Free Website Audit: Discover what's holding your digital presence back
Short answer
Give AI one task first: high volume, repetitive and easy to check, such as capturing supplier invoice data for accounts payable. Measure the task for one month before you start, then compare. Keep a person on every payment, journal and report. Expand to the close, receivables or forecasting only when the first task holds up.
A finance team of three cannot test ten AI tools at once. You have a close to finish, suppliers to pay and a board pack to send. The useful question is narrower: which one task do you give to AI first, and how do you prove it helped?
This guide is for CFOs, finance managers and owner-accountants in companies of 10 to 250 people. It assumes you already own a spreadsheet tool, an office suite and an accounting system.
Why are most CFOs not yet satisfied with AI?
Most CFOs have not yet seen the results they expected. A Bain & Company survey of more than 100 CFOs found that “just 31% of CFOs are satisfied with their AI outcomes”, and that “only 15%-25% of CFOs have scaled AI across finance functions” (Bain & Company, 2026 CFO Survey, 13 April 2026). Planning, analysis and reporting get the largest share of finance AI spend in that survey.
A smaller case shows why. The finance team at Bloom & Wild took about two months to embed an AI close tool and train staff. One feature, AI bank reconciliation, added nothing, because the team already had a good method in its system. The case gives no measured savings (ICAEW Insights, 10 November 2025).
Two lessons follow. Setup takes weeks, not days. And AI can add nothing to a step that already works well.
Which finance tasks suit AI, and which still need a person?
AI suits tasks where the input is messy text or documents and a person can check the output fast. It does not suit tasks where one wrong answer moves money or changes a number you report.
Use this table to rate your own tasks.
| Task | Fit for AI | Main risk | Human check |
|---|---|---|---|
| Accounts payable: read invoices, extract supplier, amount, VAT and due date | High | Wrong amount or supplier on a payment | A person approves every invoice before payment |
| Accounts receivable: draft reminder emails, sort customer replies | Medium to high | Wrong tone, or a reminder to a client who already paid | A person sends or approves each message |
| Month-end close: draft accrual notes, flag unusual entries, write variance comments | Medium | A missed entry looks complete | The controller reviews and posts every journal |
| Reporting: draft commentary on approved numbers | Medium | Text that misstates a figure | The CFO signs off the text and the numbers |
| Forecasting: suggest a cash flow forecast from history | Low to medium | False confidence in a number nobody can explain | A person owns the forecast and its assumptions |
Rules beat AI on structured work. If a step is “match this payment to this invoice number”, a rule in your accounting system is cheaper and easier to test. Our AI automation cost guide explains where rules end and a model starts.
How do you pick the first task?
Pick the task with the most repetitive hours and the cheapest mistake. For most small teams, that is invoice capture in accounts payable.
Score each candidate task on four questions:
- How many hours a month does it take today?
- Does it repeat in the same form every week or month?
- Can a person check the output in less time than doing the task?
- If the output is wrong, does a person see it before money moves?
Start with the task that takes the most hours and gets a yes on the other three questions. Do not start with forecasting. It looks impressive, but it is hard to check and easy to trust too much.
Local detail matters here. In Bosnia and Herzegovina, supplier documents can be in Bosnian, Croatian or Serbian, in Latin or Cyrillic script, with amounts in BAM or EUR. Test any document tool on both scripts and both currencies before you trust it. A tool that reads only Latin script will fail on part of your inbox.
Our AI automation service builds this kind of first workflow: it connects the inbox, the extraction step and your accounting system, with a review screen for your team.
What should you measure before you start?
Measure the task for one full month before you change anything. Without a baseline, you cannot show that AI helped.
Before-you-start metrics list
- Hours per month the team spends on the task.
- Volume: invoices, reminders or journals per month.
- Time per item, from arrival to approval.
- Corrections: how many items a person has to fix.
- Close days: working days from month end to final numbers.
- Forecast error: forecast cash against actual cash, per month.
- Exceptions: items the process cannot handle without a person.
Write down who measured each number and how. Then run the AI step for at least two cycles and measure the same items again. If corrections go up, stop and fix the process before you expand.
What tools does a small team already own?
You may not need a new platform. Check what your current licences already include:
- The AI features in your spreadsheet and office suite, if your plan includes them.
- The bank feed, matching rules and document capture in your accounting system.
- A general AI assistant for drafting emails and comments, with company data rules in place.
Use those first. Buy a dedicated finance AI tool only when you can name the task it does better than what you own. The Bloom & Wild case shows that a paid feature can still add nothing.
How do you keep an audit trail and human approval?
Keep the AI step out of the posting and payment path. AI drafts, extracts or flags. A named person approves, posts and pays.
- Store the source document next to the extracted data.
- Log who approved each item and when.
- Keep a list of the prompts and settings you use, with dates.
- Do not paste customer or employee personal data into tools your company has not approved. BiH’s new Law on Personal Data Protection has applied since 4 October 2025 (N1 Sarajevo, 14 October 2025).
One use is a different category. If you use AI to score the creditworthiness of people, the EU AI Act treats that as high-risk (AI Act Service Desk, Annex III). Read our guide to the EU AI Act and credit scoring before you build anything like it.
When does it make sense to bring in outside help?
Bring in help when the first task needs two systems to talk to each other, or when nobody on the team has time to run the test. A tool alone does not solve an integration.
If the first task works, the next step is a plan for the whole company, not just finance. Our AI enablement programme guide covers that. For sector context, see our finance industry page.
Write down your first task, its monthly hours and the systems it touches. Then book a call and we will tell you if it is worth automating.
Sources
- Bain & Company, 2026 CFO Survey press release — 13 April 2026
- ICAEW Insights, Bloom & Wild month-end close — 10 November 2025
- N1 Sarajevo, BiH data law — 14 October 2025
- AI Act Service Desk, Annex III — 28 October 2026
FAQ
Does AI replace the controller’s review?
No. AI can draft notes, extract data and flag unusual entries. The controller still reviews and posts every journal, and owns the numbers. Treat the AI output as a first draft from a junior colleague: useful, fast, and checked before it counts. Put the approval step in writing so nobody skips it in a busy close.
How long does it take to set up a first AI task?
Plan for weeks, not days. In the ICAEW case, Bloom & Wild took about two months to embed an AI close tool and train staff. Add one month before that for your baseline measurement. A simple invoice capture workflow on tools you already own can be quicker, but measure it before you claim a result.
Which finance task should a small team avoid first?
Avoid forecasting and anything that posts or pays without a person. A forecast is hard to check, and a confident wrong number does more harm than no number. Start with a task where a person sees every output before money moves, such as invoice capture or draft reminder emails.
Can AI read invoices in Cyrillic script?
Some tools can, and some cannot. Test before you buy. Collect a sample of real invoices in Latin and Cyrillic script, with amounts in BAM and EUR, and run them through the tool. Count how many fields a person has to correct. Use that count, not the vendor’s claim, to decide.