A UK MSP finance director should evaluate AI for accounts payable on five tests: whether it copes with the invoices that actually cause trouble in an MSP, such as cloud licensing, distributor hardware and carrier bills; whether it acts only inside limits you sign; whether every decision is recorded in a form your auditor will accept; whether it works with your ledger; and whether you can trial it on your own invoices, read-only, before committing. Demonstrations on tidy sample data tell you very little.
MSP supplier ledgers have a particular shape. A handful of suppliers account for much of the value, and their invoices are the awkward ones:
Before speaking to any vendor, pull three months of these and note which caused rework, late postings or margin surprises. That list is your test set. A tool that handles stationery invoices beautifully and stalls on a forty-line distributor invoice with three project references has not solved your problem.
Vendor demonstrations use invoices that have been to finishing school. Insist on a trial on your own supplier invoices, read-only, so nothing posts to your ledger until you have seen the decisions. Run it across one month-end as a minimum and ideally two, because month-end is where volume, cut-off and late invoices meet.
Ask for output you can check: for each invoice, the proposed coding, the confidence and the reason, and what the system would have done. Then compare it with what your team actually posted.
The record on AI projects is sobering. 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 report cites MIT Technology Review Insights research showing that only 2% of organisations rate their AI as delivering measurable business results. Pilots rarely fail because a model cannot read an invoice. They fail because nobody can say who authorised what.
So ask:
You are giving a supplier access to your financial records and your suppliers' bank details. Ask where the data is hosted and in which region, what security certifications the supplier holds (Cyber Essentials Plus is a sensible baseline for a UK supplier handling financial data), whether it is registered with the ICO, what access it needs on day one, and how encryption keys are managed. Ask what happens to your data and decision records if you leave.
Many UK MSPs run Xero or Microsoft Dynamics 365 Business Central. Check that the system reads your chart of accounts, tracking categories or dimensions, and supplier records as they are, rather than asking you to restructure them first. Ask how it copes with a mid-year change to the chart of accounts.
Avoid long commitments before you have evidence. A sensible structure is a defined proof period on your own data, followed by a rolling arrangement you can leave. Be wary of any contract under which the decision record belongs to the vendor or is hard to export.
Score each candidate on: handling of your awkward invoices; signed limits; confidence explained by factor; escalation to a named person; tamper-evident record; ledger fit; security and hosting; exit terms. If a candidate fails on the record or the limits, the rest does not matter.
Mandate was built for this buyer. Its agent, Invoice Maestro, makes the posting decision on supplier invoices inside a mandate the finance director signs, scores its confidence on five factors, 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, works with Xero and Microsoft Dynamics 365 Business Central, and is UK-hosted (AWS London, eu-west-2), Cyber Essentials Plus certified, Defence Cyber Certification (DCC) Level 0 certified and registered with the ICO. The 60-Day Proof puts it through the tests above: sixty days on your own invoices, read-only and free, across two month-ends, with a day-60 report you keep and a rolling monthly contract from day 61 only if you choose. Details at runmandate.co.uk/60-day-proof.
Long enough to cover two month-ends, so you see cut-off, late invoices and volume peaks. A trial that ends before a month-end shows you very little.
No. A read-only trial lets you compare the system's proposed decisions with what your team actually posted, with no risk to the books.
That is common, and a trial is a good moment to fix it. Writing the rules down is the first step in setting the limits that any system, or any new member of the team, should work within.