The AI tools most firms have tried so far wait for a person. You type, the tool answers, and nothing more happens until you ask again. That is starting to change, and the AI assistant that works while you sleep has arrived. On 7 July 2026 Anthropic began putting its Claude Cowork assistant on phones and the web, having kept it to the desktop until now, and the way it works points to where these tools are heading. Cowork takes a task and runs with it. It works in the background, carries on when your own device is offline, can be set to start at a chosen time, and sends a question to your phone when it reaches a decision that needs a person. More than nine in ten of the tasks it handles are ordinary office work rather than software, which is the work a law firm runs on. Cowork is one example of a broader shift, not a special case. OpenAI has given ChatGPT an agent mode that works a task through on its own, Google has built the same into Gemini, and Microsoft is putting autonomous actions into Copilot across Word, Excel and the rest of its apps. The legal-specific tools point the same way. Harvey, CoCounsel from Thomson Reuters and Lexis+ AI from LexisNexis now plan and carry out multi-step work such as bulk document review and grounded research. The price of the big-firm products still keeps some of them out of reach for a smaller practice.
What changes when the tool acts on its own
An assistant that waits for each instruction is simple to supervise. You see every prompt and every answer, and the work stops the moment you look away. An assistant that runs on its own is a different proposition. It opens files, moves through a task and produces a result while nobody is watching the screen, sometimes on a schedule set days before. The person who started it carries responsibility for what comes out, yet has not seen the steps in between. That distance between who is accountable and who watched the work is the thing a firm has to close, and it is unfamiliar ground for most practices.
The confidentiality question moves
When a fee earner can start a task from a personal phone, the boundary around client data travels with them. The question is no longer only which tool the firm has approved, but what that tool reaches, where it sends the data while it works. A task running quietly in the background can also be drawing on client files the person has not thought about. An assistant that works offline and syncs across a desk machine and a handset is convenient, and each of those conveniences is also a path for information to move. Before anyone starts, a firm needs a clear answer to two questions. What can the tool see, and what does it keep.
Supervision has to keep pace
The SRA framework already requires you to supervise the work and to account for how it was done. An autonomous assistant does not change that duty, it raises the bar for meeting it. A firm should be able to say who may run these tools and on what kind of matter, and to show how the output is checked before it reaches a client or a court. Tasks that run on a schedule or in the background need a record of what ran and what it produced, so that a file can be reconstructed later if a client, an insurer or a regulator asks. None of this argues against the tools. It argues for setting the rules before the tools set them for you.
What your firm should do
Treat the arrival of these assistants as a question of policy, not of technology. Before staff run autonomous tasks on client work, decide which tools are allowed, what data they can reach, and how a person signs off the result. Keep personal phones out of the client file unless the firm controls what the app can see on them. Put the plain questions to any provider about where data goes and how long it is held, and keep the answers on record where a supervisor can find them. The firms that write these rules now will take the benefit of assistants that do real work with less hand holding. The firms that wait will find their staff have already made the choices for them, on devices the firm never approved and cannot see.
The SRA's compliance tips on AI and technology are a short, useful companion to this piece.
Before an assistant works your files overnight, someone should decide what it reaches. That decision is a short piece of work, and we do it with firms: tell us where you stand.