Resources / Blog
Governed knowledge, in writing.
Written for the people who have to defend an AI answer. Every post cites its primary sources, and lists them at the end.
The most recent piece.
A decision log keeps the option you rejected
A decision log is the one record that keeps the options an organization considered and refused. A policy states what applies and says nothing about what was ruled out, so a retrieval assistant reads that silence as permission.
System of record ยท 6 min read
Read the piece →Everything else.
Everything else we have published.
Compliance officers, regulatory affairs leads, quality teams, and the engineers who build for them.
Financial services
A representation letter asserts what isn't there
A management representation letter is mostly a list of things management swears are not there. Retrieval answers from what it can reach, so an empty result and an empty world look identical. That gap is where the signature is exposed.
Read →Standards
Nobody runs a first article on an AI answer
A first article inspection verifies one part against every requirement before a production run is released. The assumption underneath it is that nothing changes. An AI answer breaks that assumption every day, and no control catches it.
Read →Life sciences
Design history files have no page for an answer
A design history file records the date, the revision, the participants and the result behind every design decision. An AI answer that restates the same decision records none of those. It joins the chain nowhere.
Read →AI governance
Your data dictionary stops at the column
Data dictionaries exist because one word can mean four things. They cover columns in a database. The policies, contracts and procedures an AI assistant actually reads carry the same ambiguity, with nothing pinning it down.
Read →Financial services
MNPI rules follow the person, not the answer
An information barrier is keyed to a named person and a named document. A generated answer is neither, so the one record that makes assembled research defensible, the trail of what was read, never gets written.
Read →Standards
Job aids are approved. An AI answer is not.
Every quality system treats the one page extract as a document: numbered, reviewed, dated and tied to the procedure it was cut from. An AI answer does the same job at the point of work, and enters no register at all.
Read →Life sciences
Common technical document: no module for an answer
The CTD gives every claim about a medicine a named place, down to whether a section may hold one file or several. An AI answer about the same medicine has no section, no sequence and no operation. It sits outside the structure.
Read →System of record
Offboarding revokes access, not authorship
The laptop comes back and the accounts close the same week. Sixty documents still carry the leaver's name as approver, and your AI assistant reads them every day without knowing.
Read →Financial services
Chief compliance officer can't certify an answer
A chief executive certifies once a year that the firm has processes to establish, maintain, review, test and modify its written procedures. An answer assembled on demand has no text to maintain and no version to review.
Read →Standards
Metrology measures doubt. An AI answer hides it.
Metrology spent centuries working out how to trust a number: an unbroken chain of calibrations, and a stated doubt on every result. An AI answer about your own documents has neither of those things.
Read →Life sciences
Trial master file: no artifact for an AI answer
The trial master file exists so anyone can judge how a trial was run from the record alone. An AI answer that shapes a trial decision is filed nowhere, and the TMF Standard Model has no artifact for it.
Read →AI governance
Your AI answer has no chain of custody
A forensics team can name every person who held a piece of evidence. Your assistant cannot name anyone who held the knowledge behind this morning's answer. Custody is solved for samples, timber and build artifacts. Nobody wrote a form for knowledge.
Read →System of record
Runbooks are tested. Answers are not.
Runbooks are the knowledge base nobody calls one. Their authority comes from an owner, a field test and a version number. An assistant answering from them delivers the steps and none of the three.
Read →Financial services
Compliance monitoring can't sample an answer
Compliance monitoring needs a population to sample and a benchmark to test against. An assistant answering from your own documents leaves no record of what it read, and no version of the procedure it read it from.
Read →Standards
Work instructions: nobody is trained on an answer
Work instructions are the most tightly controlled documents in a quality system: numbered, revised, signed, and tied to a training record. AI answers do the same job with none of that.
Read →Life sciences
Electronic batch records: answers have no master
An electronic batch record works because an approved master record defines the batch before it runs. Nothing defines an AI answer in advance, so there is no exception for anyone to review.
Read →AI governance
Your model didn't drift. Your documents did.
Model drift measures the model. A retrieval assistant's wrong answer usually comes from a superseded document instead, and no distribution test can see that.
Read →System of record
Lessons learned stop at the register
A lessons learned register is a written statement that something you already approved is wrong. It sits in a log nobody queries, while the procedure it corrects keeps its approved status and answers every question.
Read →Financial services
Risk appetite has no limit for a wrong answer
Every risk appetite statement is built from numbers: limits, thresholds, tolerances. Ask what the stated tolerance is for an AI assistant giving a confidently wrong internal answer, and the statement goes quiet.
Read →Standards
Nonconformance: you can't quarantine an answer
Nonconformance management rests on one unstated assumption: the defective thing is still in your hands. An AI answer is delivered the instant it exists, so there is nothing to hold, count or recall.
Read →Life sciences
IQ, OQ, PQ: nothing qualifies the knowledge
Installation, operational and performance qualification each test a system against an approved specification. Run all three on the AI assistant your team already uses and it passes, while the procedures it answers from were never qualified at all.
Read →AI governance
AI red teaming never tests the source
AI red teaming simulates an attacker and tests how the model behaves under pressure. The likelier enterprise failure has no attacker in it at all: a fair question, a confident answer, and a source nobody approved.
Read →System of record
Policy attestation stops at the reader
Policy attestation binds a named person to a specific policy version on a specific date. It is the strongest proof of knowledge most companies own. The AI answering the same policy questions signed nothing.
Read →Financial services
Key risk indicators don't watch your knowledge
Every exposure in a regulated firm has a number, a threshold and an owner. The knowledge your AI assistant answers from has none of the three. Four leading indicators that would change that.
Read →Standards
Objective evidence: an AI answer has none
ISO 9000 defines objective evidence as data supporting the existence or verity of something, and audit standards rank it by reliability. An AI answer sits outside that hierarchy entirely: no artifact, no version, no approver, no way back.
Read →Life sciences
Continued process verification stops at data
Stage 3 verifies your process forever. The procedures, interpretations and answers built on top of it were approved once and never re-verified, and an AI assistant cannot tell the difference. What continued verification of knowledge would look like.
Read →AI governance
Data classification sorts by harm, not accuracy
Most regulated companies can produce a data classification policy in a minute. Four tiers, named owners, handling rules. Every tier answers one question: what happens if this leaks. None asks whether the content is right.
Read →System of record
Your data retention policy stops at the record
Retention schedules sort records into classes, put a clock on each one and say how it is destroyed. An AI answer drawn from those records has no class, no clock and no disposal method. Two failures follow, in opposite directions.
Read →Financial services
Complete compliance documentation, wrong answers
Every framework tells you which documents to keep and how to control each one. None of them asks whether two of those documents say the same thing. An AI assistant reading your whole set is the first thing that notices.
Read →Standards
Management review never reviews your knowledge
Management review is the one forum where top management formally judges the quality system. Its required inputs all measure the machinery. None of them measures the state of the knowledge the company answers from.
Read →AI governance
Your acceptable use policy stops at the prompt
Ask most companies how they govern AI and you get the acceptable use policy. It is a conduct document: it binds the user and the tool, and it never asks whether the knowledge behind the answer was approved, current or contradicted.
Read →Life sciences
Inspection readiness never tests the answer
Inspection readiness programs test two things: whether you can produce the document, and whether your people can speak to it. Both assume someone reached for a record. Increasingly, they asked an assistant instead.
Read →Financial services
Delegation of authority stops before the answer
Your delegation of authority names who may approve what, up to which limit. No line of it covers the answer an AI assistant gives from your own documents, and accountability for that answer never moved anywhere.
Read →AI governance
AI observability can't see a stale source
A retrieval trace records which document your AI reached for and how relevant the retriever thought it was. It carries no approver, no version and no effective date, so a faithfully grounded answer from a withdrawn policy still scores well.
Read →Standards
CAPA closes on the document, not the answer
Corrective and preventive action assumes four things: a detected nonconformity, a way to contain it, a traceable cause, and a later effectiveness check. A wrong AI answer offers none of the four.
Read →Life sciences
GxP scope stops at the systems you classified
GxP compliance begins with a list of systems in scope. An AI answer is assembled across that list, from sources that were never on it, and the classification scheme never sees the answer at all.
Read →Financial services
Segregation of duties and the unsigned answer
Segregation of duties separates recording, approving, custody and reconciling. A retrieval-based AI assistant quietly occupies three of those four stages and leaves the approving seat empty. Here is where the second signature belongs.
Read →Standards
Every AI answer is an uncontrolled copy
Document control governs the document. An AI assistant answers from it and produces something with no revision number, no approver and no way back. That is an uncontrolled copy, made at machine speed.
Read →System of record
Legal hold preserves documents, not answers
A legal hold freezes the documents an AI assistant reads. It does not freeze the answers that assistant already gave, and it will not help you reproduce them.
Read →Financial services
Three lines of defense, no owner for the answer
The three lines model gives every risk an owner. Ask which line owns the knowledge an AI assistant answers from, and the question tends to go quiet.
Read →Life sciences
GAMP 5 validates the system, not the answer
GAMP 5 validates the computerized system. It was never written to govern the document an AI assistant retrieved, or the answer it assembled from that document. That gap sits outside every validation package.
Read →AI governance
Data stewardship stops at the data layer
Data stewards own tables, lineage and quality metrics. The policies, SOPs and memos your AI assistant actually answers from sit outside that remit, with no owner and no review date.
Read →System of record
Policy management ends at the policy
Policy lifecycles govern the document. They stop short of the sentence an employee is handed when they ask an assistant what the policy says, and that sentence is what people act on.
Read →System of record
Information silos aren't the problem
Every guide to information silos ends in the same instruction. Break them down, consolidate the tools, migrate everything into one place. That advice has been failing for twenty years, and location was never the real problem.
Read →AI governance
AI grounding stops where governance starts
Every grounding guide tells you to make sure the model answers from verified source material. Almost none of them say what makes a source verified, who approved it, or what happens when two grounded documents disagree.
Read →System of record
Why we built Cognatum
Cognatum did not start as a product. It started as an internal fix for thirty years of contracts that nobody could answer for. Here is what that exercise taught us, and what we built because of it.
Read →Life sciences
ALCOA principles don't cover your AI answers
Your batch records are attributable, legible, contemporaneous, original and accurate. The sentence your AI assistant wrote from them is probably none of those things. Here is how to close the gap.
Read →Standards
Change control ends where your AI answer begins
Every regulated company controls changes to its documents. Almost none control what happens to the answers already built on the old version. An AI assistant turns that gap into a daily problem.
Read →System of record
Institutional knowledge: capture is only half the job
Most institutional knowledge programs stop at capture. Writing an expert's method down does not say who approved it, when, or whether it still holds. Once an AI assistant reads it, that gap becomes the answer's problem.
Read →AI governance
Agentic AI governance covers the agent, not the answer
Every agentic AI governance framework on the first page of Google governs the same thing: what the agent is permitted to do. Almost none of them governs what the agent is acting on: who approved it, and whether it is still current. That is where the expensive failures live.
Read →AI governance
Shadow AI: a knowledge problem before a security problem
Every definition of shadow AI describes the same behavior and prescribes the same fix: find the unapproved tools and block them. That treats the supply of AI tools as the problem. The demand is what actually needs answering.
Read →Life sciences
21 CFR Part 11 and AI-generated answers: what an inspector will ask
An inspector who finds an AI assistant answering from controlled documents asks the questions Part 11 has always framed: who approved this, where is the audit trail, was it current. What a defensible answer trail looks like, mechanism by mechanism.
Read →Standards
ISO 30401 for AI governance: a clause-level walkthrough
ISO 30401:2018 is the certifiable management system standard for knowledge management. Its clauses map onto what a review board asks about AI: who approved this, when, and is it still current. A clause-by-clause walkthrough of what each requires and the evidence that satisfies it.
Read →Financial services
SEC 17a-4, FINRA supervision, and AI-generated answers
Neither SEC Rule 17a-4 nor FINRA Rule 3110 mentions AI. Both still apply to it. This piece shows where an AI answer creates records and supervision risk, and what an evidence trail looks like: who approved it, when, from what source, at which version, and how long it is kept.
Read →System of record
What is a knowledge management system of record?
It is the one place that holds the approved version itself, not a link to wherever the file happens to sit. This piece defines the term, shows how it differs from tools that only point at content, and says why AI raises the bar.
Read →AI governance
Why your AI initiative stalled at governance review
The mandate was real and the pilot worked, yet the deployment stalled the moment governance asked who approved what the assistant says. The objection is to the knowledge it reads, not the model, and it has a mechanical fix.
Read →Knowledge governed. Intelligence everywhere.
See it on your own content, in your own environment.