How do you set up AI search across your firm's internal documents?

How much time does looking for things cost a firm?

Every firm has a version of the same scene. A fee earner knows the firm advised on something similar two years ago, or wrote a report on a comparable ecology survey, or drafted a clause that dealt with exactly this point. They spend twenty minutes in the document management system, then message a colleague, who stops what they are doing to look as well. Sometimes the document turns up. Often it does not, and the work starts from a blank page.

Searching is invisible in timesheets because nobody records it as a task. To size it, ask a handful of people to note for one week each time they went looking for an existing document or answer, and roughly how long it took. Multiply the weekly hours by the number of people who search like that, by their charge-out rate, and by 46 working weeks. Count the colleague who was interrupted too, because every "do you remember where we put..." costs two people. The unbillable hours calculator turns those weekly hours into an annual figure, and the audit shows how searching compares with the firm's other unbilled work.

Why does the firm keep doing the same work twice?

Because knowledge is filed by where the work was done, not by what it was about. Advice sits inside a matter folder named after a client. A useful calculation note lives in a project directory nobody outside that team would think to open. The best version of a procedure is on someone's desktop. The firm owns the answer, but the answer is only reachable by the person who remembers where it is.

So when that person is busy, on leave or gone, the firm pays to produce the answer again. That is the double work here: not rewriting a report, but re-researching a question the firm settled years ago, and sometimes reaching a slightly different answer the second time.

What does working AI search over internal documents look like?

A fee earner types a question in plain English, such as "have we advised on overage clauses in a residential sale" or "which flood risk assumptions did we use in our last three reports for this council", and gets a short answer drawn from the firm's own material. Beneath it sit links to the documents the answer came from, so they can read the original before using it.

Getting there takes decisions, not just software:

• Which collections count. Final advice, signed-off reports, approved precedents and current procedures go in. Drafts and superseded versions are excluded or clearly marked, so a stale document cannot pass for the current position.
• Who sees what. Answers respect the permissions already on your files, including any walls between clients or matters.
• How it stays current. New documents are picked up as they are filed, so the search reflects this week's work, not a snapshot from the day it was set up.
• Where it lives. It sits inside tools the team already opens every day, rather than a separate portal people forget exists.

Why doesn't a Copilot licence already do this?

A general assistant can only be as useful as the material it is pointed at. Left to search everything it can reach, it treats a half-finished draft and a final signed report as equally valid, and it has no idea which precedent the firm now prefers. Many firms bought licences, tried a few questions, got muddled answers and quietly stopped. The missing piece was never the tool; it was the work of deciding what the firm's authoritative knowledge is. A tool is not a process covers that pattern, and what Copilot is actually good for is a fair assessment of where it helps.

What changes for the team?

Junior staff get answers without having to interrupt someone senior, which speeds up their work and spares the senior person a steady trickle of small questions. New starters reach the firm's accumulated know-how in their first week instead of their first year. Partners spend less time acting as the firm's memory. And because each answer points to its source, people build the habit of checking the original rather than trusting a summary, which is the habit you want on client work.

Which firms is this built for?

Firms whose value sits in a large body of past work. Law firms hold years of know-how, precedents and advice notes. Engineering consultancies and environmental consultancies keep design notes, survey findings and reports that answer tomorrow's questions. Accountancy firms answer the same technical queries repeatedly across clients. Insurance brokers compare policy wordings and past placements. If the firm's procedures are part of what people search for, pairing this with standard operating procedures makes both more useful.

How do we get from scattered files to working search in 30 days?

We begin by collecting the questions people really ask: from the team, from the messages that fly around asking where something is, and from the partners who get asked most. Those questions decide which document collections matter and what counts as the final version. The second and third weeks index those collections, set the permissions, and test the answers against questions the team already knows the answers to, so accuracy is judged by the people who will rely on it. In the fourth week the team is shown how to ask well and how to check a source, and the search goes into daily use. We support it for 30 days afterwards. How we work gives the detail.