Privilege, supervision, and a tool that never tires.
AI is in drafting, research and advice. The duties that govern it are the ones you already carry: confidentiality, privilege, and competent supervision of everything going out under the firm's name.
The starting point
Where AI already sits in your work
Drafting and review
First drafts, clause comparison, document review at volume. The judgement about what is missing is the part that cannot be delegated.
Research and analysis
Case and authority research where a confident, fluent answer can cite something that does not say what it appears to say.
Client and matter material
The confidentiality question is what leaves the firm. A reflective record surfaces where people are unsure, before it becomes an incident.
The duty
What actually binds you today
We would rather be precise than alarmist. The duties that bind you now are mostly the ones you already carry, not new AI-specific ones.
Professional conduct rules
Your conduct obligations apply to AI-assisted work as to any other. Responsibility for the advice stays with the firm.
Confidentiality and legal professional privilege
Privilege is not a setting on a tool. What goes in, and what the tool may retain or train on, is a professional question.
Competent supervision
Work going out under your name has to be supervised by someone competent to judge it, which means competent to judge what the tool produced.
And the EU AI Act, stated precisely
Under the EU AI Act the duty sits with the organisation deploying AI, not the vendor supplying it. The Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) rewrote the AI-literacy duty from a duty to ensure literacy into a duty to take measures supporting it, and it carries no standalone fine, though market surveillance authorities have supervised it since 2 August 2026. The Act's sharper duty, the competence of the people exercising human oversight, was deferred to 2 December 2027 and bites where AI is used in a high-risk role. It applies per person, which is the argument for starting now rather than then: a record of judgement counts because it accrues, and it cannot be assembled retrospectively.
Position as at 6 September 2026. This is not legal advice, and a firm evaluating Skilly Work will form its own regulatory view.
The record
What a quarter of it shows you
Your people spend about fifteen minutes a quarter reflecting on real moments of AI use. Each reflection is scored against five observable habits, and the result is a record per person and per team rather than a survey.
Where AI is in use
Which tools across which practice groups and matter types, including the ones outside the firm's licences.
Where the risk sits
Client or matter material entering tools that may retain it, AI-drafted work moving without review, undisclosed use, and over-reliance. Prompts in the legal pack carry a standing instruction not to identify any client or matter.
The supervision gap
Whether the people supervising AI-assisted work can still judge it. The Practise habit asks which skills AI is quietly taking over, trended by seniority so you can see the gap forming rather than infer it later.
What is working
Where AI is genuinely earning its place, and whose practice is worth spreading to the rest of the firm.
Where this stops. Skilly Work evidences how your people use AI. It is not a matter-management, conflicts or document-retention control, and it never asks for client or matter detail. It is also not monitoring software: nothing reads anyone’s screen or files, and the input is a person’s own account of their own work, which is the only reason it is honest.
The applied check
The part that is not self-report
A cycle can open with a piece of AI-drafted work with problems planted in it. Your person marks what they would not let go out unaided, before any hints and before they write anything. Then they see what they caught, what they missed, and why each one matters. It takes a few minutes and it happens once, at the start.
A research note with an authority that does not exist
A fluent note on an enforceability question, carrying an invented case name and neutral citation alongside a settled proposition that is perfectly sound. Catching the first without flagging the second is the whole skill.
An advice note that makes the client's decision
Settle or proceed, at a number, with a percentage attached to the prospects. The question is whether the reviewer sees a decision that belongs to the client rather than a paragraph that needs editing.
A prompt about to go into a public tool
Client named, documents attached, and the request drifting from help with structure to the advice itself. It also contains matter detail that identifies nobody and should stay in.
A covering email that overstates what you did
Not silence about AI, which a firm can settle for itself, but a line saying you reviewed the authorities personally, and a time entry the client will read as the story of the fee.
For the person doing it
They find out where their own eye is sharp and where it is not, on work that looks like theirs, in private. The reveal is written for them and it is the part people say they remember. Nobody is told how they did.
For the reflection that follows
They write about something that just happened rather than something they chose to remember. That is the difference between an account you hope is representative and one anchored to an event you can see.
For whoever asks
“How do you know they actually do this?” stops being answered with an adjective. Scoring is a fixed key, not a model judging anyone: we counted what they found, on a task everyone at that stage saw.
Twelve legal cases, three per habit, so a person meets a different artefact for three cycles running and the support fades as they go. The first has the kinds of problem listed; by the third there are no hints, and there may be several problems or none.
Completion carries stakes. Performance does not. There is no pass mark, no grade and no ranking, and no manager ever sees an individual’s result. Managers see team patterns only. That is a product constraint rather than a setting: the moment a result can be held against someone, the honest answer stops being the safe one, and honest answers are the only thing that makes any of this worth reading. Applied checks are also off until an organisation turns them on.
Start with one team, one quarter
A structured pilot in one team hands you a real evidence pack at the end of it. Try the interactive demo now with no sign-up, or talk to us about shaping a pilot.
Other sectors · See all four