Walk into most accountancy or law firms right now and there are two separate conversations about AI happening. One is in the boardroom. The other is in the back office. They are not talking to each other, and that gap is the whole problem.
In the boardroom it is strategy decks, a partner who went to a conference, a line in the business plan about being 'AI-enabled by 2027' and a quiet worry that a competitor is further ahead. It is exciting and it is mostly abstract.
In the back office it is a junior clearing the same anti money laundering checks they cleared last week, by hand, from half-scanned PDFs, at roughly three days a file. That is where AI is actually landing in professional services. Not in the strategy. In the grind.
This is a post about that grind, and about one specific workflow you could put live this quarter that pays for itself. No hype. One real example, the numbers behind it, and an honest account of what it takes to make it work.
The AI conversation is happening in the wrong room
The strategy-deck version of AI asks big questions. Will it replace fee-earners. What is our firm-wide position. Do we need a head of AI. These are not silly questions, but they are the wrong place to start, because they are unanswerable in the abstract and they stall for months while nothing ships.
The useful version asks a small question instead. What is the single most repetitive, rules-based, high-volume task in this firm that a person quietly hates doing. In almost every accountancy and law firm we talk to, the answer is the same: client and matter onboarding, and the AML checks wrapped around it.
"You do not need an AI strategy to start. You need one painful workflow and a way to measure whether the machine did it properly."
Why most firm AI pilots quietly die
Plenty of firms have already tried. They bought a clever tool, ran a demo that looked magical, and six months later nothing is in production. There are a few predictable reasons, and they are worth naming, because the fix for each one is exactly what separates a toy from a tool.
- The demo used clean data. Real client files are messy, half-scanned and full of edge cases the demo never had to handle.
- There was no definition of 'good'. Nobody agreed an acceptable error rate before switching it on, so trust never formed.
- It could not show its working. Regulated work needs an audit trail. A black box that cannot explain a decision is useless to a COLP or an MLRO.
- It lived outside the real systems. If it does not write back into IRIS, CCH or your practice management system, it is just another screen to copy from.
- No owner. The enthusiastic partner moved on and there was nobody whose actual job it was to keep it running.
Hold those five in mind, because the example below is built to answer every one of them.
A real example: client onboarding and AML
Here is the workflow. A new client lands. Someone collects ID, proof of address and company documents, runs the AML and KYC checks, cross-references Companies House, screens for PEPs and sanctions, records the lot for the file, and gets it past a reviewer. In most firms this takes about three days per file, passes through several pairs of hands, and still bounces back when something is missing.
It is the perfect candidate for AI, and not because it is glamorous. Because it is the opposite. It is high volume, rules-based, document-heavy and painfully repetitive. That is precisely the shape of work modern document AI is good at.
What it looks like with AI doing the first pass
- Documents are read and classified automatically, whatever format they arrive in.
- Names, addresses, company numbers and dates are extracted and checked against Companies House and your screening providers.
- Gaps and mismatches are flagged for a human, with the source document cited next to each one.
- Every step is logged, so the file is audit-ready and SAR-ready by default.
- A verified record is written back into your practice system, not stranded in a separate tool.
- Time per file drops from about 3 days to roughly 12 minutes of human attention.
- 98% of files clear first pass straight through, with people handling only the exceptions.
- Live in production in about 6 weeks, not a multi-year programme.
This is our Amlio product, and those are not aspirational figures. In one 240-person accountancy firm, a six-week pilot freed up about 22 hours a week per onboarding team and paid back roughly £120,000 in the first year. The reclaimed capacity did not get cut. It got pointed at advisory work the firm previously could not staff.
See exactly how it works, including the integrations and the audit trail, on the Amlio page, or read the accountancy breakdown with the full case study.
For law firms: the same shift, applied to contracts
If you are a law firm, swap onboarding for contract review and the logic is identical. A partner is the bottleneck on first-pass review of an MSA or a lease. Every clause matters, juniors flag and partners re-read, and the work cannot scale without more partners.
The same document-AI approach triages every clause against your own playbook and gives it a red, amber or green rating in minutes rather than hours. It is not there to replace legal judgement. It is there so the partner spends that judgement on the three clauses that matter, instead of re-reading the forty that do not.
- Clause-by-clause triage in minutes, not days, against your firm's own positions.
- 95% clause-match accuracy, with suggested redrafts you can actually send back.
- In one 140-fee-earner firm, partner review time fell by 64% with zero hallucinations reaching a client.
That is our Vern product. Same principle as the AML example: take the repetitive first pass off a senior person, keep the human on the decision, and log everything.
More on clause triage, the Vern Score and how privilege stays intact on the Vern page, or the legal overview.
The difference between a demo and a deployment
Both examples above only work because of the boring engineering around the AI, not the model itself. This is where most pilots fall down, so it is worth being specific about what actually makes the difference.
"If a vendor cannot tell you their error rate or show you the audit log, you are looking at a demo, not a deployment."
Where to start, and what it costs
You do not commit to a firm-wide transformation. You pick one workflow, prove it on your own files, and decide from there. For most accountancy and law firms that first workflow is AML onboarding or contract triage, because the pain is obvious and the payback is fast.
We run that as a fixed-scope, fixed-fee six-week pilot. At the end you see real numbers on your real data, not a slide. If it does not earn its place, you have lost six weeks and not a transformation budget.
- Pick the one workflow your team complains about most.
- Run a six-week pilot on your own files, with an agreed quality bar.
- Read the eval results, the time saved and the error rate.
- Roll it out, or walk away. No multi-year lock-in either way.
The 6-week AI pilot page has the full scope and price. If you would rather just talk it through, talk to an engineer, not a salesperson.
Who owns it after the first win
Say the pilot works. You have one workflow live and a team that just got a day a week back. The next question is the one that actually decides whether AI becomes a capability or stays a one-off: who owns what comes next.
This is the same gap that kills most pilots, reason five from earlier. Not a lack of strategy. A lack of someone whose job it is to sequence the next workflows, make the build-versus-buy calls, keep the data governance honest and stop ten enthusiastic experiments turning into ten unmanaged risks. In a large organisation that is a full-time technology leader. Most accountancy and law firms cannot justify a full-time CTO or CISO to do it.
That is what fractional leadership is for. A senior technology leader, a day or two a week, who owns the roadmap and is measured on workflows shipped, not slides produced. They are the person making sure the second, third and fourth workflow land as cleanly as the first, and that the whole thing stays inside your regulatory lines. Leadership that ships, not leadership that decks.
More on how that works on the fractional leadership page. It pairs naturally with a pilot: prove one workflow, then bring in the leadership to scale it without the mess.
The firms pulling ahead on AI are not the ones with the best strategy deck. They are the ones who picked one boring, expensive, repetitive workflow and quietly made it ten times faster while everyone else was still debating. Start in the back office. The boardroom conversation gets a lot easier once something is actually working.