AI agents · finance

The CFO Agent

It answers the question a CFO actually gets asked — why did gross margin move — by walking the transactions rather than by summarising a chart. It operates at Level 0 and 1, because analysis that cannot be argued with is worse than no analysis.

0%25%50%75%100%wk 1wk 2wk 4wk 6wk 8wk 12wk 16wk 2091%
Explains to the transactionLevel 0 and 1 onlyDrafts, you decide

What it does

Six jobs, end to end.

Each of these is work a person does today. The agent does them in sequence and stops at the first thing it is not confident about.

Explain the movement

Margin down two points is decomposed into the customers, products, projects, vendors, and labour categories responsible, ranked by contribution to the change.

Forecast cash

From open receivables weighted by that customer’s actual payment behaviour, committed payables, payroll, and contracted recurring revenue — not from a straight-line extrapolation.

Model scenarios

Hiring plans, price changes, a lost account, a delayed collection. Each modelled against the real cost base rather than against a percentage assumption.

Draft the board pack

Statements, variance commentary, KPI movement, and the two or three things that changed materially — as a first draft you edit rather than assemble.

Surface what you did not ask

A customer concentration creeping up, a vendor whose pricing drifted, a project trending over. The things that do not appear in a report because nobody built one for them.

Answer in plain language

Ask why professional services margin fell in the Austin office last quarter and get an answer with the transactions attached rather than a chart to interpret.

Authority

Where this agent sits, and who decides.

Level 0 · default

Read and analyse

Almost everything this agent does. Nothing changes, nothing posts, and no approval is required because there is nothing to approve.

Level 1 · optional

Draft documents

Board packs, commentary, and lender reporting prepared for review. A person edits and owns what goes out under their name.

Never

Post or commit anything

This agent has no write access to the ledger at all. It is an analytical role, and giving analysis the ability to change the numbers it analyses is a category error.

Why explanation beats dashboards

Every ERP has dashboards, and they answer the question of what happened. The question a CFO is actually asked in a board meeting is why — and the gap between those two is a week of somebody’s analysis, repeated monthly, mostly re-deriving the same decomposition.

Because every transaction carries its dimensions and links back to the contract, project, and vendor behind it, the decomposition is a traversal rather than an investigation. Margin moved two points; here are the eleven things that contributed, ranked, with the largest being one customer whose mix shifted toward a lower-margin service line.

A dashboard tells you margin fell. The question you were asked is why, and that gap is somebody’s week, every month.

Forecasting on behaviour, not on terms

Most cash forecasts age receivables on invoice terms, which assumes customers pay when they agreed to. They do not, and the variance is not random — each customer has a stable personal habit that is usually more predictive than the contract.

Weighting each open invoice by that customer’s own payment history produces a forecast that is materially better than a terms-based one, and it degrades gracefully: a customer with no history is weighted on terms and flagged as an assumption rather than silently included.

What it cannot do, deliberately

This agent has no write access to the ledger whatsoever — not at Level 1, not with any configuration. An analytical agent that could adjust the numbers it reports on would be unsound in a way no audit trail could repair, so the capability simply does not exist.

It also does not make recommendations dressed as conclusions. It will tell you that margin fell because of a mix shift at one account; whether that is a pricing problem, a delivery problem, or a deliberate strategy is a judgement it does not have the context to make, and pretending otherwise would make the useful parts less trustworthy.

Questions

What people ask first.

Can it produce our board pack?
It drafts one — statements, variance commentary, KPI movement, and what changed materially. A person edits it and owns it, because it goes out under their name.
How accurate is the cash forecast?
Materially better than terms-based ageing because it weights each receivable by that customer’s actual payment behaviour. Accuracy improves with history; customers with none are weighted on terms and flagged as assumptions.
Does it have access to our ledger?
Read access only, at every level. It has no write path to the ledger at all, which is deliberate — analysis that can change the numbers it analyses is unsound regardless of the audit trail.
Can it answer questions in plain language?
Yes, and the answer comes with the underlying transactions attached rather than as a summary you have to take on trust.
Will it tell us what to do?
It will tell you what happened and what contributed. Whether a mix shift is a pricing problem or a strategy is a judgement it lacks the context for, and we would rather it stay silent there than sound confident.

Ask it why margin moved.

Two periods of ledger data and one real question is enough to see what it produces.