Your documents are full of sentences like that one.
CoKnow reads where decisions actually get made — Slack, Confluence, Google Docs, the transcripts you already record — and proposes the correction when a message contradicts a line in a doc. A person accepts every edit; nothing is written silently.
“Annual plans cannot be cancelled mid-term without exception.”
“Annual plans may be cancelled mid-term at the account manager’s discretion. The unused balance is credited to the account, not refunded.”
“Annual plans may be cancelled mid-term at the account manager’s discretion. The unused balance is credited to the account, not refunded.”
“Annual plans cannot be cancelled mid-term without exception.”
“Confirmed on the call with legal — AMs can cancel an annual mid-term, we credit the remainder, we don’t refund it. The PRD still says otherwise.”
meridian.slack.com/archives/C02QK4T1P/p1709549 →The stale line is never the doc you were thinking of.
“If the nightly job fails, page Tomás Lind.”
“If the nightly job fails, page the #payments-oncall rotation.”
“Tomás left in May. The rotation owns reconciliation now — please stop paging him.”
Slack permalink →“Rate limit: 120 requests per minute, per key.”
“Rate limit: 600 requests per minute, per key, bursting to 1,200.”
“We raised this for every partner when we moved to the new gateway. The guide was never updated.”
Comment permalink →“General availability: 14 April.”
“General availability: 2 June, pending BIN sponsor sign-off.”
“June second is the date we can defend. April was always the sponsor’s best case, not ours.”
Transcript segment →It reads what your team already writes.
It flags contradiction, not resemblance.
120 requests per minute
“we raised it to 600 for every partner”
It proposes, cites, and waits for a person.
Ask it how work actually gets decided here.
Not a prediction of what anyone would say — a report of what people did say, the last dozen times something like this went through.
What would it take to get a partner data-sharing proposal approved here?
Four went to review in the last eighteen months. Three were approved; the fastest took eleven days, the slowest fifty-two.
Legal asked about sub-processor disclosure on all four. On the two that stalled, the delay was the partner’s DPA, not internal review.
Dana Whitfield engaged on all four and asked each time for the retention window in writing before approving.
1 more item matched that you don’t have access to.
The refusals are the design.
“Legal would probably ask about the retention window.”
It never predicts or simulates a person. It reports what someone actually said: “Legal asked about retention on four of the last five proposals — here are the threads.”
“Based on internal discussion, the limit was raised.”
No claim without a permalink. If the evidence can’t be linked to a message, comment, revision or transcript segment, the claim is dropped rather than shown unsourced.
“Your data improves the model for everyone.”
Never trained on your data. Ever — contractually and technically. A separate database per customer, and nothing in it leaves for training.
“Here’s what #exec-comp said about your proposal.”
Answers are filtered per asker, at answer time. Nobody sees evidence from a channel or doc they can’t already open; CoKnow says only how many more items matched, never what they were, where they live or who wrote them.
“12 documents updated automatically this week.”
No silent edits. Today CoKnow writes nothing to your documents at all. When proposed edits ship, a person accepts every change, and CoKnow records who accepted it and when.
“CoKnow has joined the meeting.”
No bot of ours is ever on your call. CoKnow reads only transcripts you already record, and posts a plain-language notice in every channel it joins.
Full history of the channels on your allowlist, public or private, including messages written before it joined. Inviting the bot is not enough on its own: a channel is read only once it is on the allowlist. It cannot read DMs between people; the only direct messages it sees are questions sent to CoKnow itself.
Page bodies, comments, version history and authors — in the pilot spaces on your allowlist only.
Bodies, comments and revision authors, in folders you explicitly share.
Google Meet transcripts your team already records, read as Google Docs from a folder you share. CoKnow never joins a call.
A 45-minute screenshare. One team, not the company.
- Your Slack admin
installs CoKnow as your own internal app from a manifest we provide. It reads history and writes nothing but its own replies and one notice per channel; you can revoke it without us. - A Confluence admin or user
approves a read-only OAuth connection. CoKnow reads only the pilot spaces on its allowlist and checks every page against it. - A team lead
shares specific Drive folders with your workspace’s own service account, as Viewer. No domain-wide delegation and no super admin. - On our side
a separate database per customer, a tested delete-everything path, and a query log of who asked what — available to you.
Only the ones on your allowlist that someone has also invited it to; an invite alone is not enough. It posts a plain-language notice in the channel when it joins, so nobody finds out later. Take a channel off the allowlist and CoKnow leaves it and deletes everything it held from it.
Not between people. It cannot — the scopes your admin installs reach only messages sent to CoKnow itself, so someone can ask it a question by DM, and there is no setting that widens that.
Yes, and the path is tested rather than theoretical: your database is dropped, the index and query log with it, then the backups and credentials, and we send you the list of what was removed.
No. Never. CoKnow reads only the Meet transcripts you already record and share with it, and if you record nothing, it has nothing to read.
Point it at one team’s documents and ask it something you already know.
Forty-five minutes to connect one workspace. Then judge it on whether the sources check out.
Book a demo