The finance director of a 35-staff consulting firm has 600 invoices in her queue on the second of the month. Her team is three bank reconciliations behind. A partner has asked her, again, why the management accounts are five days late. Last year’s audit raised a “lack of supporting documentation” point on AI-assisted invoice coding because the previous platform could not explain its own classifications. She has £18,000 of approved Q3 budget and is meeting Tipalti, Dext, and Xero JAX next week.
Her question is the right one. Not whether AI belongs in finance, but which two of the six deployable jobs to start with, which platform’s audit trail will pass next year’s external audit, and what the 90-day rollout looks like with the data hygiene gate her predecessor skipped. Finance is the function where AI evidence is best quantified and where the boundary between deployable and judgment-heavy is sharpest. The useful framing is six jobs to deploy in order, four to leave with the qualified accountant, and one quiet failure mode the brochure will not mention.
What jobs does AI do well in finance and analytics today?
Six jobs have hit the maturity threshold at this revenue band. AP automation runs at 70 to 80 percent within three months on a Tipalti-class platform. Receipt capture on Dext or Navan cuts submission time 70 to 80 percent. Bank reconciliation hits 95 to 98 percent on routine matching. Cashflow forecasting on Float replaces weekly spreadsheet work. AR dunning on Chaser cuts DSO by three to five days.
Variance dashboards on Power BI cut management-pack production 20 to 40 percent, the sixth of the deployable jobs. The numbers stack. For a £4m revenue firm processing 1,000 invoices a month, duplicate-invoice prevention is £4,000 to £10,000 a year and 0.5 to 1.5 percent of spend recovered as early-payment discounts is another £20,000 to £60,000. Receipt capture returns 24 to 27 hours a month to a finance team of two. Cashflow forecasting frees 3 to 15 percent of operational working capital, which on a £5m firm is £40,000 to £80,000 of cash released. AR automation on the same firm frees roughly £44,000 more. The prior post on where invoice AI pays back goes deeper on the AP economics.
Where are UK SMEs actually using these tools?
The tooling stack at this revenue band has settled into a recognisable shape. Xero from £10 a month is the dominant cloud accounting platform, with Just Ask Xero (JAX) as the embedded LLM assistant. Sage Accounting and Sage Intacct cover practices with older client bases. Power BI at £14 to £30 per user a month has become the UK SME default for finance dashboards because it needs no dedicated data engineering resource.
Around that core, the specialists map cleanly to jobs. Dext (£33 to £99 a month) handles receipts and invoices, integrating directly with Xero and Sage. Tipalti starts at £99 a month for AP automation. Chaser (£324 to £899 a month) covers AR with AI-driven dunning. Float (£30 a month) and Futrli own cashflow forecasting. Navan handles expense and travel at £12 per user a month with policy enforcement before submission, not after. Spendesk and Ramp add corporate cards. Cake and Monzo Business sit at the smaller end. The right pick depends on integration with your existing accounting platform, invoice volume, and the audit-trail design.
Where does AI still fall short in finance?
Four jobs remain genuinely human work. Multi-entity consolidations involve judgment calls AI cannot reliably make on inter-company eliminations, equity-versus-debt characterisation, and overhead allocation. Deferred tax and transfer pricing involve rules with exceptions that defeat current models. Judgment-heavy statutory accounting sits with the finance director under FRC guidance. Making Tax Digital edge cases involve HMRC ambiguities AI cannot resolve.
The fifth boundary is the audit trail itself, and it is the quiet failure mode the brochure will not mention. The Financial Reporting Council and ICAEW both require interpretable explanations for AI-driven decisions, not just an audit log. A model that records “Invoice classified as Repairs and Maintenance with 87 percent confidence” passes a logging test and fails an audit test, because it cannot explain why in terms the external auditor accepts. The fix is to design the audit trail before the platform choice and ask each vendor for a worked example of an auditor walking through one of their AI-coded transactions.
What does a 90-day starter rollout look like?
Four phases. Days 1 to 14 are diagnostic: document invoice volume, expenses, cashflow, and AR, then prioritise (typically AP, receipts, reconciliation, cashflow, AR in that order). Days 15 to 42 are AP automation: train the model on 100 to 200 historical invoices, configure approvals, parallel-run for one week, then migrate. The cleanup justifies a separate piece on cleaning data before financial AI goes live.
Days 43 to 70 cover receipts and expenses, with more change management and four to eight hours of staff training to communicate that the new process reduces workload rather than increases surveillance. Days 71 to 90 finish bank reconciliation and dashboards, either through built-in reconciliation in Xero or Sage or 15 to 25 hours of Power BI consultant time. Total 90-day cost lands at £8,000 to £22,000 plus £150 to £400 a month ongoing. For a £3m to £5m revenue services firm, payback runs three to nine months, with quantified annual benefit of £15,000 to £40,000 across labour savings, duplicate prevention, working-capital release, and DSO reduction.
What should you ask a finance AI vendor before signing?
Five procurement questions separate the platforms that pass an external audit from the ones that do not. Show me the audit trail for an AI-coded invoice; if the answer is “we have an audit log” without “we explain the reasoning”, ICAEW guidance is breached. What is your Making Tax Digital posture against HMRC’s compatible-software specification? How do you handle the 5 to 10 percent of transactions needing special handling?
A vendor claiming 100 percent automation is either marketing or under-counting exceptions. Fourth, what is your security posture, specifically ISO 27001, SOC 2 Type II, and GDPR Article 32 technical and organisational measures? For employee expense data and customer payment data this is non-negotiable. Fifth, can your AI explain a credit-control or expense-approval decision to a customer asking why? UK GDPR Article 22 transparency obligations apply to automated decision-making with significant effects, and “the algorithm decided” is not a defence. Pair those five with the total-cost-of-ownership multiplier before signing anything, because the platform fee is rarely the largest line in the all-in cost. If you arrived here from where to apply AI first and answered finance, this is the rollout.
If you would like a second pair of eyes on which two jobs to start with, and on whether the audit trail design holds up against next year’s external audit, book a conversation.
