The loop
One loop, eight steps, one gate.
AI captures what your company knows, cleans it, and keeps it current. A named person approves it. Nothing goes live without that.
-
01
Capture
Ingest from every source.
-
02
Structure
Normalize the format.
-
03
Clean
Update, deduplicate, reconcile, archive.
-
04
Enrich
Add context and tags.
-
05
Improve
Refine clarity, usability.
-
06
Approve
Human sign-off, audited.
A person holds this step
-
07
Integrate/Deploy
Push live everywhere.
-
08
Reuse
Serve people and AI.
The problem
Where AI projects stall.
An AI system does not go live until a review board signs it off.
Regulated enterprises are deploying AI assistants and agents they cannot prove are answering correctly, and their AI initiatives are stalling at governance review because of it.
The stall has a specific shape. Governance review rarely objects to the model. It objects to the knowledge the AI reads: nobody can say who approved that content, when, or against which source. The Knowledge Loop exists to make that question answerable for every entry, every answer, every time.
The loop, in one paragraph
The Cognatum Knowledge Loop is an eight-step process that runs continuously. AI captures knowledge from every system where it is created, structures it, cleans it, enriches it, and improves it. A human approves it. Cognatum then deploys it to every channel that needs it, and consumption signals feed the next cycle.
Approved records
318
In review
41
Flagged stale
27
Open gaps
37
- Which regions are approved for data residency? 24
- Who approves a retention exception over seven years? 19
- What evidence do we keep for a supervisory review? 17
- Which model versions are approved for regulated answers? 14
- How is a superseded procedure withdrawn from every channel? 11
The shape
Eight steps, four phases.
AI runs seven of the steps. A person holds one, and it is the one that decides what goes live.
- Intake, steps 1 and 2. Knowledge enters from every source and is normalized into a consistent, findable form.
- Refine, step 3. A continuous hygiene pass: update, deduplicate, eliminate contradictions, archive.
- Elevate, steps 4 and 5. Context, tags, relationships, and clarity, driven by how the knowledge performs in use.
- Govern and activate, steps 6 to 8. The human gate, deployment to every channel at once, and the consumption that feeds the next cycle.
Phase one
Knowledge comes in.
Step 1 is Capture. AI connects to the systems where knowledge already gets written: shared drives, chat, tickets, CRM records, wikis, and portals. Everything it finds arrives as a candidate, and a candidate is served to nobody. It exists to be worked on.
Step 2 is Structure. Every capture becomes one kind of thing: a title, a body, a link to where it came from, and an owner. Nobody can sign off on a screenshot in a chat thread, and nobody can trace one afterwards.
Intake is deliberately wide. Capturing broadly costs you nothing, because nothing captured is live. The gates come later, and they are narrow.
The question
Which retention period applies to client trade communications?
Today, without a governed home
- A slide deck from two years ago says five years.
- A policy PDF on a shared drive says six.
- A chat thread says it depends, and links nothing.
- Nobody knows which one is current.
With one approved entry
Six years, the first two in an accessible place, on non-rewriteable media.
Phase two
The work nobody staffed.
Every company agrees the cleanup matters. Almost none has ever funded it, because it competes with everything that has a deadline. The library grows anyway.
Step 3 is Clean, and it holds a phase to itself. It updates, merges, reconciles, and archives, continuously, so that what a person is asked to approve is worth approving.
- Update. AI compares each entry against its sources, and knows which entries derive from which sources. Change a source and every entry that depends on it is flagged and routed for re-check, not just the one somebody happened to notice. AI drafts each revision and queues it for the Approve gate; the previously approved version stays live, with its approval date visible, until a human signs off on the new one.
- Deduplicate. Near-duplicates are merged into one canonical entry. Every surviving duplicate is a chance for an AI system to retrieve the wrong copy, so duplication is treated as a defect, not a nuisance.
- Eliminate contradictions. Disagreements between two entries, and between an entry and its source, are detected rather than waited for.
- Archive. Entries that no longer apply leave the live library. A retired entry stops being served on every channel at once, and its full record, including its approval history, is retained for traceability.
Why this is what makes an answer defensible
It is also the step that makes an AI answer defensible. Retrieval is faithful to whatever it is given. Point it at three versions of one procedure and it will faithfully return one of them, and nobody can say which in advance or defend the choice afterwards.
Conflicts
When two sources disagree.
This is the case a search box cannot handle, and the one that costs most when it goes wrong. Two entries say different things. Both were approved. Both are live.
Cognatum compares entries against each other and against their sources, so a disagreement is detected rather than waited for. What happens next depends on whether the answer is knowable from the record.
- Where the resolution is unambiguous, because one version supersedes the other on a date the record can show, AI reconciles it and queues the result for the Approve gate.
- Where it is a genuine disagreement, the two are held apart and routed to a named person. They are never averaged, blended, or resolved by recency.
- While a flag is open, no contested version is served as approved. The question comes back as a conflict rather than as a confident answer.
Two live answers
One question, both approved
- Entry A: fee waived for retail
- Entry B: fee applies above tier 2
The record settles it
Resolution queued for approval
-
Entry B supersedes A
Effective 2026-05-01 · queued
A real disagreement
Neither is served meanwhile
-
Routed to a named person
Product · flag open
Phase three
Context, clarity, and use.
Step 4 is Enrich. AI adds context, tags, relationships, and who each entry is for. For your people that means finding things. For an assistant it means answering from the entries approved for that audience, rather than from the nearest matching text.
Step 5 is Improve. AI refines wording based on how an entry performs when people use it: what gets reused, what gets abandoned, what makes somebody ask a follow-up. Entries marked as fixed wording are exempt, at every step.
Nothing here publishes itself. Every rewrite queues as a proposed version for the approval gate, and readers keep seeing the last approved one until a person signs off. AI proposes. The gate disposes.
The gate
A person signs it.
Step 6 is Approve. AI prepares the package: the entry, its sources, its history, and exactly what changed since the last approved version. A named person reviews it and signs off.
The decision is recorded with the approver, the date, and the source. Every approved entry carries that record from then on, in a form a reviewer can read at a glance.
The word for that record is provenance: who approved an entry, when, from what source, and at which version.
“I approved this once. I should not be answering the same question eleven times a year.”
Published as the approved answer.
Approver: you · just now · sha256 c4d1…b90
Everywhere at once
Then it goes everywhere.
Step 7 is Integrate and Deploy. The approved entry goes to every connected channel at once, over MCP, the open standard AI assistants use to reach outside data, and a plain API. There is no rollout channel by channel, and no window where the portal and the assistant say different things.
Step 8 is Reuse. The entry serves your people, your applications, your workflows, and your AI systems, with that record attached to every answer. Clearance is checked on every request, so an agent answers within the clearance of the person it acts for, never through a shared service account.
A question nobody has answered
Consumption generates the signals: what was retrieved, what was reused, and where an answer fell short. A question nobody has an approved answer to is not a miss that disappears into a log. It becomes a recorded gap with an owner, routed back to Capture, which is how the library learns what it does not yet cover. Those signals feed Capture, Clean, and Improve in the next cycle. The loop closes and begins again.
Cognatum governs the entry
source · version · approver · permissions
Fixed language
Wording that must not change.
Paraphrase is what a language model does by default. It is also the exact failure that disclosure review, labeling review, and supervisory review exist to catch. A model that helpfully tightens a sentence has changed a regulated statement, and no amount of provenance on the answer repairs that.
So some entries are marked as fixed language. A fixed entry is returned word for word, in full, or it is not returned at all. It is exempt from the Improve step, it cannot be summarised into an answer, and an assistant drawing on it quotes rather than restates.
- Disclosure and risk language that was reviewed as written
- Labeling and indication text, and the claims tied to it
- Contract and policy wording that a counterparty has agreed to
Returned word for word
“These statements have not been evaluated by the Food and Drug Administration. This product is not intended to diagnose, treat, cure, or prevent any disease.”
D-102 · v4
- PDP template Unchanged
- Support macro Unchanged
- AI assistant Refused
What was refused
Asked to send “isn’t meant to” in place of “is not intended to”. Rewording approved text is refused, not corrected.
Where next
The same machinery elsewhere.
The loop leads with the governance-review problem because that is where an unproven answer costs the most. The same machinery extends without modification.
Support
Support
approved entries answer internal and customer-facing questions, instead of whatever an agent happens to find.
Onboarding
Onboarding
new hires learn from the approved version of what the organization knows, not from the first document a search returns.
General knowledge management
General knowledge management
any body of knowledge worth governing can run the loop. The regulated use cases lead; they do not limit.
Common questions
What a reviewer asks.
What happens at the Approve gate?
AI prepares an approval package: the entry, its sources, its version history, and a precise account of what changed since the last approved version. A named approver reviews the package and signs off, and the decision is recorded with the approver's identity, a timestamp, and the source. Nothing is served to any consumer, human or machine, without passing this gate.
How does provenance work on an entry?
Every live entry carries a provenance record: the named approver, the approval timestamp, the source it derives from, and its version history. When an AI system answers from Cognatum, the answer is tied to the specific entry and version it drew from, so the chain runs from the answer back to a dated, attributable approval decision. That chain is what a reviewer or auditor walks.
What happens when two sources conflict?
The Clean step detects contradictions between entries and between an entry and its sources. Where the record resolves the conflict unambiguously, for example a dated supersession, AI reconciles it and queues the result for approval. Where two authoritative sources genuinely disagree, the conflict is flagged for human review, and no contested version is served as approved until a human resolves it.
Can the assistant reword approved language?
Not for entries marked as fixed language. Those are returned word for word, in full, or not returned at all. They are exempt from the Improve step, they cannot be summarised into an answer, and an assistant drawing on one quotes it rather than restating it. Everything else is subject to the normal loop, where a rewrite is a proposal that queues for the Approve gate.
What happens when we get a question nobody has approved an answer to?
It is recorded as a gap rather than absorbed. The assistant says there is no approved answer instead of writing one that sounds right, and the question is routed back to Capture with an owner. Gaps are evidence too: what your library does not cover is a thing a reviewer can ask about, and it is better to hold that list than to find it one bad answer at a time.
How do AI assistants and agents connect?
Over MCP and REST API. An assistant or agent connects to the Cognatum MCP server and retrieves governed entries with provenance attached, scoped to the permissions of the requesting user or system. The same interfaces carry usage signals back into the loop, and write-back to source systems runs over the same connections.
What does an evidence export look like?
For an entry, a set of entries, or a review period, the export assembles the approval records with approver, timestamp, and source, the version history, the deployment record, and any retirement record. It is evidence you supply toward your own obligations under frameworks such as ISO 30401, the EU AI Act, or NIST AI RMF. Cognatum does not confer compliance with any of them; it gives you the trail your reviewers ask for.
How does an entry retire?
The archive sub-step of Clean identifies entries that no longer apply: superseded, expired, or contradicted by an approved successor. A retired entry stops being served on every channel at once, and its full record, including its approval history, is retained. An answer given while the entry was live remains traceable after retirement.
Does an approved entry stay approved when its source changes?
No. When the update sub-step detects a material change in a source, the entry is flagged and a revision routes through Clean and the Approve gate as a new version. Until a human approves the revision or retires the entry, consumers continue to receive the last approved version, with its approval date visible, so the state of the entry is never ambiguous.
How much of the loop do our people have to staff?
Humans concentrate at two points: the Approve gate, and conflicts the Clean step flags when authoritative sources genuinely disagree. AI runs capture, structure, cleaning, enrichment, improvement, deployment, and serving continuously. The division is deliberate: machine effort goes to the upkeep that grows with the library, and human judgement goes to the decisions that carry legal weight.
Knowledge governed. Intelligence everywhere.
See it on your own content, in your own environment.