Finance: AI Agents
Forecasting, invoice processing, anomaly detection and project profitability, and the payment and reporting decisions that never leave human hands.
Markdown. No sign-up, no email.
Finance has the clearest line in this framework. Analysis, forecasting and matching are pattern work at volume. Moving money is not, and the reason is not capability. It is that fraud specifically targets whatever step lacks a human, and an automated payment path is the single most attractive thing an attacker can find inside a company.
Forecasting#
| Forecast | Built from | Where it goes wrong |
|---|---|---|
| Budget | History, committed spend, plan | Assumes next quarter resembles last, which is exactly false at a turning point |
| Cash flow | Receivables, payables, payroll, tax dates | Optimistic collection assumptions. Customers pay late, consistently |
| Revenue | Pipeline weighted by stage, plus recurring | Inherits sales optimism unless calibrated against actual close rates |
| Project profitability | Time, cost and scope against contract value | Unbilled scope creep, which is invisible until the margin is gone |
Every forecast ships with its assumptions visible. A cash-flow projection assuming 30-day collection when the actual average is 47 is not a forecast, it is a wish with a chart, and the gap only becomes obvious in the month it matters.
Invoice processing#
Extraction, matching against purchase order and receipt, coding, and flagging exceptions.
Genuinely well suited to automation, with one rule: it prepares payments and never releases them. The three-way match is a control, and a control both performed and approved by the same system is not a control.
Anomaly detection#
Duplicate invoices, unusual amounts, new bank details on an existing supplier, spend outside normal patterns, round numbers just under an approval threshold.
New bank details on an existing supplier is the highest-value single flag in this list. It is the mechanic of most invoice fraud, it looks entirely routine, and it is caught by a phone call to a number you already had rather than to the one on the letter.
Project profitability#
Continuous, per project, rather than discovered at the end.
A project that will lose money usually announces this in month two through unbilled hours and quiet scope drift. Discovering it at closure means the information arrived when nothing could be done with it.
What is never automated#
- Releasing a payment. Any payment. This is the line, and it does not move.
- Changing supplier bank details. Verified out of band, by a person, to a number held before the request arrived.
- Approving spend beyond the delegated authority of the person requesting it.
- Signing off financial statements.
- Recognising revenue. Judgement, and often a regulated judgement.
- Writing off a debt or a project.
Two-person rule on anything above threshold, and the second person is not the one who prepared it. This is old, unglamorous and remains the single most effective financial control ever devised.
What good looks like here#
The measure of this function is not how much of the ledger is automated. It is whether the CEO could answer, today, without a project: what is our runway, which projects are profitable, and what are we committed to spending that has not yet been spent.
Those three answers are what the automation exists to make continuously available. Everything else is bookkeeping, and bookkeeping was never the constraint.