Automating bookkeeping: a step-by-step plan for accountancy firms

As an accountancy firm you automate your bookkeeping in seven steps: map the document flow, take the quick wins such as bank connections and e-invoices, test your scan & recognise on the difficult documents, automate the exceptions, set a review threshold, measure the share of postings that need no correction, and then scale up administration by administration. In practice, common scan & recognise software often gets stuck at around 60 per cent automation; firms that work through this plan in a structured way can move towards processing 99 per cent of the invoice flow automatically with modern AI processing. Those percentages are calculation assumptions, not guarantees: measure them in your own document flow.

This article is a general step-by-step plan; it is not a sales page. Where we describe our own approach, we say so explicitly. For the technology behind automatic processing there is a separate guide to posting invoices automatically, and for the costs a cost analysis of invoice processing.

Why a step-by-step plan and not one big project?

It is tempting to automate the whole bookkeeping process in one go: every administration, every document type, every connection at once. In practice that often fails. Automating the standard flow is not the problem; the exceptions are, and they differ per administration and per client. A step-by-step plan forces you to first know what happens today, then automate the easy work and only after that tackle the difficult work, with a measurement point after every step. That way you know at each step whether the automation really saves work or merely moves it.

The plan in seven steps

Step 1: Map the document flow

Start by counting, not by picking tools. How many purchase invoices, sales invoices, receipts and bank transactions does the firm process per month, and through which routes do they arrive: email, post, a client portal, an upload or a mobile app? Which part already runs automatically, and where are the piles? For each route, note who has to do something before the document reaches the accounting package. Also measure the lead time from arrival to posting. This overview is your baseline: without a baseline you cannot show in step 6 what the automation delivers.

Step 2: Take the quick wins: bank connections and e-invoices

Two things almost always deliver an immediate result. First, bank connections: let bank transactions arrive in the accounting package automatically instead of downloading and importing files. Almost every modern package supports this, and it structurally saves handling per administration. Second, e-invoices in UBL format: there the invoice details are already data inside the file, so recognition is not needed and the chance of reading errors disappears. Note: UBL is a file format, not a delivery channel. So ask separately how those files reach you and whether your software reads them as data instead of treating them as an image. Encourage clients and their suppliers to submit UBL wherever possible; a UBL file almost never fails on reading errors. The substantive check still applies: a UBL invoice can be held back too if there is doubt about the coding or a suspected duplicate submission.

Step 3: Assess your scan & recognise on the difficult documents

Most firms already have scan & recognise software. The question is not whether that software works, but where it stops working. Collect a test set of fifty representative documents, deliberately including the hard cases: poor scans, receipts, unknown suppliers, invoices with several VAT rates and documents that have to be split. For each document, measure whether it is posted correctly without correction. The percentage that gets through is your real rate of automation, and it is almost always lower than the recognition percentage the vendor quotes. How to set up such a test is described in more detail in the guide to posting invoices automatically.

Step 4: Automate the exceptions

This is where the biggest gain sits, because exceptions cost the most time each. What you need is software that does not just read out fields but also understands the accounting context: recognition per invoice line, knowledge of the administration and the supplier, and learning rules so that today's correction improves next month's posting. It matters that those rules can apply separately per firm, per administration and per supplier; the same supplier is sometimes posted differently for two clients. You do not have to replace your existing setup straight away: you can point a specialised solution at the exception pile first and leave the standard flow where it is.

Step 5: Set the review threshold

Posting fully automatically without a safety net is not a goal; a deliberate review threshold is. Agree when a posting may become final automatically, when a draft posting is prepared and when a human has to look. Good criteria are the confidence of the recognition, the invoice amount and the kind of deviation: a new supplier or an unusual amount deserves a look, a recurring monthly invoice does not. Set the threshold per administration, because a large client with its own controls calls for something different from a small administration. And demand of your software that doubt is visible: a parked document with a concrete reason is worth its weight in gold, quietly posting on despite doubt is a risk.

Step 6: Measure the share of postings without correction

Automation you do not measure deteriorates unnoticed. Pick a small number of fixed measurement points and track them month by month: the percentage of documents posted correctly without human correction, the average handling time of an exception, the lead time from arrival to posting, and the number of errors that only surface later, at the VAT return for instance. Compare those figures with the baseline from step 1. If the correction percentage rises, find out why: a supplier who changed their invoice layout, a new client with unusual documents or a rule that has come to apply too broadly.

Step 7: Scale up administration by administration

Then roll the automation out in a controlled way: start with a handful of administrations, preferably a mix of an easy one and a few awkward ones, and expand every week or month as long as the measurement points stay good. Scaling up per administration has two advantages: mistakes stay small and visible, and your team gets used to the new role step by step, from entering to checking. Take your staff along in this deliberately. The work shifts from typing postings to assessing exceptions and helping clients, and for most staff that is an improvement, but only if it is presented and organised that way.

How much of the bookkeeping can you realistically automate?

That differs per firm, but the patterns are predictable. Bank transactions with good matching rules and e-invoices in UBL reach almost full automation. Standard purchase invoices from known suppliers follow after that. The last few per cent sit in receipts, poor scans and composite invoices; that is exactly where the quality of your software makes the difference between getting stuck at around 60 per cent and growing towards 99 per cent. A hundred per cent without any human eye is not a realistic goal, nor a desirable one: the review threshold from step 5 belongs to a professional firm. What is realistic: that checking becomes the exception instead of the day job.

How does Agentancy fit into this plan?

From here on this is about our own approach. Agentancy is built for steps 3 and 4: the documents your current scan & recognise cannot handle. With Tandem you put Agentancy alongside your existing setup and it processes only the exception pile, with recognition per invoice line and learning rules per firm, administration and supplier. With Totaal you replace the scan & recognise layer entirely. In both cases you set the review threshold: automatic posting, draft postings or manual review, configurable per administration. Which accounting packages and scan & recognise tools we support is shown on the integrations page; the rates are on the pricing page. Hosting runs in the Netherlands and documents are stored within the EU.

Testing it comes without obligations: you try Agentancy free for 30 days or 100 documents, without a credit card. That fits step 3 of this plan: take your own difficult documents, not demo material, and measure what gets through without correction. To prepare, read how posting invoices automatically with Agentancy works.

Conclusion

Automating bookkeeping is not a software purchase but a process: knowing what happens today, taking the easy gains, tackling the exceptions, choosing a deliberate review threshold, and then measuring and scaling up. Firms that keep to this order avoid the two classic pitfalls: automation that only takes over the easy work, and automation that quietly makes mistakes. So start at step 1, by counting. The rest follows from there.