Will AI Replace Junior Lawyers or Just Change Billing Models?

What AI already does in law firms, and what it cannot

Document review that once absorbed hundreds of trainee hours now runs in minutes. Contract analysis tools flag risk clauses at scale. Research platforms produce a first draft faster than any associate.

The SRA measured the scale of this in its Risk Outlook report on AI in the legal market, published on 20 November 2023. Three quarters of the largest solicitors’ firms were using AI, nearly double the figure from three years earlier. Over 60 per cent of large firms were at least exploring generative systems, as were a third of small firms. The regulator described use as rising rapidly, and warned that staff may use public tools casually on client matters even where the firm has formally adopted nothing. Those figures are now the better part of three years old and adoption has continued since, so treat them as a floor rather than a current reading.

The distinction that matters is not between legal and non-legal work. It is between tasks where output can be checked against an external source and tasks where it cannot.

What the technology cannot do is carry responsibility. The SRA’s position is that firms should oversee AI much as a solicitor supervises a junior employee: the work can be delegated, the accountability cannot.

Which professional duties does AI use engage?

The existing ones. The SRA has not written a separate AI rulebook, and the duties that apply to AI-assisted work are the duties that apply to all work.

Four are doing most of the work in practice. The obligation to provide a competent service, which now includes understanding the limitations of the tools a firm relies on. The obligation on those who supervise or manage others to ensure the work is properly done, which extends to work produced by a system rather than a person. The duty of confidentiality, which is engaged the moment client information is typed into a platform the firm does not control. And the duty not to mislead the court, which is where fabricated citations land.

Continuing competence is the quieter one. Solicitors declare annually that they have reflected on and addressed their learning needs, and a fee earner relying on a tool whose failure modes they cannot describe has a competence gap rather than a technology gap.

None of this is novel regulation. It is the ordinary framework applied to an unfamiliar delegate, and the SRA’s supervision analogy is the most useful way to hold it: you would not sign out a trainee’s research without checking it, and the tool is less accountable than the trainee.

How will AI affect junior lawyers?

The tasks most exposed to automation are repetitive, high volume and rule based. That is precisely the work trainees have historically learned on, which is why the anxiety concentrates at the junior end of the profession.

The likelier outcome is compression rather than extinction. Junior lawyers are becoming supervisors of machine output earlier in their careers, and that demands more substantive legal knowledge sooner, not less. A trainee who cannot spot that an AI has misread a contractual clause, or invented a precedent, adds nothing to the review. The profession still needs legal expertise. It has simply shortened the runway for demonstrating it.

The training problem nobody has solved

Document review was never valuable because the documents needed reviewing. It was valuable because reviewing five hundred contracts teaches a trainee what a normal contract looks like, which is the only way to recognise an abnormal one quickly.

Remove the volume and the pattern recognition has to be built another way, and no firm has published a convincing answer to how. This matters beyond training quality: qualifying work experience under the SQE requires the opportunity to develop competences from the Statement of Solicitor Competence, and a role reduced to prompting and light checking may not deliver that. Firms signing off qualifying work experience should be satisfied the experience is real.

How firms restructure training for that shift is a leadership problem as much as a technology one, and the firms treating it as an IT procurement decision are the ones most likely to lose their best juniors.

What did Ayinde v Haringey change?

Verification stopped being an abstract skill on 6 June 2025, when the Divisional Court handed down judgment in Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin).

In Ayinde, grounds for judicial review cited five cases that did not exist. In Al-Haroun, 18 of the 45 authorities cited in correspondence and witness statements were fake, drawn from unverified research. Sitting under its Hamid jurisdiction, the court made wasted costs orders of £2,000 plus VAT in Ayinde against both the barrister and the instructing solicitor, and the matters were referred to the relevant regulators.

The court set out the range of consequences available where fictitious material reaches a courtroom: wasted costs, referral to a regulator, and potentially contempt proceedings. For junior lawyers the message lands hardest, because checking AI-generated research against authoritative sources is exactly the work now delegated to them. After Ayinde, that check is a professional duty with judicial teeth, and the personal consequences of skipping it fall on named individuals.

What has happened since

The judgment did not stop the problem. Around fifty fake citation cases have now been reported in England and Wales, and the Upper Tribunal noted a considerable increase through the second half of 2025.

Guidance has accumulated alongside the case law. The judicial AI guidance for judicial office holders was updated in October 2025 and states that legal representatives are responsible for the material they put before the court. The Bar Council issued updated generative AI guidance in November 2025, and the Civil Justice Council published an interim report and consultation on the use of AI in preparing court documents in February 2026.

One later decision widened the risk. In R (Munir) v Secretary of State for the Home Department [2026] UKUT 81, the Upper Tribunal observed that putting client letters and decision letters into an open AI tool can amount to placing that information in the public domain, with consequences for confidentiality and privilege. The full verification and supervision controls are set out in our guide to fake AI citations in court.

Can you put client material into an AI tool?

It depends entirely on the tool, and the answer needs to be decided at firm level rather than by individual fee earners under deadline pressure.

The relevant distinction is contractual, not technical. A closed or enterprise deployment where the provider is bound not to retain or train on the firm’s inputs is a different proposition from a consumer chatbot whose terms permit both. Munir puts the point beyond argument for the second category.

The questions worth asking any vendor, before procurement rather than after an incident:

• Where is the data processed and stored, and under which jurisdiction
• Is client input used to train the underlying model, and can that be contractually excluded
• What is the retention period, and can it be set to zero
• Who at the provider can access inputs, and under what circumstances
• What happens to the firm’s data if the contract ends or the provider is acquired
• Does the arrangement satisfy the firm’s obligations under UK GDPR as well as its professional confidentiality duty

The SRA’s warning about casual use of public tools is the practical risk here. A formal procurement decision protects nothing if a fee earner pastes a witness statement into a free chatbot at nine in the evening because the approved tool is slower.

Do you have to tell clients you used AI?

There is no general regulatory requirement to disclose AI use to clients, but the question is increasingly being answered by clients rather than regulators.

Institutional clients are adding AI provisions to outside counsel guidelines and panel terms, ranging from disclosure obligations to outright prohibitions on particular tools or particular categories of work. A firm that has not read its panel terms recently may already be in breach of a contractual commitment it did not notice.

There is also a pricing dimension. Where a client has agreed an hourly rate on the assumption of human effort, and the work is substantially machine-assisted, the transparency and fairness questions arrive whether or not disclosure is mandatory. Firms that get ahead of this by repricing tend to control the conversation; firms that wait tend to have it during a fee dispute.

Will AI kill the billable hour?

This may prove the larger disruption. When a task that took forty hours takes four, time-based billing becomes difficult to defend to a client who knows the tools exist, and the billable hour had critics enough before AI arrived.

Firms are already experimenting with fixed fees, subscription arrangements and value-based pricing. Each hands the efficiency gain to the client in exchange for predictability, and each forces a firm to measure profitability by matter rather than by hours recorded.

The leverage problem underneath it

The economics are more awkward than the pricing conversation suggests. The traditional firm model depends on leverage: a partner’s time is sold at a high rate, and profitability comes from associates and trainees billing volume work at rates above their cost. Automate the volume work and the pyramid narrows.

That produces a genuine tension for firms committed to training. The commercially rational short-term response is to recruit fewer juniors, and the commercially rational long-term response is not, because the supply of mid-level lawyers in five years depends on the juniors recruited now. Firms are resolving this differently, and it is worth knowing which way a firm has gone before joining it or recommending it.

For junior lawyers the knock-on effect lands on appraisal. If recorded hours stop being the currency, contribution has to be measured another way: outcomes delivered, client feedback, reliability as a supervisor of AI output. Some juniors will find that liberating. Most firms will find it administratively harder than it sounds, because a decade of performance infrastructure is built on the timesheet.

What about professional indemnity insurance?

Worth raising with your broker before renewal rather than after a claim.

An AI-caused error is, for insurance purposes, an error: a negligently prepared document is negligent whether a person or a tool drafted it, and the firm’s cover responds to the firm’s negligence. What is less settled is how insurers will treat firms with no verification controls, and whether proposal forms will begin asking about AI use, tool inventories and supervision arrangements in the way they now ask about cyber controls.

The defensive position is the same one the regulator wants: a written policy, named sign-off, an inventory of approved tools, and evidence that verification actually happens. Firms that can produce those documents will have a straightforward renewal conversation. Firms that cannot may find the conversation harder, and should establish their insurer’s position rather than assume it.

What UK firms should do now

Five decisions are worth taking before the next insurance renewal or tool purchase:

• Adopt a written AI policy. Name the approved tools and the prohibited uses, including casual use of public chatbots on client matters, which the SRA specifically flagged as a risk even in firms that have adopted nothing formally.
• Assign sign-off. Every AI-assisted document needs a named person responsible for verifying it. After Ayinde, that name is where the personal risk sits.
• Decide which tools may touch client material. Open, public systems raise a confidentiality question distinct from accuracy, and staff need a stated position before they are working to a deadline.
• Train juniors to verify, not just to prompt. Checking every citation and quotation against the primary source is the core skill the courts now expect, and it is a better use of training budget than another prompting workshop.
• Reprice before clients ask. Identify the matter types where AI collapses the hours and move them to fixed or capped fees on your own terms, rather than under pressure at a panel review.

Ownership matters as much as content. Marketing owns the tools, IT owns the procurement, and compliance often sees neither until something goes wrong, which is why the COLP needs visibility of what has actually been deployed. Firms reviewing their wider systems alongside AI adoption can start with our guide to what practice management systems must prove in 2026, and can benchmark their current position with our AI Compliance Readiness Score.