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.
Cognatum Team · Oct 20, 2026 · 6 min read
Three letters, three protocols, one conclusion. The equipment was installed as specified. It operates across its range. It performs the same way batch after batch. That conclusion is the backbone of qualification in every GxP plant.
Cognatum governs the entry
source · version · approver · permissions
Now point the same three protocols at the AI assistant your team already asks questions of. It will pass. It was installed as specified. It responds within its range. It performs consistently under real conditions, with real users and real data.
It can still tell a supervisor that the hold time is four hours, from a procedure that was superseded in March.
What the three stages test
Installation Qualification confirms the system is installed and configured as specified. Operational Qualification confirms it performs as intended across its defined operating range. Performance Qualification confirms it performs consistently under real operating conditions, with real users and real data. That is how one vendor's quality officer summarised the three stages this month.
Each stage runs against an approved protocol. The protocol names the test, the thing being challenged, the data to collect and the acceptance criteria. Greenlight Guru's guide states the IQ criteria plainly. They confirm that equipment, utilities, software versions, drawings, manuals, certificates and installation conditions match approved specifications.
Always against a specification
One word appears in every stage. Specifications. Qualification is a comparison. One side is the thing in front of you. The other side is a document that says what it should be.
The calibration rule
There is a prerequisite nobody skips. Calibration is confirmed before testing begins. The reason is stated directly. Qualification data is only as reliable as the instruments used to generate it. Every gauge, sensor, timer and scale used in a protocol needs a current calibration record first.
That rule is this whole argument, moved up one layer. An answer is only as reliable as the documents used to generate it. Nobody checks those documents before the assistant answers.
A qualified process has qualified inputs
PQ comes closest. It is run in the actual facility, with trained personnel, using the utilities, equipment and control procedures that will produce commercial batches. The recommended PQ protocol lists the manufacturing conditions, the operational parameters and the raw materials used.
Raw materials. In a plant, material does not enter a validated process on trust. It arrives with a specification. It is tested. It is released. The qualified process assumes qualified inputs.
A retrieval assistant has raw materials too. They are the procedures, the deviation reports, the validation summaries, the decks and the mail threads it reads before it answers. None of that is released to anything. It is simply present in a folder.
The requalification list names the machine
Qualification is not permanent and the industry knows it. The triggers are well rehearsed. Equipment relocation. Major repair. New tooling. Software or firmware changes. Utility changes. Calibration failures. Repeated deviations. Performance data drifting out of control.
Read the list again and look for the knowledge. Not one trigger is a document. A procedure can be revised, withdrawn or contradicted by a second procedure, and no stage of qualification notices.
What the guidance already expects
The expectation exists. It is simply not written as a qualification stage.
ICH Q10 names knowledge management as an enabler of the pharmaceutical quality system, alongside quality risk management. It defines it as a systematic approach to acquiring, analysing, storing and disseminating information related to products, manufacturing processes and components. It says product and process knowledge should be managed from development through the commercial life of the product. EU GMP 1.4(ii) says the same thing in one line. Product and process knowledge is managed throughout all lifecycle stages.
An obligation with no protocol
The Q9(R1) questions and answers settle what an inspector will ask for. There is no regulatory requirement for a formal knowledge management system. The second half of that answer is the part that matters. It is expected that knowledge from different processes and systems is appropriately utilised.
So the duty is real and the mechanism is left open. MHRA's guidance sets out the core elements of a compliant data governance system across all GxP sectors, and data governance is about records. Qualification is about systems. The knowledge sits between the two, and EudraLex Annex 15 gives it no stage.
Accuracy is declared, not verified
AI regulation has the same shape. The EU AI Act asks high-risk systems to achieve an appropriate level of accuracy and robustness throughout their lifecycle. The accuracy metrics are then declared in the instructions for use. That is a property of the system, measured once and stated.
Grounding the system in your own documents does not close the gap either. A preregistered study of commercial legal research tools built on retrieval found that the leading products hallucinated between 17 and 33 percent of the time. Their vendors had advertised hallucination-free citations. Retrieval lowers the error rate. It does not sign anything.
What a fourth stage would test
Call it knowledge qualification. It asks four things of an answer, and all four are ordinary questions on a plant floor. Which approved entry did this come from. Who approved it, by name. When. What has happened to that entry since.
Seven steps, then a person
The Cognatum Knowledge Loop is built around those four questions. Eight steps run the knowledge and AI runs seven of them. It finds material across the systems where it already lives, structures it, cleans it, enriches it and surfaces it. The eighth step is Approve, and a named person holds it. The AI does not decide what is true. It routes the item to the person who does, and that person can rework the answer before approving it.
When a source moves
Where two documents disagree, both are presented and the disagreement is identified, rather than one being picked quietly. When a source changes, Cognatum notifies the owners of the entries built on it. It does not rewrite them on its own. It also records where the knowledge stood when an answer was given, so a question asked in March can be reconstructed in October.
A test you can run this week
Pick a process you have qualified. Take a question the floor actually asks about it. A hold time, an action limit, a sampling frequency. Put that question to the assistant your team already uses.
Then ask for three things. The approved entry the answer came from. The person who approved it. The date.
Your IQ, OQ and PQ packages are complete. If nobody can produce those three inside a minute, the packages qualified a system that answers from material nothing qualified. Your company's knowledge isn't missing. It's unusable. Cognatum changes that.
Sources
- A Basic Guide to IQ, OQ, PQ in FDA-Regulated Industries (thefdagroup.com)
- IQ, OQ, PQ: Complete Guide to Process Validation in Medical Devices (greenlight.guru)
- IQ, OQ, PQ: What Software Validation Means for Pharma QA (informait.com)
- EudraLex Volume 4, Annex 15: Qualification and Validation (health.ec.europa.eu)
- Is Knowledge Management a Part of the Pharmaceutical Quality System? (gmp-compliance.org)
- Guidance on GxP data integrity (gov.uk)
- EU AI Act Article 15: Accuracy, robustness and cybersecurity (europa.eu)
- Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (arxiv.org)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
Do IQ, OQ and PQ apply to an AI assistant?
They can be run on it, and that is the problem. A retrieval assistant is a computerized system, so it can be installed to specification, tested across its operating range and shown to perform consistently with real users and real data. All three stages examine the system. None of them examines whether the documents it reads are approved, current or in agreement with each other.
Is knowledge management a GMP requirement?
Not as a formal system. The ICH Q9(R1) questions and answers state that there is no regulatory requirement for a formal knowledge management system, and then add that knowledge from different processes and systems is expected to be appropriately utilised. ICH Q10 lists knowledge management as an enabler of the pharmaceutical quality system and EU GMP 1.4(ii) asks that product and process knowledge be managed throughout all lifecycle stages. The duty is there. The protocol is not.
Does grounding an assistant in our own SOPs make its answers reliable?
It helps and it is not sufficient. A preregistered evaluation of commercial legal research tools built on retrieval found hallucination rates between 17 and 33 percent for products marketed as hallucination-free. Grounding anchors an answer to retrieved text. It does not establish that the retrieved text was approved, that it is the current version, or that a second document does not say the opposite.
Can Cognatum make us compliant with the EU AI Act or NIST AI RMF?
No. Software does not confer compliance on an organisation. Cognatum aligns with those frameworks and supplies the evidence an assessor asks for, including a named approver, a date and a source on every approved entry, plus a record of where the knowledge stood when an answer was given. The compliance decision stays with your quality and regulatory functions. ISO 30401 certification sits with Atokah Tech and holds while running on Atokah servers.
Do we have to move our documents into Cognatum first?
No. Information stays where it lives, in the departmental repositories your teams already use. Cognatum maintains a bidirectional flow with those systems rather than housing the content itself, so a validation team keeps its files where they are. The departmental split is normal. What is missing is a governed layer above it.