Judgement

A junior lawyer, accountant or consultant used to spend a large part of the week turning a large pile into a smaller pile.

Read 300 pages of contracts. Reconcile 20,000 transactions. Interview twelve managers. Find the pattern. Draft the memo. Build the slides. Send everything upstairs so somebody more senior can ask why page 47 uses a different shade of blue.

That production layer paid for the pyramid. It also trained the people climbing it.

AI compresses much of this work. A field experiment with 758 Boston Consulting Group consultants found that people using GPT-4 completed suitable tasks more than 25 per cent faster and produced work rated more than 40 per cent better. On a task beyond the model's capabilities, those using AI were 19 per cent less likely to reach the correct answer.

The study captures the operating problem for professional services: much faster production, uneven reliability and a human being who still owns the result.

The technology question is becoming the easy part. Firms now have to redesign their economics, their training and the meaning of professional value.

The middle of the work gets compressed

Law firms can use AI to compare contracts, extract clauses, search case law and produce a first draft. Accounting firms can classify transactions, identify anomalies, read reporting standards and prepare working papers. Consultants can assemble market maps, analyse interviews, build scenarios and draft a respectable presentation before lunch.

These tasks will not vanish in one clean sweep. Some depend on poor data. Some contain facts the system cannot access. Some sit outside the model's competence while looking almost identical to tasks it handles well. The BCG experiment called this a jagged frontier, which is a useful description because the edge moves and it is difficult to see from your desk.

The production layer still shrinks because research, comparison and first-pass drafting require fewer human hours.

This puts pressure on the traditional pyramid: many juniors producing work, fewer managers reviewing it and a small group of partners selling trust at the top.

The new firm will look more like a collection of small expert teams. A senior professional owns the matter. A few people combine subject knowledge, client context and the ability to direct and check AI systems. Specialists join when the risk or complexity demands it. Shared technology, data and quality controls sit behind them.

Large firms will still have advantages. They own deep archives, client relationships, specialist knowledge and the money to build secure systems. Small firms gain something valuable too. A small expert team can take on assignments that once demanded a much larger production group, provided they know where the machine becomes unreliable.

A document is no longer the product

Clients never woke up wanting a thick report. They wanted confidence before an acquisition, a defensible tax position, a contract they could sign or a decision they could explain to the board.

For years, the visible labour around those outcomes helped justify the price. Research took time. Drafting took time. Review took time. A thick report made the effort tangible, even when half of it was a well-dressed appendix.

That logic weakens when the first draft becomes cheap.

The valuable work moves towards framing the problem, testing the evidence, choosing among imperfect options, explaining the trade-offs and accepting responsibility. Relationships matter more because clients can generate a plausible memo themselves. Domain knowledge matters more because a plausible memo can still be wrong. Taste matters because somebody must decide which ten pages deserve to survive.

Professional firms therefore need to sell a decision and the confidence around it. The deliverable may include a document, but the document cannot carry the fee on its own.

The billable hour has a meeting with reality

The billable hour rewards time spent. AI is designed to reduce time spent. Those incentives point in opposite directions.

The American Bar Association's Formal Opinion 512 gives a wonderfully concrete example. If a lawyer spends fifteen minutes using an AI system to draft a pleading, the lawyer can charge for those fifteen minutes and for the time needed to check the result. The lawyer cannot invoice the hours the task might once have taken. The same opinion says fees must remain reasonable even under fixed-fee arrangements.

Rules differ across countries and professions, but the commercial pressure travels easily. A client will not enjoy learning that its adviser completed the work quickly, billed as if the old workflow remained and then added an AI technology fee for the privilege.

Expect more fixed fees, subscriptions, managed services and prices linked to a defined outcome or risk. Routine work becomes cheaper and more standardised. Difficult judgement, urgent access and responsibility command a premium.

Firms also need to show where the saving went. Thomson Reuters surveyed more than 1,800 professionals in 62 countries for its 2026 Future of Professionals report. Among corporate buyers of professional services, 78 per cent said AI-enabled quality improvements were very important or essential. Only 6 per cent said they were receiving them from most or all providers.

Clients have heard the efficiency speech. Now they would like the efficiency.

The apprenticeship problem is serious

Junior work was repetitive for a reason. Reading contract after contract teaches you what an odd clause looks like. Rebuilding cash-flow models teaches you where assumptions hide. Sitting through interviews teaches you the difference between what management says and what the organisation does.

AI can remove the repetition before the junior has developed the pattern recognition needed to check its output.

PwC's 2026 AI Jobs Barometer found that the most AI-exposed junior roles were seven times more likely than the least exposed to demand skills traditionally associated with senior work, including leadership. Employers are asking people to arrive higher up the ladder while removing some of the lower rungs.

Firms cannot solve this with a prompt-writing workshop and a cheerful certificate.

Training has to become deliberate. Juniors need review sessions where an experienced person explains why an AI draft fails. They need simulations, client exposure and responsibility for small decisions with fast feedback. They should learn how to trace a claim to evidence, challenge an assumption and recognise when the tool has crossed its competence boundary.

Senior people will have to teach more visibly. That may feel inefficient in the short term. The alternative is a generation of professionals who can produce polished work at astonishing speed and have no idea whether it is good.

Trust becomes an operating system

Every serious firm needs to know which systems staff use, what data enters them, where outputs are stored and who signs off before work reaches a client, court, regulator or auditor.

The lawyer still owes duties of competence, confidentiality and candour. The accountant still owns professional judgement and scepticism. The consultant still has to tell a client when the evidence is weak, even if the slides look marvellous.

This requires boring but useful infrastructure: approved tools, access controls, source links, review thresholds, testing, incident reporting and a named person responsible for the final answer.

It also creates new work. Clients will need help setting up continuous controls, testing models, monitoring regulations, redesigning processes and checking decisions made by their own AI systems. Some advisory firms will become part software company and part expert service. Others will remain small, human and expensive, called in when the consequences are high.

Both can work. The large undifferentiated middle will struggle.

What I would change now

Map the work at task level. Job titles are too broad. Find where time goes, where errors matter and where client context changes the answer.

Name the accountable reviewer before adding AI. “Human in the loop” on a dashboard means little unless a named person owns the decision.

Rewrite pricing while the savings are still theoretical. Decide how clients share in the gain and how the firm gets paid for judgement, speed and risk.

Rebuild junior development around practice and feedback. Protect the experiences that create good professionals, even when a machine can produce the immediate output faster.

Show clients the improvement. Measure turnaround time, error rates, coverage, decisions improved and hours returned to useful work. Count tool licences as an expense. Measure the result.

Clients will keep paying for somebody who can say, “I checked this, I understand the trade-off and I will put my name on it.” They will become less patient with the hours of formatting that used to appear before that sentence.