AI Readiness Audit for an Accounting Practice: What It Covers, What You Get, and When to Skip It

Roughly half the readiness audits I run end with the same recommendation: build nothing this quarter. Fix two things, re-measure in the autumn, and put the budget somewhere it will actually earn.

Which is a poor advert for a paid audit. It is also the best argument for buying one. The alternative costs roughly ten times more: a three-month automation project that dies in month two because nobody ever wrote the booking rules down — and it takes the team's appetite for a second attempt with it.

What follows is the whole process. I have written it for accounting practices and finance teams, because that is where nearly all of the demand sits now. Wholesale or manufacturing? Everything below still applies. Only the document flows change.

What the audit actually answers

Four questions, in this order. Which of your document flows can an agent take over. How much of each. What has to be fixed before it can. And what the whole thing is worth in hours per month. Every other page in the report exists to defend those four answers.

Notice what is not on the list. Nothing about your AI strategy. No survey of the market. An opinion on large language models is free — every newsletter in your inbox has one. What nobody hands you free is a person sitting down with your documents and measuring the straight-through rate before you sign for anything.

The audit runs three weeks and starts at €2,500 (full price list), credited against the implementation if you go ahead with a project it identifies.

Skip the audit if you are below this size

If you are a three-person practice with one ERP and a predictable client portfolio, do not buy an audit from me. You already know where the hours go. You will also confirm it faster, and more honestly, than I would.

Here is the free version. For two weeks, log every document flow in a spreadsheet with four columns: what arrived, which client, how many minutes it took, who touched it. Nothing else — the moment the log needs a fifth column, people stop filling it in. At the end of the fortnight, sort by total minutes. The top three rows are your automation candidates. The rest is noise. That is essentially what my discovery phase produces, minus the measurement in week two.

Then check the arithmetic before you go shopping. In what an accounting AI agent costs I work the break-even through to the end: with a pilot from €1,900 and a monthly subscription, the volume where it repays inside a year sits around 700 documents a month. Below roughly 400 it never really repays at all. An audit does not shift that line by a millimetre. It only tells you more precisely which side of it you are standing on.

Buy the audit when the answer stops being obvious. Several ERPs across the portfolio. Partners who disagree about what to automate first. A vendor quote you cannot evaluate because you have nothing to compare it against. Those are worth €2,500. Where do our hours go is not — a spreadsheet answers that one.

The four things I score

1. Data — what arrives, in what shape, measured not described

I score this on a sample. Never on a description. Every practice I have worked with underestimates how much of its inbound is still unstructured, and the gap is usually a factor of two.

Since the KSeF mandate, purchase invoices arrive as structured XML, and in that channel there is genuinely nothing left to recognise — the data comes in already typed. But XML is never the whole picture. Alongside it sit PDFs mailed straight to the client's accounts address, receipts photographed in a car park, statements from four banks in three formats, and a housing-association charge that has arrived as the same scanned sheet of paper for eleven years. An agent's ceiling is set by the worst channel, not the best one. So the mix is what I measure: what percentage arrives structured, what percentage semi-structured, what percentage as pixels.

Then master data, which is where the unpleasant surprises live. Duplicate contractors carrying the same NIP under three spellings. A chart of accounts that has quietly drifted apart across clients, so the same fuel invoice books to different accounts depending on who set that client up back in 2019. Cost centres filled in by half the team and ignored by the other half. None of it stops an automation project. All of it decides whether calibration takes a week or six — and you should hear which before you buy, not halfway through.

If the phrase "our data is fine" came up in the kickoff, I quote it back in the readout, next to the sample results. It is the single most reliably wrong sentence in this line of work. Nobody says it in bad faith. It is just that nobody has ever counted.

2. Process — what is written down, and what lives in someone's head

The test I use is blunt: could a competent new hire book a fuel invoice for your third-largest client, correctly, without asking anyone? If the answer is no, an agent cannot either. An agent is not smarter than your documentation. It is faster than your typing.

So I map the top flows and mark every rule as written, partially written, or resident in a person. That last category is not a dig at anybody. Every practice has a Marta who knows that this client's transport invoices split across two cost centres, and that the reason has something to do with an inspection back in 2021. But an unwritten rule is an automation blocker with a name attached to it, and the honest version of that finding reads exactly like this: three of your top five flows depend on one person, so the automation project is really a documentation project with software bolted on.

Exceptions get a section of their own, because that is where the hours actually hide. Not in the 700 invoices that book cleanly — in the 40 that do not. I want the taxonomy: how many exception types, how often each one occurs, how long each takes to clear, and who clears it. A vendor who has not asked about your exception queue has not thought about your month-end. Worth remembering when you line the quotes up side by side.

3. Technology — what is reachable through an API, from where

This is the shortest section of the audit. Occasionally it is the one that ends it.

I look at three surfaces. Your ERP — enova365, Optima, Symfonia, wFirma or something bespoke — and specifically whether its API layer is licensed, switched on and reachable, which is not the same thing as existing in the brochure. enova365 is the friendliest of the bunch, which is why its REST layer shows up in so many of my projects; the pattern moves in weeks rather than months. Your banks, where MT940 files and PSD2 access sit at very different levels of convenience and the details matter more than they should — the mechanics are in automating bank-statement reconciliation. And KSeF, where the interesting question after the mandate is no longer whether you can download invoices but how the retrieval is scheduled, deduplicated and reconciled.

Two practical things drive integration cost harder than anything above. Hosting first: an on-premise ERP running on a server under a desk in the back office is a different project from a hosted one, mostly for network reasons rather than technical ones. Permissions second: I want to know whether a read-only technical account is politically possible in your firm. Every deployment I run starts read-only — the agent proposes, a human posts — and if read-only access alone means a three-month battle for approvals, that timeline belongs in the roadmap now, not in a nasty surprise in month two.

4. People — who approves, and who owns the exception queue

Two named individuals, or the project underdelivers. Not roles. Names.

Somebody has to approve entries in the first weeks, while the agent proposes and a human confirms every single one. Somebody has to own the exception queue afterwards, because an unowned queue fills up and quietly becomes a second inbox that everyone resents. When I ask who those two people are and the room looks at each other, that is a finding. A bigger one than any data-quality issue, frankly, because data gets fixed with a script.

I also score capacity against your real calendar. A practice that spends the middle of every month on filings — the 15th, the 20th, the 25th — has roughly one usable week for a rollout, and pretending otherwise is how go-live dates slip by a quarter. Plan the pilot into the quiet window from the start and the problem disappears — and it is worth reading what that quiet window can actually absorb before you schedule anything.

What the score actually means

Three bands, and what each one should make you do.

Green — pilot now. Structured inbound dominates, top flows are documented, the ERP has a reachable API, and two people have their names against approval and exceptions. Skip the rest of the roadmap conversation and start one flow, in draft mode, end to end. This is the case where I say plainly that you overpaid for the audit and should have bought the pilot.

Amber — two to six weeks of preparation first. The usual pattern: the flows are fine, the master data is not, or one high-volume rule exists only in spoken form. The preparation is unglamorous and cheap — deduplicate the contractor list, tag a category nobody has been filling in, write down the four rules Marta carries in her head. Then pilot. Most practices land here, and the preparation is nearly always work they can do themselves, which is why I price it as advice rather than as a project.

Red — do not automate booking yet. Inbound is mostly pixels, the chart of accounts differs per client with no documented reason, or nobody will own the queue. Automate on top of that and you have bought an expensive machine for generating exceptions.

Red does not end the conversation, though, and this is the finding that surprises people most. There is usually something worth building that never touches the books at all — a read-only agent that answers questions from live ERP data instead of posting entries. That is exactly what the AI accounting agent at LUBART Keg Market does: it looks up receivables, ageing, and what a contractor has bought since January, and it cannot change a thing. It earns trust in a week because the risk is structurally zero — and it buys you the time to fix the data underneath.

The three weeks, concretely

Week 1 — discovery. A 90-minute kickoff with you and with whoever actually does the work, three 30-minute interviews, a data inventory filled in asynchronously, and the sample export: about 200 documents per flow, ideally a full month for two or three representative clients. Plus read-only test credentials for the ERP, or an anonymised export if the credentials need more time than the audit has. Total demand on your team: about six hours.

Week 2 — analysis, and this is the part that separates an audit from an opinion. I run your samples through the same pipeline a live agent would use, and count what comes out the other end. How many documents pass with no correction. Where the failures cluster. What the exception taxonomy really looks like against a real month rather than a remembered one. The output is a measured straight-through estimate on your own documents — not the vendor's 95% from a slide, and not my 70% planning assumption either. Yours.

Week 3 — readout. A 90-minute session with the report in hand and a deliberate challenge round, because a number you have not argued with is a number you will never act on. If you can break my estimate, I would much rather it broke in that room than in month two of an implementation.

The four deliverables — and the one that saves the most money

1. Readiness score across the four dimensions, each with the sentence that actually matters: what has to change, and roughly what that costs in time.
2. Opportunity matrix — your flows ranked by hours saved per month against effort, with the measured straight-through rate beside each one rather than a guess.
3. Roadmap for six to twelve months, sequenced so the first project funds the argument for the second.
4. The do-not-automate list. Flows I recommend leaving alone, each with its reason: too variable, too low-volume, or too tied to a judgement call somebody should be making personally.

The last one is the deliverable clients quote back at me a year later. Selling somebody four automations is easy. It is far more useful to tell them that two of the four will produce an exception queue costing more than the work it replaces — and to say it before the invoice rather than after.

The three questions that are not on the scorecard

Who is liable when it books something wrong? You are. Exactly as you are today when a junior books something wrong. That is precisely why every deployment starts in draft mode with human approval, and why the straight-through rate is a number to watch every month rather than a promise you accept once and forget.

Where does client data get processed? A fair question, and the first one a serious practice asks. It belongs in the audit, in writing: a data-processing agreement, named sub-processors, the region where processing happens, and what is retained versus what merely passes through. If a vendor cannot fit that answer on one page, the answer is not encouraging.

What happens during an inspection? Every proposed entry keeps its source document, the rule applied, the model version, and the human who approved it. In practice the audit trail from an automated flow ends up better than the manual one, because nothing has to be reconstructed from memory six months later.

If you want the shorter, more practical version of where to begin before any of this, where to start with AI in an accounting practice covers the first three steps and costs nothing to read.

Starting point

The first step is a free 30-minute call. We walk through your document flows and I tell you honestly which of three things applies: buy the audit, buy the pilot instead, or go and keep that spreadsheet for two weeks and come back to me afterwards. The last answer comes up more often than the first.

Book a 30-minute call →