AI and knowledge
What can an AI agent do for a small firm?
Direct answer
An AI agent is useful in a small firm when it does one bounded job against the firm's own material: a first draft from approved answers, a list of what is still missing, a summary of a file a colleague must pick up, or an answer to a staff question from a maintained set of sources. It is not a qualified person, and it should not send advice, file anything, or change a client record without someone accountable in front of it.
The jobs that fit a small team
A law firm can have an agent prepare the document checklist for a new matter type and draft the chasing note, while a lawyer decides the matter is one the firm should open. A CA firm can have an agent read what a client uploaded and mark the obvious gaps against the engagement list, while the accountant decides the working papers are complete. A clinic can have an agent turn a phone note into a structured request the desk confirms.
These are patterns, not claims about a particular client. The shape is the same: the agent reduces reading and retyping inside a process the firm already understands. It does not invent the process.
What it should not be asked to do
Do not ask it to give legal, medical, tax or financial advice and then send that advice. Do not ask it to decide a conflict, a diagnosis, a filing position, or whether a drawing is approved. Do not give it the ability to email a client, move money, or overwrite a file as its default action.
The failure mode is confidence. A fluent paragraph that is wrong costs more than the slow human draft it replaced, because the person who was supposed to check it trusted the tone. Design the workflow so the unchecked answer cannot leave.
Sources, a review step, and a log
The agent should answer from an approved set: the firm's own checklists, the current engagement, the documents on that matter. It should be able to show which source it used. An agent browsing the open internet to answer a client question about this firm is a different, riskier product.
Someone owns the review. The system keeps a log of what was drafted, what was changed, and what was sent. That log is how the firm learns which prompts are safe and which tasks should go back to a person. Without it, the firm cannot tell a useful assistant from a lucky one.
Start narrower than the demo
Pick one repeated task with a visible before-and-after: time to assemble a missing-document list, or the number of times a colleague re-reads the same file. Run it on real examples the firm chooses, including the awkward ones. Expand only when that task is dull and reliable.
Sunbeam builds this kind of assistant as part of a workflow, with the source limits and the review written into the system. A chatbot in a corner of the website, answering from nowhere in particular, is a different and usually worse idea.