Accountable AI for finance operations is AI that makes routine finance decisions, such as posting a supplier invoice, under four conditions: it acts only inside written limits the firm has signed, it scores its own confidence on every decision, it records each decision in a tamper-evident log that cannot be edited after the fact, and a named person answers for the outcome. If any of the four is missing, the AI may still be useful, but it is not accountable.
That definition is deliberately narrow. Plenty of software uses machine learning to read invoices or suggest a nominal code. The question a finance owner should ask is not whether there is AI in the product, but who answers when it gets one wrong, and how anyone would find out.
Capability is rarely what stops AI reaching production in finance. Reading a PDF, extracting a VAT number and matching a supplier to a purchase order are solved problems. What stalls projects is the moment someone asks: who signed this off, on what basis, and can we show the auditor?
The wider evidence points the same way. Databricks, in its State of AI Agents 2026 report, cites MIT NANDA research from 2025 finding that 95% of generative AI pilots fail to reach production. The same Databricks report finds that companies that actively govern AI put 12 times more projects into production. Governance is not the brake. It is what lets the car leave the car park.
A mandate is a written statement of what the AI may do on its own. For supplier invoices it typically covers which suppliers, which value bands, which nominal codes and cost centres, and which conditions force a stop. The finance owner signs it. It changes only by a further signed revision, and the old version is kept, so you can always show what the system was permitted to do on a given day.
This is the same principle as a delegated authority matrix for people. Nobody lets a new AP clerk approve whatever they like in their first week. An AI system should not be treated more generously.
Every decision carries a score that says how sure the system is, and why. A useful score is built from named factors rather than one opaque number: the quality of the source data, how closely the invoice matches previous patterns, whether the proposed posting follows the firm's coding rules, whether it is consistent with the supplier's history, and whether any anomaly check fired. The score decides the route: approve, flag for review, or stop.
Each decision is written to an append-only log. Append-only means entries can be added but never changed or deleted. Tamper-evident means that if anyone did try to alter an entry, the alteration would show. The record holds the invoice, the evidence used, the score, the rule applied, the version of the mandate in force and the outcome. A log that can be edited is a diary, not evidence.
Somebody, by name, is accountable for the posting decision. Not "the system", not "the vendor", not "finance". The AI escalates unclear cases to that person, and that person can see everything the AI decided within the limits they signed.
It is not a promise that the AI will never be wrong. Bookkeepers make mistakes too. The difference is that an accountable system makes its mistakes visible, bounded and traceable to a rule someone signed.
It is not a chatbot bolted onto a ledger. Conversation is a poor control surface for posting decisions.
It is not a reason to remove human judgement. It moves human judgement to where it is worth having: setting the limits, and handling the cases that fall outside them.
Ask four questions, and expect documents rather than adjectives in reply:
If the answer to any of these is a demonstration rather than an artefact, keep asking.
Mandate builds accountable AI for finance operations, starting with accounts payable. Its agent, Invoice Maestro, makes the posting decision on a firm's supplier invoices inside a mandate the finance owner signs, scores its confidence on five factors, approves, flags or stops each invoice, escalates unclear cases to a named person and writes every decision to a tamper-evident, append-only audit trail. It is delivered as a managed service and works with Xero and Microsoft Dynamics 365 Business Central. The practical way to judge it is the 60-Day Proof: sixty days on your own invoices, read-only and free, across two month-ends, with a day-60 report you keep. Details at runmandate.co.uk/60-day-proof.
Not quite. Explainable AI tells you why a model produced an output. Accountable AI adds the limits it worked within, the record that proves what it did, and the person who answers for it.
No. It removes the need for a person to approve every routine item, and sends everything outside the signed limits to a named reviewer. Human judgement moves to setting the rules and handling the exceptions.
Usually the finance director or financial controller, because they already own delegated authority for spend. Where neither role exists, the managing director signs, and ownership is reviewed once a finance owner is in place.