01

AI in legal workflows

On 5 October 2026, the Law Society published a foresight study into agentic AI in the justice system of England and Wales. It examines how systems capable of planning and carrying out several steps could change legal work. One concern is that human oversight may remain formally in place while becoming thin in everyday use.

There are already concrete examples of AI entering legal workflows. In its published account of working with Lawfront, a group of regional law firms, Jylo describes a platform through which teams can build reusable playbooks for document analysis, review and generation. Those tasks include finding clauses, extracting information and flagging risks.

Clifford Chance described another approach when announcing its wider AI rollout in February 2024. Its framework included having AI-generated legal output identified and validated by a qualified lawyer in the relevant jurisdiction.

These examples make the next question quite practical: how does a workflow preserve that lawyer’s approval as a document moves towards release?

02

The approved version and the outgoing version

Consider a hypothetical firm using an AI assistant to prepare a letter of advice. A solicitor checks the sources, revises the wording and approves version seven for a named client, with one agreed attachment.

The assistant then prepares the email. It rewrites a paragraph to make the advice easier to read and adds a background document it thinks will help. The workflow still shows the matter as “approved”.

The solicitor has reviewed one package. The assistant is about to send another.

Even a small edit can change the strength of a recommendation, remove a qualification or make a conditional statement sound certain. An additional attachment can disclose information the reviewer never intended to release. A different recipient changes who receives the advice.

The review has to remain connected to those details throughout the remaining steps.

03

Keep approval connected to the release

Firms already manage document versions, reviews and release permissions. Adding an agent makes it necessary to express those rules clearly enough for software that may continue working after the reviewer has moved on.

A useful approval would identify the document version, the intended recipient, the attachments and the action authorised. Approval to prepare an email could be recorded separately from approval to send it. The firm could also specify any routine formatting changes permitted after review.

Immediately before release, the workflow would check the actual outgoing material against that approval. A substantive edit, new attachment or changed recipient would return for the review required by the firm’s rules. Withdrawing approval would prevent a queued message from being released under the earlier decision.

The reviewer would need a clear account of what changed. Another “Approve” button, presented without the altered paragraph or added attachment, gives the person too little information to make that decision.

04

Test one document workflow

This is where a small trial could produce useful evidence.

A firm could use synthetic client files and a mock email queue. First, test an unchanged, approved letter. Then alter one detail at a time: revise a paragraph, add an attachment, replace the recipient or withdraw the approval while the message is waiting to send.

The trial should establish whether each case follows the firm’s intended process. It should also show how easily a solicitor can understand an exception and whether a reviewer can later reconstruct what happened.

At PF Systems, we would propose testing this at the point where the assistant requests release. PF Kernel could evaluate that request against the recorded scope of approval. PF Core could link the decision to the evidence supplied by the document and email systems.

That would be a proposed evaluation of a specific workflow. Any live integration would need its own testing, including whether every relevant sending route observes the control.

The record should let a reviewer compare the version the solicitor saw with the version accepted by the email system, alongside the recipient and attachments. Delivery and readership would require their own evidence.

The solicitor would still need to assess the advice, check its sources and exercise professional judgement. A release control addresses what happens to that reviewed work afterwards.

For a firm considering a more capable AI assistant, this is a manageable place to begin. Take one document workflow and follow an approval all the way through it.

By the end, the firm should be able to put the reviewed letter beside the sent letter and explain every difference.

05

Sources

Public sources supporting the factual statements in this perspective. Reported statements and company or vendor-reported results are identified in the article.