Automation
Every company has a handful of jobs that nobody wants and that a person still does by hand: the same details typed into two systems, information chased by email, a report assembled from three exports at the end of every month. They are rarely anyone's whole job, which is why they never get fixed, and they are where a surprising amount of a small team's time goes.
Making those jobs disappear is usually the most useful thing I do on a project, and it doesn't need a separate "AI strategy". It comes out of the discovery work: once I've seen how the work actually flows, the repetitive parts are obvious, and most of them can be handled with a proper integration between the systems you already have. Some of them are a good fit for AI, and where that's true I'll say so and build it in.
What it looks like in practice
Incoming enquiries and documents read, sorted and filed to the right place. Information pulled out of PDFs, emails and old spreadsheets into a proper database. Routine questions answered from your own records, with anything unusual passed to a person. A monthly report that assembles itself. A draft of the client update written from the numbers and left for someone to check and send.
Where a person stays in the loop
Anything that goes to a client, moves money or changes a record someone relies on gets a human check before it happens. That is a design rule, not a limitation: the point is to remove the tedious work, not the judgement. It also means nothing fails silently at night.
Your data
For financial and legal firms in particular, the first question is where client data goes. The answer depends on the job, but the principles don't change: data stays in your own systems wherever possible, anything sent to an AI model is the minimum needed and is not used to train it, and I'll set the choices out in writing before anything is built.