Why does the line run through a process rather than between jobs?
Firms usually ask this at the level of whole jobs. Can you automate advice? Should you automate the audit? Put that way the answer is always no, and the conversation stops. Which is a pity, because the useful answer sits one level down.
Almost no professional job is a single act. A piece of advice is a sequence: gather the facts, check them against the file, find the relevant precedent or policy, work out what it means here, decide what to recommend, write it in house style, get it approved, send it, record that it went. Of those steps, two or three carry real judgement. The rest is retrieval, assembly, formatting, filing and chasing.
So when a partner says the work cannot be automated, they are right about the decision and wrong about the nine steps around it. The honest question is narrower: which steps must stay with a person?
What should always stay with a person?
Four kinds of step.
The decision itself. Choosing between two defensible courses of action is what a client is buying. A system can set out the options, the precedent and the consequences. It cannot carry the weight of picking one, because it has nothing to lose if the choice goes badly.
Anything the firm signs and stands behind. The opinion, the certificate, the valuation figure, the recommendation on suitability, the statement that a file is clean. These are not documents, they are undertakings. A named person has to read the thing and mean it. Drafts can be produced. Approvals cannot.
Difficult conversations. Telling a client the deal is off, the claim will not be covered, the inspection turned up something expensive, or that the firm made a mistake. Tone, timing and the ability to answer the next question matter more than the wording.
Deciding whether the inputs are true. A system works faithfully from what it is given. Someone who knows the matter has to notice that the figure in the schedule contradicts what the client said on the call, or that the plan attached is the superseded version. That belongs with the person holding the context, not with a tick box at the end.
Where do firms draw the line in the wrong place?
Usually too generously, and usually around craft. Someone will insist every report is different, every proposal bespoke, every note particular to the matter. Read twenty of them and the content varies while the shape barely moves: the same sections in the same order, the standard paragraphs recurring with small edits, the difference being the facts poured in.
The craft defence protects typing, not thinking, and it usually protects the slowest and most expensive typing in the firm. That pattern runs through the partner bottleneck and the work done twice.
The line gets drawn wrongly in the other direction too. Firms hand over the parts that look mechanical, then treat the output as finished because it reads well. That is the risk in copy and edit: a fluent draft invites a lighter review than it deserves.
Does regulation settle the question for you?
Not in the way people hope. Bodies such as the SRA, the FCA, ICAEW and RICS concern themselves with competence, supervision, client care and who is accountable for the outcome. They do not generally list which software may touch a draft. They do make clear that responsibility sits with a person and cannot be handed to a process.
Read that way, regulation is a design instruction rather than a list of banned tasks. Every output going to a client or a regulator needs a named owner who reviewed it and can explain it. Writing an AI policy covers putting that on paper. Data protection in AI client work handles the separate question of what information should be going anywhere at all.
How do you test a single step?
Take one step of one process and ask four things.
- Is there a right answer, or a chosen answer? Retrieving the correct clause has a right answer. Deciding whether to rely on it does not.
- If this went wrong, would the firm point at a person or at a document? If it is a person, that person stays in the step.
- Does the step need something only the human in the room knows? The client's appetite for risk, the history with the other side, the remark made off the record.
- Would you be content to tell the client exactly how this step is done? If the description makes you wince, the line is in the wrong place.
Steps that fail all four are the ones to rebuild first. That is the ordering logic behind what to automate first and the process priority scorer.
What does this look like in a real process?
Take a lease report in a property practice. Reading the lease, pulling key dates, rent review mechanics and break conditions into the firm's standard schedule, and assembling a first draft in house style, is retrieval and assembly: document review. Deciding what the break condition means for this tenant's plans, and whether to advise them to use it, is judgement, and stays put.
The same split appears in file notes and case notes, where the record can be drafted from the meeting but the characterisation of what the client instructed cannot. Sector by sector the vocabulary changes and the split holds. Law firms and surveying practices both see the same thing once a job is mapped properly: the protected part is smaller than expected and the unprotected part slower than anyone admitted.
So where should a firm start?
Not with a policy about what is forbidden. Map one process at step level and mark each step with who owns it and why. The steps nobody can justify owning are your answer. The steps with a name and a reason beside them are the work that stays human, and they get easier once the rest is cleared away.
That is the point the argument keeps losing. Protecting judgement is not a reason to leave a process alone. It is the reason to rebuild it, so the people paid for judgement spend their week on it. The audit takes about three minutes and shows every assumption it uses.