Most board packs now mention AI. Far fewer finance teams can show where it changed invoice matching, the close checklist, or a cash forecast that operators trust. For mid-sized firms, AI in the finance function only matters when it shortens real work and tightens controls, not when it decorates a slide.
The pattern is now measurable. In a McKinsey survey of 102 CFOs, 44 per cent said they used gen AI for over five use cases in 2025, up from 7 per cent in the previous year. 65 per cent of respondents said their organisations will increase gen AI investment in 2025. That is a jump in tooling. It is not yet a jump in scaled value.
The same McKinsey article points to a State of AI 2025 finding that nearly two thirds of respondents said their organisations have not yet begun scaling AI across the enterprise. Pilots stall when they meet messy data, live invoices, and the month-end calendar. Mid-sized firms feel that stall sooner, because the team that runs AP is also the team that closes the books.
What AI in the finance function actually changes
McKinsey groups the useful work into three domains: planning and control, cash and working capital, and cost. That map is more helpful than a vendor catalogue. Start where the hours sit, then add a control that a human already owns.
Planning, commentary, and the close
Decision-support tools pull variances, draft first-pass commentary, and run scenarios in language a non-finance manager can follow. Across functions that have adopted this well, McKinsey has observed finance professionals spend 20 to 30 per cent less time crunching data. The saved hours only count if they go into a conversation with operations, not into another dashboard.
A consumer-goods example in the same article saved an estimated 30 per cent of finance professionals’ time on budget-variance insight. Treat that as a case, not a target you can copy. Your close still needs a named owner, a checklist, and a rule for what the model is allowed to post versus what it may only draft.
Payables, receivables, and leakage
Agentic workflows can read contracts against invoices and flag missed rebates, unused early-payment discounts, and tiered pricing that never hits the AP file. A global biotech case in the McKinsey survey identified contract leakage equal to approximately 4 per cent of total spend. That is not a promise for every ledger. It is a reminder that leakage often hides across many invoices, not in a single outlier.
Receivables benefit from the same discipline. Ageing is a timing report, not a collections strategy. An agent that ranks which invoices are late for a reason you can act on is useful. An agent that sends the same chase email to every overdue account is noise.
Spend and cost taxonomy
A European financial institution in the McKinsey article classified invoice-level spend into a four-level taxonomy, then looked for waste in energy, travel, and facilities. The work helped reduce costs by approximately 10 per cent of a multibillion-euro spend base. Mid-sized firms will not have 400 subcategories. They can still insist that free-text invoice descriptions get mapped to a short cost tree before anyone talks about AI savings.
This is also where first-party data systems earn their keep. If supplier names, entities, and cost codes are a mess, the model will classify the mess faster, not fix it.
Deployment is not the same as return
Deloitte’s Finance Trends 2026 press release, dated 8 October 2025, reports that 63 per cent of finance teams have fully deployed and actively use AI, while only 21 per cent report clear, measurable ROI. 14 per cent of respondents have fully integrated AI agents. Among the agent opportunities named, working capital optimisation (46 per cent) sits beside sales and profitability management and expense management.
That split should change how a mid-sized CFO sequences work. A fully deployed tool that nobody measures is an IT project. A narrowly scoped AP matching pilot with a weekly leakage report is a finance project. Moving agentic AI from pilot to production is the same problem in a different coat. Scope, data rights, and a human approval step have to be named before the first live invoice.
Deloitte also records data privacy concerns among 57 per cent of leaders in an advanced implementation stage. Privacy is not a late-stage luxury. If invoice PDFs, payroll files, and bank exports sit in a shared drive with a public model, you do not have an AI programme. You have a control failure. Keep the first use cases inside systems you already treat as confidential, with access logs a controller can read.
Skills, not slogans
The same Deloitte release finds 64 per cent of finance leaders plan to infuse more technical skills, such as AI, automation and data analysis, into their teams over the next two years. Traditional close skills do not become optional. The mix changes. Someone still has to know why a VAT code is wrong. Someone also has to know when a model has invented a supplier name.
Mid-sized firms rarely hire a full data-science bench. They can pair one technically fluent finance hire with a documented prompt and review standard, and they can refuse tools that cannot show their source rows. That is closer to scaling AI agents beyond pilots than to a transformation programme with a three-year slide.
McKinsey lists five stalls that match what we see in practice: waiting for perfect data, trying to rewire the whole function at once, launching pilots with no road map, skipping change management, and automating a fragmented process. None of those need a new statistic. They need a sequence. Pick one domain. Clean the inputs you already have. Put a named reviewer on the output. Measure hours returned and errors caught. Then expand.
A workable sequence for a mid-sized finance team
- Week 0. Write the two processes that consume the most hours, usually AP matching and month-end commentary. Name the control owner for each.
- Weeks 1 to 4. Run one assisted workflow on historical invoices or last quarter’s close pack. Compare the machine draft to the human file. Log misses.
- Weeks 5 to 8. Go live on a bounded slice, for example invoices under a stated value, with a mandatory human post. Report leakage found and hours saved, not AI adoption as a slogan.
- After that. Add cash forecasting or spend classification only if the first slice has a stable error rate and a privacy path you would show an auditor.
African mid-sized firms face the same process problem, plus thinner specialist benches and messier bank-file formats. AI investment priorities for African mid-sized firms should therefore start in finance operations, where the work is repetitive and the control standard is already written, rather than in a greenfield insight layer.
What to do next
Ask for one number this month. How many hours did AP and the close actually consume? Then pick one use case that can cut that number without posting unsupervised journals. Keep gen AI inside a system with access control. Measure leakage found or hours returned. If you cannot measure it, you are still in a demo.
AI in the finance function is a controls and capacity decision. The survey numbers show tools spreading faster than proven return. Mid-sized firms win by starting small, keeping a human on the posting path, and refusing any workflow that cannot show its source rows.
Key takeaways
- McKinsey’s CFO survey shows gen AI use for five-plus cases rising from 7 per cent to 44 per cent, while most organisations still have not begun scaling AI across the enterprise.
- Useful work sits in planning commentary, invoice-to-contract matching, and spend classification, not in a generic chatbot on the intranet.
- Deloitte finds 63 per cent of finance teams fully deploying AI and only 21 per cent reporting clear ROI. Treat deployment as a lagging vanity metric.
- Privacy and posting rights belong in the first design, especially once invoices and payroll files are in scope.
- Sequence one bounded process, measure hours and errors, then expand. Do not rewire the whole function in one programme.
FAQ
Where should a mid-sized firm start with AI in the finance function?
Start in a high-volume, rules-heavy process such as AP matching or close commentary. Keep a human on posting. Measure hours returned and errors caught before you add a second use case.
Does more gen AI use mean the finance team is more effective?
No. McKinsey shows a sharp rise in multi-use-case adoption. Deloitte shows a wide gap between full deployment and measurable ROI. Effectiveness is hours and control quality, not tool count.
Can AI post journals without a reviewer?
Not as a first step. Drafting commentary or flagging leakage is useful. Unsupervised posting creates a control hole that most mid-sized audit files cannot absorb.
Why do finance AI pilots stall?
They stall on messy master data, unclear posting rights, and processes that were never standardised. McKinsey’s stalls, waiting for perfect data and automating fragments, match that pattern.
How should privacy be handled for invoice and payroll files?
Keep those files in systems you already treat as confidential. Deloitte records privacy as a live concern for advanced implementers. Access logs and a named owner are part of the use case, not a later add-on.
Sources
- McKinsey, “How finance teams are putting AI to work today”, 3 November 2025. mckinsey.com/…/how-finance-teams-are-putting-ai-to-work-today
- McKinsey, “The state of AI in 2025”. mckinsey.com/…/the-state-of-ai-2025
- Deloitte, “Finance Trends 2026” press release, 8 October 2025. deloitte.com/…/finance-trends-2026-survey-release.html
This article is general information, not investment advice. See section 4 of our Terms & Conditions.
This is the kind of decision we advise on. If you are weighing it for your own business, see how we work or start a conversation.
