The short answer
Start with a business licence on a language model with a data processing agreement, use the AI features already inside your existing package (Exact, AFAS, Twinfield) before buying anything, and then automate the preparation of case files and correspondence. Judgement, review and client due diligence stay with an employee. That is not caution, that is the requirement.
Where the time actually goes
Not in the bookkeeping itself, because your package has largely handled that for years. The time goes into everything around it: gathering documents, checking whether a file is complete, tracing differences, chasing clients for missing paperwork and typing the same answer for the fiftieth time.
That is exactly the work a controlled workflow can prepare: collect, compare, flag what is missing and put a draft ready. The employee reviews and signs.
The order for a firm
| Order | What | Why here |
|---|---|---|
| 1 | Business licence on a language model, with a data processing agreement | Without that agreement no client data may go in. A free account does not have one |
| 2 | The AI features your package already has | Exact, AFAS and Twinfield build these in themselves. Try them before buying anything |
| 3 | Case-file preparation and document checks | This is the largest recurring time cost, and the steps are easy to record |
| 4 | Correspondence as a draft | The same question, the same fifty answers. Never send automatically |
Anti-money-laundering work: what is allowed and what is not
A workflow may prepare client due diligence: gather documents, compare them against a fixed list, make differences visible and prepare a file with source references. What a workflow may not do is determine the risk, accept the client or file a report. That is and stays a human judgement, and that judgement has to be demonstrable.
In practice that means recording every step, a stop rule for deviations, and a log showing afterwards who saw what and when. That is the same requirement as with any other control. See the Wwft workflow audit and the supervision figures.
What not to do
- Putting client data into a free or personal AI account. No data processing agreement, no legal basis, no story for the supervisor.
- Letting a model decide whether a client carries elevated risk. That is a judgement with consequences and it belongs to a person.
- Sending correspondence automatically. A draft is a gain; a sent mistake is a case file.
- Buying a tool for a process you have not written down yet.
Frequently asked questions
May I put client data into ChatGPT?
Only with a business licence where you have signed a data processing agreement, training on your data is off and you know where the data is stored. With a free or personal account the answer is no.
Can AI take over client due diligence?
No. A workflow may collect, compare and prepare. Determining the risk, accepting a client and filing a report stay a judgement by an employee, and that judgement has to be demonstrable.
Does this work with Exact, AFAS or Twinfield?
Yes, and check first what those packages already offer. What falls outside them are the steps between mail, documents and the package; that is where I build the workflow.
What does this deliver for a firm of ten people?
That depends on your volume and your current time per file. Calculate it with the net process value tool, or start with the free process diagnosis; then it is on paper before you spend anything.
