Where AI genuinely helps
The useful applications share one property: the output can be checked against a source. That is what makes them safe to operate in a finance function.
- Commentary drafting — narrative generated from figures the model is handed, not figures it recalls, then reviewed by the analyst who owns the number
- Anomaly and variance detection — flagging what deserves investigation, with a human deciding what it means
- Document extraction — pulling structured data from invoices, contracts and statements, with confidence scores and an exception queue
- Reconciliation matching — proposing matches for human disposal, which removes volume without removing control
- Close task automation — the deterministic, repetitive steps that consume the first working week of every month
In each case the model proposes and the control is the verification step. That framing is what keeps AI useful and auditable at the same time.
Where it does not belong
Language models produce fluent, confident text regardless of whether the underlying figure is correct — and fluency is exactly what makes an error hard to catch in review. So:
- Never as the system of record for a number
- Never performing arithmetic a deterministic calculation can do
- Never recalling a figure when it could be handed one
- Never producing board, regulator or auditor output without a verifiable trail to source
The risk in finance is not that the model is wrong. It is that it is wrong fluently, in a format that reads as though someone checked it.
The governance you will need
- A documented map of AI-assisted steps in the process, so nothing is quietly automated.
- A named accountable human for every output, unchanged from the pre-AI process.
- Retained prompts and responses for audit.
- A defined verification step before anything reaches an external audience.
- A data policy covering what may be sent to which providers, and where it is processed.
Organisations subject to external audit should expect to evidence all five. Building them in at the start costs considerably less than reconstructing them under audit.
What it means for the team
AI reliably removes assembly work — gathering data, formatting packs, drafting routine commentary — which in many finance teams occupies the majority of the calendar. It does not remove judgement about what matters, the business context that makes a variance meaningful, or accountability for the number.
The realistic outcome is a shift in the ratio: less time preparing, more time analysing. That is the change most finance functions have been trying to make for a decade, and automation is the first lever that has moved it materially.
How to start without creating risk
Begin where output is internal, checkable and repetitive — commentary drafting on a management pack, or variance flagging for analyst review. Run it in parallel with the existing process for a full cycle so quality is measured rather than assumed. Keep a person accountable for anything leaving the team. The pilots that cause problems are the ones that start with external-facing output and no verification step.
AI in finance FAQs
Where does AI genuinely help?
Commentary drafting, anomaly detection, document extraction and reconciliation matching — cases where output is checkable against a source.
Where should it not be used?
As a system of record, for arithmetic, for recalling figures, or for board/regulator/auditor output without a verifiable trail to source.
What governance is required?
A documented map of AI-assisted steps, named accountability, retained prompts and responses, a verification step, and a data policy.
Will AI replace the FP&A analyst?
It removes assembly work, not judgement, context or accountability. The shift is less preparation, more analysis.
How do we start safely?
Internal, checkable, repetitive output first; run parallel to the existing process for a full cycle; keep a human accountable for anything external.
Where this fits
Automation works on a governed foundation — where definitions are agreed and data is reconciled. Where that is not yet true, start with data strategy and integration. For statistical prediction rather than language tasks, see predictive analytics & ML.
