Thred

AI adoption in accounting firms: a four-week pilot, step by step

Manoj Tutika · Published 16 November 2026 · Last checked 16 November 2026

Working on this yourself? Talk to Mark, our co-founder. We’re here to solve it with you.

Book a demo

How do you adopt AI in an accounting firm without the risk?

Most AI projects in firms do not fail on the technology but on the approach. Someone tries a tool for an afternoon on a test company, the result is mixed, and the project fades. What works is a short pilot on real clients, with one task, measurements agreed up front and a decision at the end. Four weeks is enough to know whether it works.

What the four-week pilot looks like

An AI pilot in four weeks
WeekWhat you doWhat the agent may doWhat you measure
1Choose clients, connect books, baselineRead onlyHours per client for the task without the agent
2Agent runs the task alongside your teamRead only, report findingsOverlap between agent and team, missed errors
3Your team works from the agent's proposalsBook after approval of each entryTime per client, share approved unchanged
4Write rules, evaluate, decideApproval, plus rules for fixed kinds of entriesTime saved, quality, team view

What do you do in week 1?

Choose ten clients. Take simple and difficult books, clients who book themselves and clients you book for. Choose one task that recurs for every client, takes many hours and has a clear right or wrong. A weekly check on suspense accounts, VAT codes and duplicate entries is a good start.

Take a baseline and note how long that task takes per client today. Agree who leads the pilot and who reviews the findings. Connect the books with read-only access.

Settle the data questions too. Where does the data sit, is there a data processing agreement, who has access? Ask for EU hosting if your clients are in the EU.

What do you do in week 2?

The agent runs the task for all ten clients while your team does the same work as usual. At the end of the week you put the two lists side by side and ask three questions. Which findings did both spot? Which did only the agent find, and were they right? Which did only your team find, and why did the agent miss them?

That last question matters most. An agent often misses something because it does not know a rule of your firm, such as a client that always uses a particular clearing account or a supplier with a standing arrangement. Write those rules down. They become the agent's instructions.

What do you do in week 3?

Now the agent may book, but only after approval of each entry. Your team stops searching and reviews the agent's findings and proposals instead. Measure time per client again. Count how many proposals you approve unchanged, how many you edit and how many you reject. A high share approved unchanged means the agent knows your rules.

What do you do in week 4?

Look at which kinds of entries you always approved unchanged in week three. For those you can set a rule that the agent books them without asking, with a record. Everything you hesitated on stays on approval per entry.

Then line up the numbers. Time per client before and after, errors missed, share approved unchanged, and what your team thought. Decide on one of three outcomes: scale to all clients, add a second task, or stop. Staff shortages are the top challenge for many firms, as ICAEW's 2025 mid-tier report shows, and Wolters Kluwer found in May 2025 that 98% of Dutch accountants say automation adds to their enjoyment of the work. Ask your team whether it felt that way for them.

How do you scale after the pilot?

  1. Switch the task on for every client on the same ledger, with the same instructions.
  2. Choose read-only or approval per entry per client, and keep difficult clients on approval at first.
  3. Add a second task, such as payment matching or the period close.
  4. Write your AI use into your firm manual: which tasks, who reviews, how you record.
  5. Train your team. The EU AI Act requires that staff who work with AI have sufficient AI literacy.

Read AI for bookkeeping firms for the tasks agents already do, and automating compilation engagements for the next step.

What Thred's agents do in a pilot

Thred is built to start exactly this way. You sign up your firm, connect your Exact Online environment and choose read-only or approval per entry per client. New firms start with 100 credits, enough for a pilot on real clients. In one chat across all your clients you ask questions, and the agent finds the right books.

The Bookkeeping Quality agent runs the weekly check per client, with evidence for each finding. The Bookkeeper proposes corrections, payment matches and depreciation as approval cards with the client's name. The Close agent works the period-end checklist when you widen the pilot. You write your firm's rules into the instructions, and for fixed kinds of entries you set a rule that they may go through without asking. Every action can be traced, so your week-four evaluation is ready when you are. Thred runs in the EU (AWS Frankfurt), with a data processing agreement in every contract. See Thred for accounting firms.

Frequently asked questions

How do accounting firms adopt AI?

Start small and measurable with one task, ten clients and four weeks. Let the agent read only at first and compare its findings with your team's. Then switch on approval per entry, measure time and quality, and scale only when the numbers hold.

Which task should an AI pilot start with?

Pick a task that recurs for every client, takes many hours and has a clear right or wrong. A weekly check on suspense accounts, VAT codes and duplicate entries is a good start. Payment matching and booking depreciation come next.

How many clients should a pilot include?

Ten is a good number. It gives enough variety to see patterns, and few enough that you can review every finding. Mix simple and difficult books, and clients who book themselves with clients you book for.

What should you measure in an AI pilot?

Measure time per client for the task, before and during the pilot. Measure quality: how many of the agent's findings were right, how many errors your team found that the agent missed, and how many proposals you approved unchanged. Ask your team what they thought.

What do you need to record when you use AI?

Record which tool you use, for which tasks, what data goes in, who reviews the output and how. Professional guidance on AI asks for human review and documentation, and the EU AI Act requires staff who work with AI to have sufficient AI literacy. An agent that records every action per client makes this easier.

Should clients be told the firm uses AI?

It is sensible to say so, for example in your engagement letter or data processing agreement. Clients want to know where their data sits and who decides. Explain that the agent prepares and your firm reviews.

Sources

  1. Regulation (EU) 2024/1689 (AI Act)
  2. NBA and NOREA, AI guidance for practitioners (Leidraad AI toegepast, June 2026, Dutch)
  3. ICAEW, Evolution of mid-tier accountancy firms 2025
  4. Accountancy Vanmorgen, Vier op vijf accountants vinden werkdruk te hoog (Wolters Kluwer research, 6 August 2025)

Read next

Where Thred comes in

Accounting Firms

Make your firm more efficient with AI, so you have more time for relationships, new customers and advisory work.

Finance, handled.

Book a demo, or request a callback and we will call you.

Book a demo