Can AI do bookkeeping and bank reconciliation?
Yes — the reconciliation, the tracing and the drafting. Not the deciding. Marge was given governed read-and-draft access to a three-entity construction and design group with roughly 199 ledger accounts and twelve live bank and card feeds. She cleared a six-month backlog in under a day, found about $110,000 in phantom cash, and posted nothing without explicit owner approval.
The situation
A closely held construction and design group, three legal entities, in the middle of a capital raise. The founder was acting as his own CFO, which is a normal state of affairs in a company that size and a dangerous one during a raise. Reconciliation was six months behind across twelve live bank and card feeds.
Six months behind is not a bookkeeping inconvenience. It means the balance sheet you are showing investors is a guess, the cash number in your head is wrong, and you cannot tell the difference between a slow month and a hole.
What we built
Marge is an AI controller, not a reporting dashboard. The distinction is the access model.
She was given governed read-and-draft access to the accounting stack: she can query the ledger, trace a discrepancy through the transaction history, and draft the correcting entry. She cannot post it. Every entry sits in a queue until the owner explicitly approves it, and every action she takes — including the ones nobody approved — is written to a permanent, auditable run log.
Two hard boundaries were set before she touched anything. Nothing in the closed fiscal year gets modified, ever. And no entry posts on her own authority, no matter how obvious the fix looks.
What it produced
| Result | What it was |
|---|---|
| ~$110,000 | Phantom cash found and corrected — money the books said existed and did not. |
| $27,388.88 | Stale prior-period payables cleared in a single approved batch. |
| ~$8,358/week | Debt service surfaced by a rebuilt five-loan debt register — real payments nobody was tracking in one place. |
| Under a day | To clear the entire six-month reconciliation backlog across twelve feeds. |
| $95k–$170k | Modeled first-year economic impact. |
The debt register is the finding people underestimate. Five loans, each individually known, none of them ever added up on one page. $8,358 a week is $434,000 a year of debt service that existed but had never been looked at as a single number.
Straight talk
Zero unapproved entries, by design. Marge posts nothing without a human yes. That is not a limitation we are apologising for; it is the reason a system like this is safe to point at a general ledger at all.
The $95k–$170k is modeled, not audited. It combines recovered cash, avoided controller cost and the value of decisions made on correct numbers. The $110,000 and the $27,388.88 are actual ledger figures. We would rather split those apart for you than blur them together.
She did not decide what was true. She found the discrepancies and proposed the fixes. A person still had to look at each one and say yes.
Where this pattern fits
This is the right shape of build when the books are behind, when multiple entities make consolidation painful, or when you are the CFO by default and it is costing you the thing you are actually good at. It is also the right shape when you are about to raise, sell, or refinance, and someone is going to look closely at numbers you have not looked at closely yourself.
It is the wrong shape if your bookkeeping is current and clean and the real problem is that you do not read the reports. A system will not make you read them.
Bring your books. The audit maps where your hours and your cash are actually going, ranks the top five fixes by return, and gives you a written plan — whether or not you build with us.
Book Your AI Opportunity Audit — $1,000, credited toward your build