In the first week of February 2026, Baker McKenzie confirmed it would cut up to 10 percent of its global business services workforce, reported at between 600 and 1,000 roles, and named its growing use of artificial intelligence as one of the reasons. The functions affected were research, know-how, marketing and secretarial support, with dozens of roles in London and Belfast and hundreds more across the firm’s offshore centres. The firm’s own words were that it had been “rethinking the ways in which we work, including through our use of AI”. It was reported as the largest cut attributed to AI in the legal sector to date, and it was not the first. Clifford Chance cut around 10 percent of its UK business services staff in November 2025. Freshfields cited AI when it let paralegals go from its Manchester support hub the September before. Irwin Mitchell removed its litigation assistant roles. Baker McKenzie employs about 4,600 lawyers across 71 offices in 45 countries. Every one of them still practises. The people who checked their work are the ones who left.
What research and know-how staff actually did
Outside the profession, “business services” sounds like the back office: payroll, facilities, the people who order the coffee. Inside a large firm it means something more specific. Research staff found the authority. Know-how lawyers and professional support lawyers kept the precedent bank current, wrote the internal guidance notes, and read the judgment before the partner had to. A citation that did not exist, a clause copied from a superseded template, a point of law that had moved since the last deal: these were caught in that layer, quietly, before the document reached a client or a court.
That layer has not been automated so much as relocated. The firm’s public reasoning is that AI tools now do the searching and the first drafting, so fewer people are needed to do it for the lawyers. Which is true. What is also true is that the lawyer is now the person operating the tool and the person checking its output, and there is no longer a second pair of eyes between the two.
The duty did not move. The safety net did.
A solicitor’s obligations did not change in February. Competence, supervision, candour to the court, confidentiality: every one of them sits exactly where it sat before. What changed is how much of the work behind those obligations is now done by the person who carries them, alone, at speed, with a tool that produces confident prose whether or not the authority is real.
Courts on both sides of the Atlantic have already disciplined lawyers for filings built on citations that did not exist. In the published cases the lawyer relied on an AI tool, did not verify the output, and signed. That was the failure mode when firms still had research teams. It is a fair question what the failure mode looks like when they do not.
The employer has removed the check. The professional’s duty to have checked is unchanged.
None of this is a criticism of Baker McKenzie, or of any firm making the same decision. The economics are not mysterious. A tool that does in seconds what a know-how lawyer did in an afternoon will win that comparison every time, and clients have been asking for the saving. The point is narrower and, for anyone who sets standards for a profession, more uncomfortable: the check has gone, and the duty to have checked has not.
The other answer to the same question
Not every firm read the year the same way. McDermott Will & Schulte, with about 1,700 lawyers, told Law360 in January that its earlier AI training had been videos on how to use ChatGPT, and that in 2026 it intended to do “12 times that amount of training”, including an academy of nine to twelve weeks with a certification at the end, targeted by practice area, and possibly compulsory for associates in time.
Two firms, one year, opposite conclusions. One removed the people who checked and trusted the tool plus the lawyer. The other decided the lawyer needed nine to twelve weeks of structured practice before that trust was earned. Both are answering the question the technology has put to every professional employer: when the checking layer goes, who is competent to replace it? Neither answer has been tested yet. The first is cheaper this year.
What this means for a professional body’s CPD programme
A professional body cannot decide how its members’ employers organise their support functions. It can decide what its members are required to be able to do, and it can decide what its CPD programme teaches and records.
For most bodies, the honest answer today is that the CPD available on AI is awareness. A webinar on what the tools are. A guidance note on confidentiality. A statement that members remain responsible for their work. All of it true, none of it practice. A member who has watched an hour on generative AI has not verified an output under time pressure, has not decided whether a client can be told that a draft was machine-assisted, and has not had a judgement of theirs scored against a standard. That is the skill the research team used to supply, and it is now the member’s own.
Three things follow for whoever owns the CPD programme at a body whose members work in firms like these.
The first is that verification is now a competence to teach and assess, not a value to remind people about. It has a method: knowing what the tool cannot know, checking the thing that matters most first, stopping when the cost of a mistake exceeds the cost of the check. A method can be practised. A reminder cannot.
The second is that the evidence has to be at the level of the individual, not the cohort. When a firm has removed its checking layer, the question a client, an insurer or a regulator will ask is not whether the profession was trained. It is whether this person was, and whether they demonstrated it. Attendance does not answer that. A scored decision does.
The third is that the body is now the only institution in the picture with a reason to hold the line. The employer’s incentive runs one way. The regulator acts after the fact. The body that licenses the professional is the one that can say, before the error reaches the court, what a competent member must be able to do with these tools and prove it.
Where the practice comes from
LearnFrame builds and licenses AI for Professional Practice, a CPD programme for regulated professionals that teaches exactly this: verification, confidentiality, judgement and disclosure, learned by working through the decisions a professional actually faces with an AI tool in the room, and scored so the record shows what each person decided. Module 1 is live and can be completed in about twenty minutes at learnframe.com/course/ai-for-professional-practice/module-1/. A professional body can license it under its own name, for its own members, with the evidence record attached.
If your members’ employers are making Baker McKenzie’s choice, the check they used to rely on is gone. The duty is not.
Want to know what your own programme could tell you that it currently does not? The Programme Design Diagnostic is a fixed-fee, independent read on a single programme: a board-ready diagnosis and a costed build scope, in two to three weeks.
Sources: The Global Legal Post, “Baker McKenzie cites AI as it prepares to cut business services roles”, 6 February 2026, including the firm’s statement, the departments affected and the Clifford Chance and Freshfields cuts. The Lawyer, 6 February 2026, for the up to 10 percent figure. Legal Cheek, 6 February 2026, for the roughly 600 roles estimate. Bloomberg Law, Wake Up Call, 5 and 9 February 2026, for London, Belfast and the offshore centres. Reports of up to 1,000 roles via RollOnFriday, February 2026. Law360 Pulse, “McDermott Plans AI Shakeup With Rigorous Lawyer Training”, 8 January 2026, for the McDermott programme. Baker McKenzie lawyer and office counts from public firm data, 2025. All retrieved 22 September 2026.
For the employer’s layer of AI training and the profession’s duty layer, see Santander (2026): AI tools for all 185,000 staff, mandatory AI training, and a €1 billion target. For an employer that trained first and then required use, see Accenture’s AI upskilling programme (2026): 550,000 staff trained. For how a professional’s judgement can be scored, see One test, five ways of scoring it. See more insights from LearnFrame.