I am a B2B marketing consultant, not an accountant. That is worth stating at the top, because it is rather the point of this post. If someone whose day job is lead generation and campaign performance can prepare a corporate tax return without it turning into a disaster, the barrier to AI accounting for small businesses is lower than most people assume.
Proximite runs lean. I keep the books myself, updating them between client calls, and I know every transaction in them because I made them.
For FY2024 the corporate tax filing went the usual route: a consultant to prepare it, an external auditor alongside. That path works and there is nothing wrong with it. The thing is that the fee is built for businesses with considerably more moving parts than mine. For a single operator with a modest transaction count, the cost sits well out of proportion to the actual work involved.
This year I wanted two things. To bring that cost back in line, and more importantly, to properly understand my own numbers rather than hand them over and wait for an answer. There is something slightly uncomfortable about not being able to explain your own return.
I also had a professional reason for trying it. Most of my client work involves automating marketing operations, and I had been recommending AI workflows to clients without having tested one on something where being wrong carries a real penalty. Doing my own books was a useful way to find out where the honest limits are.
So for FY2025 I prepped the return myself, with AI doing the heavy lifting on the parts that used to eat a weekend. It went better than I expected. Here is what the process actually looked like, which tools did the work, and how I checked everything before it went near the FTA portal.
Want to set something like this up?
Whether it is your books, your CRM, or anywhere else in your operations, I am happy to talk through what would actually work for your setup.
Email me Know more about meThe stack
Zoho Books holds the books, and has for a few years. Invoices, expenses, bank reconciliation, and a trial balance I can export cleanly. That last part matters more than it sounds. An AI assistant is only as good as the data you can hand it, and software that lets you pull a tidy export is doing half the job before you start.
Claude did the analysis, connected directly to Zoho Books so it could read the ledger, the trial balance and the P&L without me exporting and re-uploading files all day. That connection is what makes the whole thing practical. Asking a follow up question and having it pull the relevant figures itself, rather than hunting for the right export, is the difference between a workable process and an afternoon of file shuffling.
The bank side I handled myself. Statements got entered and reconciled by hand into Zoho before any of the analysis started, so the AI was always reading from books I had already checked. Good discipline regardless of tooling: the analysis layer is only worth as much as the data underneath it.
EmaraTax is where the return actually gets submitted. No AI involvement there, by design. That is a government portal and you type into it yourself, carefully, once you are confident in your numbers.
If the data feels too sensitive
Worth saying plainly, because it is the first objection most people have, and it is a fair one. Financial records are about as sensitive as business data gets, and not everyone will be comfortable connecting their books to a cloud service.
If that is you, the same workflow runs locally. Pull your trial balance and ledger as a file, point a locally hosted model at it, and nothing leaves your own hardware.
One thing matters more here than in most tasks though: use a capable model. This is not the place for the small ones. A 7B or 8B model will happily categorise your expenses and sound completely confident doing it, and a meaningful share of those calls will be wrong in ways you will not spot without checking every line yourself, which defeats the point. Financial data punishes weak models precisely because the output looks reasonable either way. There is no obvious tell.
Run something in the larger open weight classes instead. The 30B and up range handles this properly, and the current generation of open models is genuinely good at structured reasoning over tabular data. You will need hardware to match, which is the real cost of going local, but if you already have a machine with a decent GPU or plenty of unified memory, you are most of the way there. For anyone in a regulated sector, or simply with a preference for keeping financial records in house, this is a proper alternative rather than a compromise.
Either way, read the terms of whatever you use and understand what happens to your data. That is worth ten minutes of your time before you connect anything.
What AI accounting actually handled well
Three things, and they happen to be the three things that used to take up most of the time.
Expense categorisation at volume. A year of transactions where half the descriptions are bank reference codes and merchant strings. Going line by line through that is slow work, and I am noticeably worse at it by hour four than hour one. Handing over the ledger and getting a first pass categorisation back, flagged by confidence, turned a long grind into something I could review rather than produce. That shift from producing to reviewing is the whole benefit in one sentence.
Spotting my own inconsistencies. The same supplier booked to two different expense heads in different months. A software subscription in office costs in January and in marketing in June. Nothing dramatic, just the normal drift of someone doing their own books in gaps between client work. Those are exactly the things that make a return look untidy, and they are genuinely hard to catch by eye. The AI found them in minutes.
Explaining the rules against my actual numbers. There is a real gap between reading guidance on something like Small Business Relief and understanding how it applies to your own position. Being able to ask follow up questions against my own figures, as many as I wanted, got me to the right questions far faster than reading guidance cold. That was where most of the value sat.
How I reviewed it
This is the part that made me comfortable publishing any of this, so it is worth spelling out. The checks below are what turn a fast draft into something you can confidently file.
Reading yes, writing no. The connection was there so the AI could read my books, not change them. Every adjustment to the actual ledger was made by me, by hand, once I agreed with the reasoning. This is worth being deliberate about, because a connected assistant can often write as well as read. Keeping that line clear means you always know exactly what state your books are in, and anything that changed, you changed.
Everything tied back to the trial balance. After the categorisation pass, totals had to match the Zoho trial balance exactly. Not approximately. If the expense total came out even slightly different, something had been dropped or duplicated somewhere. This is the single most useful check in the whole process, and it takes about two minutes. If the numbers tie, you can trust what sits on top of them.
Rule claims got verified against source. Any time something came up about thresholds, relief eligibility, deadlines, or treatment of a specific cost, I went and found it in the FTA’s own published guidance before acting on it. Worth doing as a habit regardless of where the information came from. A confident answer and a correct answer read identically, so checking is the only way to tell them apart.
I spot checked the categorisation myself. Pulled a random sample across the year and worked through them against the bank statements, looking at whether the reasoning held up rather than just whether the label looked plausible. The hit rate was high, which is what gave me confidence in the rest.
I kept a record of judgment calls. Anywhere a decision could reasonably have gone another way, I noted what I decided and why. If a question ever comes back, I can explain my reasoning properly. Good practice whether or not AI is involved.
None of that took long. The review was maybe a quarter of the total time, and it is time I would have spent reading a consultant’s output anyway.
Where my own judgment still mattered
A few things stay firmly with you, and knowing where that line sits is what makes the rest work.
Intent is the big one. The AI can see that a payment went to a supplier. It cannot know whether that was a client cost you rebilled or a straightforward business expense. Only you know that, so those calls need your eyes every time.
Edge cases need a second opinion. Anything unusual in your structure is worth putting to a qualified professional rather than settling with a chat window. A short paid consultation on one specific question is a very different spend from a full engagement, and it is money I would happily put down again.
And responsibility sits with you either way. Your name is on the return. That is not a drawback so much as the thing that shapes how you use the tool: as something that does the preparation brilliantly, while you stay the one making the decisions.
Would I do it again
Without hesitation.
For a single operator with clean books and a modest transaction count, the maths is easy. A professional fee against a weekend of my own time, most of it spent reviewing rather than grinding. That is a trade I will take every year, and I came out of it understanding my own position better than I would have by handing it over.
It would be a different calculation with multiple trading entities, transfer pricing, or free zone questions in the mix. At that point you are buying judgment and professional liability rather than data entry, and neither of those is something you can download. Scale the business up and I would bring a professional back in without hesitating, and would not think of it as money wasted.
The honest framing is this: AI accounting did not replace an accountant. It cleared the hours of ledger work that come before an accountant’s expertise actually starts to matter. When you are small enough that the preparation is most of the job, that turns out to be most of the cost.
The wider lesson I took from it, and the reason I now say this to clients, is that the useful question is never whether AI can do a task. It is which part of the task it should do, and what you check before you act on the output. Get that split right and it works on marketing operations the same way it worked here.
Next year should be shorter still. I am keeping the books current through the year rather than reconstructing in September, which makes the whole exercise easier regardless of what tools I point at it.
If you are thinking about setting something like this up, for your own books or anywhere else in your operations, I am happy to talk it through. You can reach me at reuben.findingcities.in.
I am a B2B marketing consultant, not a tax advisor. This describes what I did for my own business, not advice for yours. If your situation has any complexity in it, talk to someone qualified.
