big

Audit and assurance is a large, recurring, regulation-anchored spend TODO: verify market size. The bigger prize is what comes after it: once a system has verified every transaction, it holds the most trusted copy of a company’s financial truth, and lending, insurance and M&A diligence can all be built on top of that.

What has to be true

1. Error and fraud detection beats sampling on recall, provably, on real books.
2. At least one regulator-accepted path exists for machine evidence, or CPA firms adopt it as a tool before the rules move.
3. The wedge customer (mid-market companies or their CPA firms) pays for faster closes before it pays for “assurance.”

Biggest risk

Distribution. The incumbents control the signature on the audit opinion. The likely path is selling to CPA firms first, which can be slow.

How I’d diligence it

Ten calls with mid-market controllers and five with audit partners, plus a backtest: run the system on a closed year’s books and compare what it flags with what the human audit found.