AI for accountants: where it actually pays, and where it is theatre

Every accountancy practice in the country is being sold AI right now. Most of the pitches start with the technology, which is the first sign nobody has looked at your practice.

So here is the honest version. There are places in a practice where AI genuinely pays, and they are narrower and less glamorous than the demos suggest. There are places where it is theatre. And there is a third category that matters more than either: problems that look like they need AI and are actually about how the firm makes money.

Where it genuinely pays

The pattern is consistent. AI earns its place where three things meet: high volume, repeated judgement of a low-stakes kind, and a human reviewing the output before it matters.

Transaction coding and record capture. The bookkeeping end has been quietly automating for years and it works. Bank feeds, receipt capture, rules that learn from correction. This is the least exciting item on the list and the most reliable.

Chasing records. In most practices the January bottleneck is not the return. It is the clients who have not sent anything. That is a communications problem with a documentation problem attached: knowing exactly what is outstanding per client, and asking for it repeatedly without it eating a person's week. Drafting, sequencing and tracking that is squarely inside what AI does well.

Onboarding. Engagement letters, professional clearance, identity and AML checks, the initial data request. A lot of it is the same paperwork with different names in it, and a lot of the delay is nobody having time to start it. Note the boundary: the assembly is automatable, the verification decision is not, and the regulatory responsibility never delegates to a tool.

First drafts of advisory work. Turning a set of accounts into a page of talking points before a client meeting. Summarising a sector's position for a review. Drafting the letter you will then rewrite. The value is not the output, it is that the work starts at a draft instead of a blank page.

Searching your own knowledge. Most firms have already answered the question in front of them, for another client, two years ago, in a file nobody can find. Being able to ask your own document estate what the firm has previously advised is a genuinely underrated use, and it gets better the longer the firm has existed.

Notice what all five have in common. None of them changes what the practice sells. They change how much of the week gets spent producing it.

Where it is theatre

Anything carrying professional judgement to the point of signature. A tax position, a going concern view, a disclosure decision. AI can assemble the material. It cannot hold the responsibility, and any pitch that blurs that line is selling you a liability rather than a tool.

A chatbot on the website when the real constraint is that nobody is having the fee conversation. This one is common. It is visible, it demos beautifully, and it moves nothing.

Content generation for a firm that grows by referral. If nearly all of your new work arrives through existing clients and the local network, an AI content engine is solving a demand problem you do not have.

Any tool bought before somebody owned the change. A tool that sits beside the old process rather than replacing it produces two processes and a licence fee. The tools that get quietly abandoned are rarely bad tools.

The question underneath the question

Here is the part the vendors will not raise.

Take two practices with identical revenue and identical headcount. Both feel stretched. Both are considering AI.

In the first, the team is genuinely at capacity. Compliance work is arriving faster than it can be produced, year-end and the January crunch are managed by overtime, and WIP sits because nobody has time to bill it. Capacity is the constraint. AI helps, immediately and measurably, because faster production converts directly into either more work or fewer hours.

In the second, the team is busy but the practice is not capacity-constrained at all. It is realisation-constrained. Work is done and under-billed. Scope creeps and nobody re-quotes. Fees on the long-standing clients have not been properly reviewed in years because the conversation is uncomfortable. There is a tail of small clients absorbing partner time at a fee that never made sense.

Make that second practice twenty per cent faster and you get twenty per cent more unbilled work. The bottom line barely moves, and in six months somebody concludes that AI does not work.

That is the real finding. The AI question sits downstream of a business question, and almost nobody asks the business question first, because the AI question is the one with tools attached to it and a supplier queue behind it.

Three questions before you buy anything

1. If everyone in the practice had ten hours back next week, what would happen to those hours? If the honest answer is that they would absorb into existing work, efficiency is not your constraint. If the answer is a specific piece of advisory work you keep failing to start, you have a case.
2. What proportion of the work you do actually gets billed? Realisation and lock-up tell you more about the health of a practice than utilisation does. If either is poor, fix the commercial process before you accelerate the production one.
3. Which clients do you wish you did not have, and why do you still have them? Client mix is the slowest problem to fix and the one AI touches least. It is also usually where the money is.

If those three come back clean and the constraint really is production volume, buy something. Start with the narrowest possible use, on one process, with one person accountable for it. Broad rollouts of general-purpose tools are how firms end up with licences nobody uses.

What we will and will not claim

We should be straight about our own position. Orange Jelly has not run this work inside an accountancy practice, so nothing above is drawn from experience in your sector. It is drawn from how the mechanism works and from what is verifiably true about the shape of the business.

What we can speak to is our own venue, The Anchor, where the numbers we publish came from changing the commercial decisions rather than the technology. Menu economics grew food revenue by 98% in three months. Search intent grew Google Search visibility by 828%. Changing how bookings were confirmed cut no-shows by 89%. AI was in all of that work, because a two-person business cannot do that volume of research, drafting and analysis otherwise. It was the reason the work was fast enough to be worth doing. It was not the reason the numbers moved.

That distinction is the whole point. AI is a very good multiplier of a decision that was already right.

The short version

AI belongs in an accountancy practice, in specific places, at the volume-and-repetition end, with a human on the other side of it. That is worth doing and it is not particularly hard.

But "should we be using AI" is almost never the question that is actually pressing on the practice. Capacity, realisation, client mix and pricing sit underneath it, and only one of those responds to a tool.

Work out which one is holding you back first. It is a cheaper piece of work than any implementation, and it is usually a conversation rather than a project. Let's talk, whether or not AI turns out to be any part of the answer.