The problem #
“Any tool that forces a user to change their process will not be adopted”.
Document drafting is a small component which gets most of the attention. The actual business of law occurs around the enquiry, intake and matter lifecycle.
To test the limits of how an AI system can handle this, we run a complete intake by fully automating our system. Doing so underlines some key patterns, and helps us think about exactly where we can optimally place human guardrails.
Skip to conclusions
Enquiry to closed file in 40 minutes #
The client did own onboarding, the agent did the legal work, and the partner mostly moved the matter along.
In this example, a small company is hiring its first Head of Sales and asking a firm it had never used for an employment contract with a 12-month non-compete.
The client filled in the firm's enquiry form. One link then took her through her instructions, identity checks, the engagement letter and money on account. The agent drafted the engagement letter from the firm's precedent, researched the law and drafted the contract.
The client read the draft in her portal and asked for two changes, which came back as tracked changes. Both sides signed electronically. The partner's captured time became a bill, paid from the money on account, and the matter closed.
The agent pushed back #
Asked for a 12-month non-compete, the agent drafted nine months and said why.
Before drafting, it researched the current law on restrictions for senior sales staff, using UK legislation, government papers and case law. It concluded that twelve months sits at the outer edge of what courts enforce. A shorter restraint, reduced by any garden leave, is safer.
It drafted that, explained the choice in a comment on the clause, and kept the twelve-month non-solicitation the client wanted. The client accepted the recommendation. When she later asked for a new start date and more holiday, the agent changed only those terms.
This is the part of legal work an agent should do well: weigh the instruction against the law and say so. It also made a mistake. Its research described a coming change to unfair-dismissal rights as applying from day one, when it applies after six months of service. That change affects how long the contract's probation should be, and the agent did not flag it. A partner reading the advice should catch an error like that.
Time and the bill #
The bill came to a fraction of the estimate.
It came to £118.50 against an estimate of £900 to £1,400, because only the partner's minutes counted. When an agent does the drafting, the lawyer's time no longer measures the work. A fixed fee fits a matter like this better than the clock.
Was it a real intake? #
It had the shape of one. It would not pass a supervising partner's file review.
| Check | Verdict | What happened |
|---|---|---|
| Conflict check | Missing | Not run before the client was taken on |
| Who the client is | Partly | Checks were made on the director, not the company that was the client |
| Identity | Partly | Uploaded photos, checked by eye |
| Engagement terms | Met | Costs, scope and complaints set out and signed |
| Advice on the key risk | Partly | Given as a comment on the draft, unreviewed by the partner |
| Partner review | Missing | The draft went to the client after a glance |
| Execution | Failed | The signed contract still contained a placeholder |
| Client money | Failed | The matter closed with a balance still held |
We can see that some trivial mistakes slipped through the cracks (a [template] field and payment processing mistake were made). These critical issues should be gated behind deterministic hooks.
The client's side was the strongest part. One link and steps that unlock in order took her through everything without a phone call, and the employee signed without creating an account. The firm's side worked, but was spread across too many screens.
Where people should step in #
The partner touched this matter eleven times without once reviewing anything. A better design has four gates, placed where a mistake is hard to undo or the decision belongs to a lawyer or the client. Automatic checks guard everything in between.
Only the second gate is long: reading the draft and the advice before either reaches the client. That is where the placeholder and the research error would have been caught. The others take minutes, because the work before them has already been checked.
The other side was missing from this run. The employee signed what he was sent; a new Head of Sales would usually take advice on a non-compete. That negotiation belongs after the client has approved the draft and set her limits, and before the final version is settled. Our agent would negotiate within those limits against the employee's agent, and only a request outside them would stop for the partner and the client.
Our Conclusions #
Mechanistically, a small matter can already run end to end, and the agent's legal work was relatively credible. The failures were around it: a final check before signing, client money at the end, a conflict search at the start, and supervision that is real but sparse.
But the key question is, where is this going?
We could see agent-mediated negotiations (multiple rounds of redlines) occurring before a partner ever even looks at a matter.
We could see more artifacts being generated for QA or checking, such as:
- A coverage matrix and branch gallery.
- Approved requirement list with versioning.
- Evidence triage queue and read-back.
- Assembly receipt and regression matters.
- API integration with real gov/regulatory systems.
Legal work is broad, and routine transactional work from a well-established and up-to-date document set is relatively trivial to automate. As the floor continues to rise it is unclear exactly how and where the limits of legal work will lie, but the common pattern is that most of the work that is being eroded is at the lower, rather than the partner tier.
Return to method