toca.io
Talk to us

Learn › Guides ·

Supervising AI on volume files after Mazur (2026)

What the Court of Appeal and the SRA now expect from firms that delegate, and how to show it on every file.

By Mat Rule, Founder and CEO, Toca · Last updated

The short answer. Since the Court of Appeal’s ruling in Mazur on 31 March 2026, your team can do litigation work that an authorised person directs, controls and stays responsible for. The SRA’s updated supervision guidance of 12 June 2026 expects you to be able to show that supervision, area by area and file by file. Work done with AI is no different: a solicitor or other authorised person stays accountable, the output gets proper human review, and you keep the evidence. On a volume caseload, the hard part is not the principle. It is the record.

This piece sets out what changed, what “proper supervision” looks like when you run thousands of files, where AI fits, and what to keep. It is a practical summary, not legal advice: read the judgment and the SRA’s guidance, linked at the end, and take advice on your own arrangements.

What changed, in three steps

When What happened What it meant for volume firms
16 September 2025 The High Court (Sheldon J) held in Mazur v Charles Russell Speechlys that an unauthorised employee could not conduct litigation, even under the supervision of an authorised person. Many firms paused or restructured work done by paralegals and case handlers.
31 March 2026 The Court of Appeal allowed CILEX’s appeal ([2026] EWCA Civ 369). “Carrying on” the conduct of litigation turns on direction, control and responsibility, so an authorised person can delegate tasks to unauthorised staff and stay responsible for them. Delegation is lawful again, provided supervision is real.
12 June 2026 The SRA updated its Effective supervision guidance to reflect the ruling, with case studies of good and poor arrangements. Firms are expected to design, record and be able to evidence their supervision arrangements.

Two things did not change. The authorised person is still responsible for the work. And the regulator still expects supervision to be proportionate to the risk of the work and the experience of the person doing it.

What proper supervision looks like on a volume file

The Court of Appeal described the test as proper direction, management, supervision and control. It did not give a complete list of what falls inside or outside the conduct of litigation, and it left the level of oversight to the context and to the regulators. The SRA’s guidance fills some of that gap, and it is clear that there is no single model. For a firm running high volumes of personal injury, credit hire or other claims, the practical reading is:

  • Direction. Each type of task has a written instruction: what the handler may decide, what they must refer, and to whom. The SRA expects firms to record the rationale behind those arrangements.
  • Control. Steps that carry real risk (issuing, serving, settling, anything that commits the client) are gated, so they cannot happen without the right person’s approval.
  • Supervision. A named authorised person oversees each handler’s work, with reviews in proportion to the risk, and some direct discussion, not only a system.
  • Responsibility. For every step, you can say who was responsible and show that they actually exercised it.

On a file or two, a supervisor’s notes and emails will show all of that. Across thousands of files, the evidence has to be produced as the work happens, or it will not exist when you are asked for it.

Where AI fits

The SRA’s guidance treats AI as one more way work gets done under supervision, not as a way round it. An authorised individual keeps ultimate responsibility for legal services delivered with AI’s help, and AI output needs appropriate human review, scrutiny and professional judgement. The guidance asks firms to think about what effective supervision of AI-assisted work looks like in practice.

One question is still open. The Law Society has asked whether an AI system making key decisions could itself amount to conducting litigation, and neither the judgment nor the guidance answers it. Until someone does, the safe design is the obvious one: AI reads, drafts, checks and flags; a person decides anything that matters, and the record shows who.

In practice, that means four things for any AI you use on a file:

  1. The output points to its source. A summary of a medical record or a bank statement should carry the page it came from, so a reviewer checks the document rather than trusting the summary.
  2. Doubt goes to a person. When the AI is unsure, or two models disagree, the finding stops and goes to someone qualified, with the disagreement shown.
  3. The thresholds are yours. Your firm decides which findings may go straight to the file and which always need a review, function by function, and can tighten them at any time.
  4. Every approval is logged. Who approved what, when, on which passage, and which model versions assessed it.

What a reviewer will ask for, and what to keep

A reviewer asks What to keep How Toca records it
Who was responsible for this step? The named authorised person for each task type and file The approver’s name on every approved step
Was the work directed? The written instruction for the task, and when it last changed The workflow that ran the step, versioned
Was the risky step controlled? Evidence the step could not run without approval Approval gates on the steps you choose
What did the AI do, and on what? The source passage behind each finding The page each finding came from
Did a person review the AI’s output? Who reviewed it, and what they decided The approval, the approver and the passage, logged
Can you show it for every file in this batch? An export, not a search of inboxes Exportable for SRA file reviews, costs challenges or your PI broker

A short checklist for a COLP

  • Map your task types against the seven kinds of work the Court of Appeal thought unlikely to be the conduct of litigation (paragraph 193 of the judgment), and against the steps that clearly are.
  • For each task type, write down who may do it, who supervises, and what must be referred. Record why you chose that arrangement.
  • Gate the steps that commit the client or the court.
  • For every AI tool on a file, check it shows its sources, routes doubt to a person and logs approvals.
  • Check you can produce the evidence for a whole batch of files in an afternoon, not a fortnight.
  • Revisit it when the SRA settles the changes it has proposed to its statement of solicitor competence, which cover supervision and AI.

When you don’t need a platform for this

If your caseload is small, or your fee earners run their own files end to end, careful file notes, supervision meetings and a simple log will do. The SRA’s guidance does not require any particular system. Software earns its keep when the volume is high enough that evidence written by hand falls behind the work, and when AI is reading documents at a scale no supervisor could check line by line.

How Toca does it

Toca runs the operational work around a case for volume firms, and its Spectral engine reads the documents on the file. More than one model assesses every finding; where they disagree, the finding stops and goes to a lawyer. Every approval is logged with the source passage, the model versions and who signed it off, and the record can be exported. Toca is certified to ISO/IEC 42001 for AI management and ISO/IEC 27001 for information security. See the governance detail.

Questions

Can unauthorised staff conduct litigation after Mazur?

Yes, if an authorised person directs and controls the work and stays responsible for it. That was the Court of Appeal’s ruling on 31 March 2026, reversing the High Court’s decision of September 2025.

What does the SRA expect firms to show?

That supervision arrangements are designed for the risk, written down with their rationale, and working in practice. The SRA’s guidance says putting arrangements in place is necessary but not enough on its own: firms should be able to evidence them.

Can we use AI on litigation files?

Yes, under the same principles. An authorised person keeps ultimate responsibility, and AI output gets appropriate human review. Whether an AI system making key decisions could itself be conducting litigation is still an open question, so keep the decisions with people.

Does a case management system count as supervision?

Systems and controls can be part of it. The SRA recognises that delegation can be built into a case management system, but it still expects direct discussion between supervisor and supervisee, not a system alone.

Is this the final word?

It is the position as of November 2026. The SRA has also proposed changes to its statement of solicitor competence that cover supervision and AI. Check the latest guidance before relying on any summary, this one included.

Sources

Published on news.toca.io, 11 November 2026. Shown on toca.io, toca.legal.

← Learn