Skip to content
Live

Hoyes Michalos had a cited answer from its own material the day after kickoff

A knowledge brain over the firm's own material: cited answers, an honest no when the material is silent, and corrections the firm controls.

Client
Hoyes Michalos
Sector
Professional services, mid-sized firm
System
Knowledge brain, Knowledge vault, Custom agents
Live since
Jul 2026
Last measured
15 Oct 2026
The acceptance eval of the Hoyes Michalos knowledge brain, replayed one question at a time: each cell is one real question, filled when the answer was grounded in a matching source, marked when the brain correctly declined, and outlined in amber for the three misses. No question text is shown.

The acceptance eval, replayed in the order it ran. Every cell is one real question. No question text is shown.

  1. Years of guidance sat in public and nobody could ask it a question.

    It was all readable. None of it was askable, so staff asked a colleague instead.

  2. The same question got answered from scratch, by a person, more than once.

The answers were public. Finding them was the job.

Hoyes Michalos 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 its custom work: the prompts, the data, and the pipeline that built the index. The Advizr platform underneath is licensed. If we disappeared tomorrow, the firm would still hold everything it needs to rebuild the brain.

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.

Knowledge systemsDocuments and SOPs feed one index. Every answer arrives with its source attached, and a question the corpus cannot answer is routed to a person with the gap logged.DocumentsSOPsIndexAnswerA personSource attachedNo source, gap logged
The brain's two exits. An answer with its source attached, or a named person and a logged gap.

Beside the brain sits 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, and it has not taken a live call yet.

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.

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 weekly refresh keeps the corpus current and picks up new recordings from the firm's own feed. Two workflows run the ingestion and the recording watcher. For ten weeks this summer the refresh failed almost every run and nobody was told. That is the first thing under what broke, below.

The gold set is how changes are gated. It held 59 questions by the acceptance run of 17 July, each anchored to a real document, and it runs 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: jobs 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.

Graded before launch, used from the first week

The acceptance eval at the top of this page ran on 17 July 2026: 59 questions, graded automatically by our harness. The firm's own grading sheet is still blank, so we call it our measurement, not their sign-off. Every cell in it is one real question, and the three misses stay on the page.

8 of 10
provisioned accounts active in the 30 days to 20 August 2026

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. The brain answered 509 questions from its first answer on 2026-07-10 to that day, 223 of them in the last 30. 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.

The systems keep working whether anyone asks or not. Scheduled jobs, curator passes and the refresh ran 630 times in the 30 days to 15 October 2026.

What we measured

48 of 51answers grounded in a matching sourceSystem loggedby Advizr, 17 Jul 2026
How this was measured

Every answer was graded by Advizr's harness: grounded, cited and drawn from the source that holds the answer. Of the three misses, one cited the wrong source and two declined a question they could have answered. The client's own grading sheet is blank, so this is our measurement, not their sign-off.

Measured at Hoyes Michalos knowledge brain, acceptance eval

8 of 8questions it had no source for, declinedSystem loggedby Advizr, 17 Jul 2026
How this was measured

Eight questions had no source anywhere in the firm's material. The brain declined all eight: five with a plain refusal and three by saying what the material does not hold. An answer nobody can source is a liability, so a refusal is a pass.

Measured at Hoyes Michalos knowledge brain, acceptance eval

0.9 secondsmedian time to an answer in the acceptance evalSystem loggedby Advizr, 17 Jul 2026
How this was measured

Half the answers in the acceptance eval came back faster than this, timed from the question to the finished answer. The slowest took about half a minute.

Measured at Hoyes Michalos knowledge brain, acceptance eval

8 of 10provisioned accounts active in the 30 days to 20 August 2026System loggedby Advizr, 20 Aug 2026
How this was measured

Accounts with activity since 2026-07-21, out of 10 provisioned on the client's domain.

223questions answered in the 30 days to 20 August 2026System loggedby Advizr, 20 Aug 2026
How this was measured

Answer rows written between 2026-07-21 and 2026-08-20. 509 since the first answer on 2026-07-10.

509questions answered since the first answerSystem loggedby Advizr, 20 Aug 2026
How this was measured

All answer rows from 2026-07-10 to 2026-08-20.

630automated runs in the 30 days to 15 October 2026System loggedby Advizr, 15 Oct 2026
How this was measured

Scheduled jobs, curator passes and the corpus refresh, with Advizr staff and test accounts left out.

17 of 788scheduled agent runs that failed, all timeSystem loggedby Advizr, 20 Aug 2026
How this was measured

Most runs are scheduled jobs and curator passes, not a person asking a question.

What broke, and how we know it is fixed

Twenty-two people were invited and none of them could accept.

Why
Every accept link carried the client slug twice, so it redirected to a login wall. Four earlier rounds of fixes missed it because every test built its URL from the database token rather than from the body of the email that was actually delivered. The tests were green and the emails were broken.
The fix
A canonical URL helper, a middleware rescue so links already sitting in inboxes started working, build-time validation of the link shape, a nightly integrity check and a daily probe across the fleet.
How we know
All 22 emails were confirmed delivered, so this was never a delivery fault. The first validated fleet probe came back clean on 5 of 5 clients.

The weekly refresh failed almost every run for ten weeks, and nothing said so.

Why
One archived source document failed to download once, and the job stopped on it at every run after that. A failed run was logged where nobody looked, so the brain kept answering from the material it already had and no one was told.
The fix
The job now sets aside a source that fails, reports it and carries on with the rest (merged 27 September 2026).
How we know
Every weekly run since has succeeded: 27 September, and 4, 11 and 18 October 2026, read from the workflow's run history on 19 October.

What we did not measure

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. When the firm signs a baseline, the first hours number goes here with its date.

Built by

  • Andy ChenEngineer of record

What we would do again

  • If you have a large public body of guidance and staff who still ask the same questions, we would index the public material first. This brain answered inside 1 day 6 hours because nothing had to be retyped.
  • If you expect to correct the system, we would build the correction loop before launch. A validated note that outranks the website is what makes a correction stick.
  • If you want to own the system, we would put the ingest toolkit in your repository on day one, with a runbook your own engineers can run without us.

More stories

Every story

Run your whole company on AI.

Tell us how your business runs. If we're not the right fit, we'll say so on the call.

Your custom strategy and systems, delivered in under a month. Your named people using them every week by day 30. Or we keep working for free until both are true.

The page ends on a hand drawn chart of the sky that draws itself as it comes into view, with a small sail ship crossing its open sky from a far star to a beacon and a small planet.

© 2026 Advizr AI Inc. · Vancouver, BC, Canada

Delivered in under a month. Used by your named people every week by day 30. Or we work free until both are true.