Situation
The answers were public. Finding them was the job.
The firm has spent years putting its guidance in public. A public website. Training material. Years of recorded talks. The legislation its field runs on. Anyone could read all of it. Almost nobody could find the right passage while a client was on the phone.
So the question went to a senior person instead. The same questions, week after week, answered from memory by the people whose time costs the most. A new hire asked a partner because asking beat searching. The material was all there. Nobody had a map to it.
What the firm wanted was narrow and hard. A system that answers only from the firm's own material, shows its source every time, and says so when the answer is not there. The principal put the last part in writing. The honesty gate, and an abstain message that names a person to ask, are both in the statement of work.
One more condition came with the brief. The firm owns everything: the code, the prompts, the data, and the pipeline that built the index. If we disappeared tomorrow, the brain would keep running and the firm's own engineers could rebuild it.
What we built
A brain that cites, refuses, and takes correction
The brain is a retrieval system over the firm's corpus: the site, the training material, the recorded talks, the legislation. 12 source feeds in total, each with its own ingest path. Ask a question in plain language and it searches the corpus, reads the passages it finds, and writes an answer with the source attached. The first rule is citation. Every answer carries one the reader can open.
The second rule is refusal. If the corpus does not hold the answer, the brain says so and names the person to ask instead. That behaviour cannot change without going back to the contract. A logged gap costs one question. A confident guess in this field could cost a client.
The third piece is the correction loop. When the firm's principal finds an error, the correction is written as a validated note, locked and pinned. From then on the note outranks the firm's own website on that question. The firm's chief-of-staff agent can propose a note. A person approves it. Nothing enters the trusted layer without that approval.
Two more surfaces sit beside the brain. The firm's engineers build an internal records tool of their own. We added a search surface over it, with sanitised query handling, masked snippets, and deletion restricted to the record's owner or an admin. And we built a voice intake lane for the phones: a retrieval-only route that can reach public document classes and nothing else. A validated internal note never crosses the phone line. The lane is built and proven in testing.
Around all of it run 7 custom agents built for the firm's own workflows, with a human approval step on anything that writes.
The ingest toolkit lives in the firm's own repository with a runbook. The firm can rebuild its brain without us, which was the point.
How it runs
Ask, read the source, correct, repeat
A person types a question in the portal. The brain searches, reads, and streams an answer, with sources shown as soon as they are known rather than at the end. Topic chips let the asker scope a question to one subject area. Every answer has a copy link and a thumbs up or down with a reason attached. Threads persist and can be pinned.
A nightly refresh keeps the corpus current and finds new recordings off the firm's own feed. It raises an alert when it fails. Two workflows run the ingestion and the recording watcher.
The gold set is how changes are gated. 49 real questions, each anchored to a real document title, run against the live portal before a change ships. When a question fails, we fix the document rather than the retrieval thresholds. Write the conclusion in the asker's words. It is cheaper, and it survives the next ranking change.
Behind the person sit the scheduled runs: missions that re-read the corpus and curator passes that review answers. 17 of 788 of those runs have failed since launch. We count the failures because a run that fails silently is how a brain goes stale.
What we measured
Adoption, usage, and how fast it went live
The firm provisioned 10 accounts on its domain. 8 of them were active in the 30 days to 2026-08-20, read from the master user table. Adoption is the number we watch first, because a brain nobody asks is a brain that failed.
The brain has answered 509 questions since its first answer on 2026-07-10, 223 of them in the last 30 days. Both counts come from the tenant's answer table, not from a dashboard estimate.
The repository was created on 2026-07-08. The first cited answer was written 1 day 6 hours later. The public material was already there, and the index was built from it.
What we did not measure: partner and staff hours before and after. No timesheet baseline was recorded before launch, so an hours-back number would have nothing to sit against. The first number is due October 2026, against a baseline the firm signs first.
We also do not print the acceptance score here. The gold set runs before every change, but the runs are not yet written to a ledger in the tenant database, and we publish only what we can read back.