A junior fee earner may now produce a draft that looks complete, careful and ready for senior review, partly because GenAI has helped with the research, structure or wording. The result may be fluent and well presented. But that does not prove that the work has been properly checked, understood or supervised.

After Mazur and the SRA’s renewed focus on supervision, this matters. The issue is not only whether GenAI can produce an incorrect answer. The deeper problem is that AI-assisted work can look reliable before the firm can see how it was produced, what sources were checked, what was generated and what was independently verified.

The answer is not to prohibit GenAI by default, nor to rely on a final read-through. The practical answer lies in clear AI governance: the firm’s rules for which tools may be used, for what purpose, with what data, with what source verification and under what level of supervision.

GenAI is not ordinary software

Many firms are used to software that processes information according to defined rules. A spreadsheet may contain errors if the formula or data is wrong, but the user generally understands that the system is applying a set process.

GenAI is different. It can produce fluent, structured and confident answers without those answers necessarily being complete, current, supported or correct. It can sound right without being right.

For legal work, that matters. A well-written paragraph may still rely on a misunderstood source. A good structure may hide an unsupported conclusion. A confident summary may leave out the point that changes the advice.

Juniors need rules, not blame

Junior fee earners should not be blamed for this risk. They are learning legal practice, professional judgement and the firm’s internal standards. It is not reasonable to expect them to design their own safe system for using GenAI, especially when many senior lawyers are still developing their own understanding of the technology.

The risk increases when juniors are left to improvise. Some may use public or unapproved tools too freely. Others may avoid AI completely, even where an approved tool could help them work more efficiently. Others may use AI for research, drafting or polishing without knowing where the boundaries are.

Firms should take responsibility for giving junior fee earners clear rules, approved routes and realistic supervision expectations. That is not only about risk avoidance. It is also about helping juniors produce better work, learn faster and use technology in a way that benefits the firm.

What supervision should look like in practice

Effective supervision of AI-assisted work should not begin when a senior lawyer receives a finished-looking draft. By then, the most important questions may already be hidden: what tool was used, what information was entered, what sources were relied on, what was checked and what was simply accepted.

A practical approach starts earlier.

First, the firm should define the permitted route. This means deciding which tools may be used, which tools are restricted or prohibited, what types of work can be assisted by AI, what information must never be entered into unapproved tools and what level of review is required.

A junior fee earner should not have to decide alone whether client-related information can be placed into a public AI tool. That decision should already be answered by the firm’s AI-use rules.

The rules do not need to ban everything. A blanket prohibition may drive use underground or prevent useful, low-risk applications. A better approach is to distinguish between different uses. Using AI to improve the structure of a non-confidential internal note is not the same as using it to analyse client documents, generate legal authorities or produce client-facing advice.

Second, the task should be classified before AI is used. Is the junior using AI for issue spotting, research support, summarising documents, drafting, checking tone or polishing language? Does the task involve confidential or identifiable client information? Could the output reach a client, a court or a regulator? Does it affect a legal conclusion?

This classification matters because not all AI use carries the same risk. A style improvement task may require one level of control. A legal research task requires another. A draft that will be sent to a client requires more care than an internal brainstorming note.

Third, AI should be used to map the issue before the legal judgement is formed. GenAI can help identify possible questions, organise a research plan, suggest themes or produce a first structure. Its value is in helping the junior see what may need to be checked, not in replacing the checking itself.

Fourth, the source pack should come before the finished draft. This is one of the most important controls.

The unsafe route is to ask AI for a polished answer and then look for sources afterwards to justify it. That reverses the proper order of legal work.

A safer route is to identify possible sources, check that they exist, confirm that they are relevant, verify that they are current, read enough to understand what they actually say, discard weak or doubtful sources and then draft from the verified material.

This is not just “fact-checking”. It is a different workflow. The source base comes first. The draft follows.

Fifth, where appropriate, firms may prefer source-based AI workflows. This means using AI in a way that works from a defined set of reviewed materials, rather than relying only on an open prompt. Some systems can produce answers by referring back to the materials provided, making it easier to see which source supports which part of the output.

This does not remove the risk of error. It does not remove the need for legal judgement. But it can make the work more reviewable.

Sixth, the junior should not deliver only the draft. For AI-assisted research or drafting, the supervisor may require a short review pack with the work. This could include the draft, the sources used, an indication of which source supports each key paragraph or assertion, the approved tool or route used, and confirmation that no confidential client information was entered into an unapproved system.

If a point cannot be verified, it should not be hidden inside confident wording. It should either be removed, checked further or clearly escalated.

Seventh, polishing should come after substance. GenAI can be useful for improving clarity, structure and readability. But firms should avoid polishing uncertainty into confidence. A document should not be made more persuasive before the underlying research and reasoning have been tested.

Finally, the supervisor should review the route to the answer, not only the answer itself. A final read-through may catch tone, structure and obvious errors. It may not reveal whether the junior relied on an invented authority, misunderstood a source, used the wrong tool, entered confidential information or accepted an AI-generated conclusion too quickly.

In AI-assisted work, supervision should test the process: the tool used, the information entered, the source base, the verification performed, the reasoning applied and the evidence of review.

That is what practical AI governance is for. It turns supervision from a final check into a working method.

Conclusion

GenAI can save time and improve the first draft of legal work. But it does not remove the need for professional judgement. In some cases, it makes judgement harder, because incomplete research or weak reasoning can arrive in a fluent and confident form.

The firm’s people should remain in control of the route to the answer, not only the final wording. That means giving junior fee earners clear rules, approved tools, source-based workflows where appropriate and meaningful review.

AI can support legal work. But SRA accountability remains with the firm and its people. The practical challenge is to make AI-assisted work visible enough to be properly supervised.

This is a guest post by Jorge Sánchez, Founder & AI Risk Advisor, Akrivium www.akrivium.com

Akrivium is an independent AI advisory firm for law firms. A former senior software engineer in regulated environments (AXA, Virgin Media O2, University of Cambridge, Public Health England), Jorge helps firms assess AI governance, risk and readiness before they commit budget or reputation.