Automating administration with AI: where do you start?

Recently I spoke with a tax adviser who showed me how he sets up a new client. He typed the same name, address, and Chamber of Commerce details into six systems: payroll, financial administration, tax software, CRM, invoicing, and bookkeeping. Half a day per new client. And for an entrepreneur with four holding companies he did it four times, because the systems don't talk to each other.

That is not an exception. This is how administration runs at most SMEs: manual work nobody enjoys and that is not allowed to go wrong anywhere. So the right question is not whether AI can take this over, because it can. The question is where you start, and what you leave alone.

What does automating administration mean?

Automating administration means repetitive back-office work no longer happens by hand: processing invoices, retyping data between systems, answering standard emails, compiling reports, drafting documents. Part of that is pure rule-based work that always runs the same way. Classic automation is strong there, often under the label RPA. Another part needs interpretation: which ledger account does this belong to, what is this client actually asking. That is where AI comes in.

The goal is not AI because it is possible. The goal is to win back the hours now spent on work a computer does just as well or better, so there is time left for the work you actually get paid for.

The biggest time saving is rarely where you expect

Ask ten entrepreneurs what they want to automate and nine say "the invoices". Understandable, because that feels the most like drudgery. But the real pain often sits elsewhere: in data that has to go into five systems at once, in the mailbox that has become a second task list, or in the monthly report that takes half a day to assemble.

That is why every engagement starts with measuring, not building. Where do the hours actually go now? Sometimes it is invoice processing. Sometimes it is client data that is connected nowhere. Only once you know that do you know what it is worth to tackle it.

Four processes that automate the fastest

Invoice processing. Not the OCR of the past that misses every other line, but recognition that suggests, based on previous bookings, which account an invoice belongs to, recognises a new supplier, and flags a deviation. You check and approve instead of retype.

Retyping data between systems. The classic example from that tax adviser: one client across six packages. A connection or a bot that records the master data once and pushes it through to the rest saves exactly that half day per client.

Standard correspondence. Payment reminders, quote confirmations, answers to "where are my annual accounts". AI writes the draft in your tone, you review and send. If that kind of mail is a third of your inbox, you win back a whole afternoon per week.

Reports. Bringing figures from different sources into one readable overview. What is now a manual copy-and-paste job then runs in the background and is ready when you need it.

Automating accountancy is a special case

For accountants and bookkeepers there is an extra layer. The work is repetitive and bound by rules: invoice recognition, VAT preparation, Wwft and KYC checks, file review. AI can do the first draft, but the responsibility stays with you, so a review step always belongs in the loop. I wrote about that separately, with the tools that genuinely work in Dutch practice, on the page about automating accountancy with AI.

One important detail: client data is sensitive. Payroll and financial figures don't belong in some random American training set. Work with EU-only tools that don't keep your data and don't use it for training. That is not a precondition, it is condition one.

RPA or AI: what fits when

In short: use classic automation (RPA) for the stable, rule-based steps, and AI for the steps that need interpretation. In practice you often build a hybrid: a bot that handles the structured actions, with AI in the places where judgment is needed. Which tool fits, from Power Automate to a custom pipeline, depends on your existing systems, not on what a vendor happens to sell. More on that is on the page about automating administration with RPA and AI.

What not to start with

Not with "writing an AI strategy". A strategy without a pilot is not a strategy, it is a report in a drawer. Start with one process that demonstrably costs too much time.

Not with your whole administration at once. A bot that crashes the week before the filing deadline is not an annoyance but a business risk. Every automated step should have a manual fallback.

And not with a tool that cannot explain what it does. If a system says an entry belongs to account 4500 but not why, you cannot defend it toward your client, the tax authority, or your accountant.

So how do you start?

Start small and measure. Pick the process where the most hours disappear into manual work now, automate that one piece, and work out what it saves in time before you scale further. If it doesn't work, you stop without a big investment. If it does, you expand to the next process.

Want to know where your biggest time saving sits, without committing to anything yet? The free AI scan gives a first analysis based on your own situation. From there you can start in a focused way, with the process that returns the most rather than the one that shouts the loudest.