What these assistants do well
Start with the strength, because it is real. An assistant embedded in the file sits close to the data. It can answer questions about the matter in front of you, summarise a long attendance note or a chain of correspondence, draft a paragraph or an email from the open document, and surface information that would otherwise mean scrolling through the file. For a fee earner working a single matter, that is time saved on the spot, and the integration is clean because the vendor built both halves. These features are a sensible baseline and most firms will benefit from turning them on.
The line they draw
The line they tend to draw is around scope. As a general characteristic, an in-file assistant works with the matter you have open. Its frame of reference is the document or the file in front of the user, not the whole caseload at once. That design makes sense for the job it does, and it is not a fault. It does mean the assistant is answering rather than acting. It responds to what you ask about the file you are looking at. It is not scanning every open matter overnight to find the ones where a limitation date is approaching, a client has gone unanswered for three weeks, or a step in the workflow has quietly been missed. The work that depends on reading across the entire caseload sits outside what a file-scoped assistant is shaped to do.
Two further points deserve care, and both should be checked against the specific product rather than assumed. The first is sourcing. When an assistant answers a question of law or procedure, a practitioner needs to know where the answer came from, because an answer without a traceable source cannot be relied on and cannot be checked by the fee earner who signs it off. Vendors differ in how far their tools cite to an authority you can open and read. Ask to see it on a real question before you form a view. The second is supervision. A tool that drafts or answers still needs a qualified person to approve the output, and firms differ in how deliberately that sign-off is built into the workflow rather than left to individual habit. Neither point is a criticism of any named product. Both are questions you should put to any vendor, including us.
Assistant or agent
The distinction that matters for a buyer is between an assistant and an agent. An assistant waits to be asked and works within the file you show it. An agent works across the whole caseload on a standing basis, initiates routine tasks, and presents them for a fee earner to check and release. The first improves the minute in front of a single user. The second changes what the firm can keep on top of. A firm can run both, and many will. The in-file assistant your CMS vendor sells is a good reason to switch it on, and a poor reason to conclude the question of AI is settled.
This is the space the Bracton AI Assistant is built for. It reads across the caseload rather than one open matter and prepares routine work. Every research answer cites a source you can open, every output is held for a qualified fee earner to sign off before it leaves the firm, and the audit trail records the lot. Hosting is in the UK and your client data is not used to train a model. The assistant is built into Bracton, the case and client management platform, and a firm free to change system buys that. A firm tied into a contract with another provider takes the assistant on its own, running inside the system you already have through the same API the vendor exposes, and what it pays for the assistant is credited against the platform in full if it moves within twelve months of starting. On that route it works alongside the assistant your vendor built rather than replacing it, and does the part a file-scoped tool is not shaped to do.
The fair way to judge any of this is to watch it work, and the demonstration runs the platform on an invented firm you can inspect at your own pace. The judgement that binds anyone happens on your own money, in writing: book a client account review.