The last manual year

The accounting software layer is not choosing a model vendor. It is embedding itself into every AI surface a client touches, and the direction of travel is unambiguous: the machine your client asks will increasingly have their actual figures in front of it.

Now put a date against the capability curve. In May 2025, Anthropic's chief executive warned that AI could eliminate half of all entry-level white-collar work within one to five years, naming finance among the most affected fields.2 The prediction was widely read as being about jobs. It is more precise, and more useful, to read it as being about tasks: the rules-based, template-driven, data-processing tasks that sit at the junior end of industry finance and at the revenue core of public practice. Bookkeeping, reconciliations, VAT returns, transaction coding, first-draft accounts, routine tax preparation. In a corporate finance team, automating those is a reshuffle. In a firm that earns most of its fees from them, it is structural.

The profession next door shows the speed. Law got its serious AI agents first, and within roughly eighteen months the leading legal platform reports agents deployed at more than two-thirds of its firms, with a fifth running over fifty agents in production and heavy users saving eleven hours a week.4 Accounting is on the same curve a few months behind: Thomson Reuters' research in June has AI adoption in tax and accounting up from 9% to 41% in a single year.3 Not five years. One.

And there is a final property of this transition that the profession keeps underestimating: nobody will send a warning. The platforms gave no notice before their models started answering cash flow questions. They will give no notice before the day the release note simply reads: it now produces a full set of accounts.

There will be no announcement that the timeline has moved. There will only be the day it has.

An engine takes a trial balance and a general ledger of varying quality (not a pristine cloud file; varying quality is the point), plus whatever supporting documents exist, and produces a complete statutory file under FRS 102. Not just the accounts. The tax computation, the lead schedules at full granularity, the checklists, the analytical review, the points for partner, and a log of every decision taken along the way. The preparation itself takes seconds.7

The revealing design choice is that the process is deliberately slowed to about an hour, because the firm should rightfully stay in control of it (a pace the firm can override if it chooses). The workflow runs in four stages: planning the work (the client's profit intent and the deadlines), then documents in and a completeness check with reconciliations, then preparation (adjustments, tax and the rest) with issues raised and decisions taken along the way, then partner review and sign-off. So the first question it asks is what the accounts are for: profit minimised for tax, or presented for exit, funding or growth. Where judgement is required it does not assume; it presents the options with a recommendation and asks. Over time it learns the firm's own decision patterns and records every deviation for the file.

And the hour that remains is a different kind of hour. The old hour was a cold review, ensuring the firm's procedures and systems were followed; the new one is spent on the client: the records conversation up front, a care call when the accounts are ready, the review itself. The engine prepares the conversation too, because the points for partner and the analytical review amount to a ready-made story of the client's year, sitting there for whoever picks up the phone. So the client's experience of compliance improves at the same time as its cost collapses, and every one of those conversations is a natural doorway to the value work. Compliance stops being the thing that crowds out the relationship and becomes the thing that opens it.

Two more observations matter. First, the architecture is mostly not what people fear. The bulk is automation and machine learning built on compliance workflows that ran real, award-winning firms for years and were refined in practice long before any of it was automated. The build order is the point: perfect the workflow first, then automate it, because automating a poor process only produces the wrong file faster. The language model is invoked solely at decision points, and client data never leaves a locked-down environment. The sound use of AI rather than the maximal use of it. And nothing about the professional obligations moves: the file is prepared under the firm's control with every decision logged, the partner signs off exactly as before, and the resulting trail is fuller than most manual jobs ever kept, which makes for a better conversation with a regulator or a PI insurer, not a worse one. Second, the honest limits are part of why I take it seriously: it does not yet file directly with Companies House or HMRC, and its builders say so plainly rather than promising it is coming soon. It is deliberately agnostic at both ends instead. Coming out, it provides the workflow and the exports, entries included, to drop the finished file into whichever filing and practice software the firm already runs. Going in, it needs no integration with the client's bookkeeping platform, because the reports it works from are easily produced from whatever the client uses.

Treat the numbers with the caution beta numbers deserve. In beta conditions the whole job, review included, runs in around a tenth of the traditional time. Assume the field result is only half that saving and the economics of a compliance job still break in half. And the proof mechanism requires no faith at all, because the same engine can cold-review work a firm has already delivered: hand it ten recently completed jobs and see whether they could have been done better. Every firm can run that test on its own files.

Accounts preparation is one task. Bookkeeping, VAT, payroll processing and routine tax computation are the same shape of task, and the same wave reaches each of them. This is not one clever product; it is what the task-shaped half of the profession's revenue looks like when the tooling matures. The build-out is visible elsewhere too, from agent platforms entering accounts preparation and review in the US to the reconciliation and close automation already in production in industry.3

The question partners ask at this point is why the ledger platforms will not simply do all of this themselves. In time they will do some of it. But their gravity points at the client rather than the firm, and the genuinely hard part was never the accounts. It was the file: the working papers, the judgements, the trail, the way a practice actually gets a job out of the door. That knowledge lives inside firms, not platforms, which is why waiting for your software vendor to hand you this capability is a plan for arriving second.

Run the arithmetic on a typical firm. A £1m practice earning 80% of its revenue from compliance holds £800k of exposure, and fee compression of 20 to 40% on that base is not a margin problem. It is an existential one, which is why the exposure across the profession is so asymmetric: a corporate finance function automating its junior tasks reshuffles a team, while a firm automating its revenue base has to find new revenue.2

The labour market has already priced some of this in. UK graduate vacancies fell by a third in the last twelve months, the large firms are cutting graduate intakes while redesigning entry routes, and 68% of firms expect AI to reduce demand for early-career compliance roles.5 The training pyramid the profession has run for a century (juniors learn by doing the work that is now first in line for automation) is being dismantled by hiring decisions before automation even arrives in most firms.

Set against that a fact from the other side of the ledger: ACCA research finds 74% of firms cannot take on more billable work because they cannot find the staff, and 45% report severe shortages.6 The profession is capacity-starved and automating at the same time. Which means the honest framing of compliance automation is not redundancy. It is the arrival, at last, of the capacity the profession has spent a decade failing to hire. The delivery model changes shape (no junior on the file at all; one senior person spending the hour they already spent on review, with quality control from the moment of upload), and the same arithmetic reaches offshoring, whose premise was always cheaper hours for the same task-shaped work; the providers that thrive from here will be the ones that reposition around review, quality and capacity rather than keystrokes.

One more piece of arithmetic, and it is the one partners should sit with. The firms that move early get a window (one to three years) in which the market still prices compliance at manual rates while their cost of delivery has collapsed. Once automation is universal, competition passes the saving to clients. In between, the margin belongs to whoever moved first. Windows like that come along roughly once a professional generation.

The early movers get one to three years in which the market still pays manual prices for automated work.

The strategic question is what the freed capacity buys, and the answer is sitting in plain sight: the conversation your clients are already paying someone else to have.

Small business owners are taking their financial questions to AI tools that answer confidently with no sight of their actual numbers and no stake in the outcome, and to business coaches at £1,000 or more a month who never see the real figures either. The platforms can increasingly tell a client what their numbers are. They cannot tell the client what to decide, because decisions need context, accountability and someone who will still be in the room next quarter. That gap is the accountant's to take, and the evidence that clients want it taken (and will pay for it) keeps mounting.8

This is where the two halves of this essay join. The seven patterns that erode small businesses are diagnosable in data the accountant already holds; what has always been missing is the time to run the diagnosis and hold the client to the actions.8 Compliance automation is not the threat to that future. It is the funding for it. The hours that come off the compliance job are the hours the advisory relationship has been waiting for, and the firms that route them there will own both sides: the automated file and the human judgement on top of it.

Full disclosure, stated plainly in the notes as well: I am not a neutral observer of this. I have an interest in one of the compliance engines described in Part Two, and Clarity exists to help firms build the advisory layer. I would rather declare the position than pretend this argument fell from the sky. Judge it on whether the logic holds.

1. Price the compression into your plan now

   Model your own book at 20 to 40% compliance fee compression over five years. If the result is survivable, you have a strategy conversation. If it is not, you have a deadline. Either way you have stopped pretending.

2. Stop hiring for the old pyramid

   Every role you recruit this year should be defensible in a world where the file prepares itself: review, judgement, client work. And have the redeployment conversation with your existing team before the rumours have it for you. The honest version is better than the one they are imagining.

3. Pilot with your biggest sceptic

   When you trial automation, do not give it to your enthusiast. Give it to the most accuracy-obsessed person in the building, run one job dual-track (engine and human, same file), and cold-review a batch of recently completed work. If it satisfies your hardest marker, the rest of the firm follows. If it fails, you have lost nothing but a pilot.

4. Route the freed hours into the relationship

   Decide, in advance and in writing, where recovered capacity goes: which clients get a structured review conversation, on what rhythm, delivered by whom. Capacity without allocation evaporates into email. The firms that captured the advisory opportunity planned the redeployment before the hours existed.

5. Move before the announcement

   The pressure point is closer than the consensus thinks; on the legal profession's curve, measured in months. Being early looks expensive right up until the release note lands, at which point it looks like the only position worth holding.