AI Finance: Planning and Analysis Automation

Status: not built. Finance roles exist in our internal organisation experiment; no financial system has been built or run. Nothing here describes a product.

Finance is the worst fit for a probabilistic system, and the most tempting#

The appeal is obvious: financial analysis is document-heavy, repetitive, and full of work that looks like summarisation. The problem is equally obvious once stated — a language model produces a plausible number as readily as a correct one, and in finance a plausible wrong number is the definition of the failure.

Worse, wrong figures in finance propagate. A number in a report becomes a number in a board pack, becomes a decision. There is rarely a point where the chain is re-derived from source.

This does not mean AI has no place in finance. It means the boundary has to be drawn precisely, and the rule is straightforward:

The system may find, structure, explain and draft. It must not be the source of a number.

Every figure that reaches a report should be traceable to a system of record, and the arithmetic should be done by something deterministic.

Where it genuinely helps#

Explaining a variance. The number comes from the ledger; the question is why. A system that retrieves the relevant transactions, contracts and notes and drafts an explanation saves real time, and the explanation is checkable against sources it cites. See RAG.

Reading unstructured documents into structure. Invoices, contracts, statements, expense receipts. This is genuinely useful — with confidence thresholds and human review on anything below them, because extraction errors are silent.

Contract and obligation review. Finding the payment terms, renewal dates, penalty clauses and indexation across a pile of agreements nobody has read since signature. Finding is a good fit; concluding is not.

Drafting commentary. Management reporting narrative is largely formulaic. Draft it from figures the finance system produced, then have it reviewed. The draft saves time; the review is not optional.

Anomaly flagging for human attention. Not "this is fraud" but "this is unusual relative to history, here is why it was flagged". The output is an ordered queue for a person, not a decision.

Scenario structuring. Helping build the model, define the drivers and challenge the assumptions. The arithmetic stays in the spreadsheet.

Where we would not go#

Generating figures. Any number a model produces from reasoning rather than from a system of record is unverified, however confident the presentation.

Automated posting or payment. An irreversible financial action taken without human confirmation is the textbook case for the blast-radius discipline in AI agents.

Forecasts presented without their basis. A forecast is a set of assumptions. Delivering the output without them produces false confidence, and nobody can challenge what they cannot see.

Anything touching statutory reporting without a controls review. The requirements here are not negotiable and predate the technology.

The reconciliation problem#

This is the practical obstacle most projects underestimate.

Finance data lives in several systems that disagree — the ledger, the billing platform, the bank, the spreadsheet somebody maintains. Reconciling them is most of the real work, and it is not an AI problem. It is a data engineering problem: sources, ownership, freshness, and checks on values rather than on job exit codes. See data engineering.

A system built on top of unreconciled sources produces confident answers from whichever source it happened to read. The failure looks like an AI failure and is not.

The order that works: reconcile first, then automate. The order that is usually attempted is the reverse, because it demonstrates faster.

What we would require before any figure reached a report#

• Every figure traceable to a system of record, with the source shown
• Arithmetic performed deterministically, never by the model
• Sources reconciled, with disagreements surfaced rather than silently resolved
• Confidence thresholds on extraction, with human review below them
• A named person accountable for the output — automation does not transfer accountability
• An audit trail showing what was read, what was produced and who approved it
• Behaviour at period end tested, because that is when it is under load and the deadline is real

FAQ#

Do you offer financial automation?#

No. This is a problem study, not a product page.

Can a model do the calculations if it is careful?#

Do not let it. Arithmetic belongs in code or a spreadsheet, where it is deterministic and testable. Use the model to decide what to calculate and to explain the result — not to be the calculator.

What about reading invoices and receipts?#

One of the better fits, with two conditions: a confidence threshold with human review below it, and awareness that extraction errors are silent — a wrong date or a transposed figure looks exactly like a right one. Sample-check continuously rather than trusting an accuracy figure from evaluation.

Is anomaly detection worth it?#

Yes, as a queue for humans. The value is in the ordering and the stated reason, not in a verdict. Anything presented as a conclusion about a person or a transaction needs a human decision behind it.

Why is reconciliation such a problem?#

Because the systems genuinely disagree, for legitimate reasons — timing, scope, definitions. An automated layer on top will answer from whichever it reads, confidently. This is the single most common reason finance automation projects produce numbers nobody trusts.

Could AI replace an analyst?#

It can remove a large share of the retrieval and drafting. The judgement — which variance matters, which assumption is fragile, what to tell the board — is the job, and it is not the part that is automatable today.

What about audit and compliance requirements?#

They apply unchanged, and controls frameworks increasingly ask specific questions about automated processing in financial reporting. Any deployment touching statutory reporting needs a controls review before it starts, not after.

Related Articles#

See data engineering for the reconciliation layer underneath, RAG for answering from documents with sources, and AI agents for why irreversible actions need confirmation.

What else is coming for AI Finance