Validated. Or not in production.
In a GxP environment a system is not finished when it works. It is finished when the evidence that it works has been written, reviewed, approved and filed where an inspector can find it. Altuon builds document intelligence, validated software and market intelligence for companies that manufacture, test and sell medicines, and it writes the validation package as part of the work rather than after it.
Ce qui est en jeu
The regulation reaches into the software itself. Good practice obligations, GxP, decide how a computerised system that touches product quality, patient safety or data integrity must be specified, validated, operated and changed. In the European Union, Annex 11 of the GMP guide requires that such systems are validated, that their audit trails are reviewed, that access is controlled and that changes go through formal control. In the United States, 21 CFR Part 11 sets the conditions under which electronic records and electronic signatures are accepted as equivalent to paper: secure, time-stamped audit trails, controls on who may sign and how, and records that remain accurate and retrievable for the retention period. Swissmedic, the EMA and the FDA inspect against these expectations, and a system that cannot produce its own evidence is a finding waiting for an inspection date.
Data integrity is the lens through which the inspector reads everything. The ALCOA principles ask that a record be attributable, legible, contemporaneous, original and accurate, and the extended set adds that it be complete, consistent, enduring and available. A model that summarises a deviation report, a workflow that routes a CAPA, an extraction that lifts a value from a batch record: each produces or transforms a record, and each has to be shown to preserve those properties. The question a quality function asks of any new system is not whether it is clever but whether it can be validated, and whether the validation will survive the next model update.
The cycle is slow by design and fast at the edges. A core quality system, a laboratory information system or a manufacturing execution system was validated once at great expense and is changed reluctantly, because every change carries a documentation obligation and a risk assessment. Around that core, the volume of documents grows without pause: deviations, change requests, supplier audits, regulatory correspondence, submissions, labelling variations. The opportunity is in the documents. The risk is in touching the core without a change-control record the inspector will accept.
The buyer is the quality function as much as the technology function. The head of quality wants the validation plan, the risk assessment and the periodic-review schedule before a demonstration. The head of regulatory affairs wants to know which jurisdictions' submission formats the system understands and which it does not. The chief information officer wants an architecture that will still be supportable when the validation is due for review. Commercial leadership wants intelligence on rivals' pipelines and pricing, and campaigns that stay within the advertising rules of each market. And every one of them is answerable, in the end, to an inspector who did not attend the sales meeting.
Pratiques applicables
- Intelligence artificielleSignature, là où Altuon mèneQuality and regulatory document intelligence under GxP: models that read deviation reports, CAPA records, change requests, supplier audits and regulatory correspondence, extract and classify with a person approving wherever the procedure says one must, and leave an audit trail per output. Each model has an owner, an evaluation set built from the company's own records and scored by its quality staff, a version, and a validation package written in the form a computerised-system validation function expects. Patient and healthcare-professional data are handled within the revised Swiss FADP and the GDPR, and where a model touches a regulated decision, the EU AI Act's obligations are mapped in the same register.
- Conseil en technologieCœurValidated-system strategy for estates in which every change carries a documentation obligation: which systems are GxP-relevant and which are not, how the validation approach is risk-based rather than uniform, how data governance meets the ALCOA principles, and what has to be true of the organisation before a model can be admitted to a quality process. The result is a roadmap the quality function has signed, not a slide the technology function presented.
- Intelligence de marchéCœurPipeline, pricing and regulatory intelligence in which the sources are public and the reading of them is the whole advantage: trial registries, regulator agendas, pricing and reimbursement decisions, patent expiries and competitor filings, read continuously and reported at board cadence with every claim evidenced to its source. Where a launch, a licensing decision or a price negotiation depends on what a rival will do next, the board can show what it knew and when.
- Publicité et croissanceCœurDisease-awareness and permitted product communication within the rules of each jurisdiction, where advertising prescription medicines to the public is prohibited in Switzerland and the European Union and permitted with conditions in the United States. Medical, legal and regulatory review is built into the publishing path rather than appended to it, audiences are separated by rule between healthcare professionals and the public, and every claim carries its approver and its reference.
- Marque et identitéCœurCorporate and product identity under regulatory constraints on naming and claims, and on the separation between the two. The corporate mark, the product brands and the disease-awareness properties are designed as one architecture so that the company can speak where the regulation permits without lending the corporate voice to a product it may not promote, and so that the identity survives the acquisitions and divestments the industry runs on.
- Ingénierie logicielleEn soutienQuality and batch-record systems where change control and audit trails are conditions of operating, not features. Integrations and extensions are built against the validated core through controlled interfaces, each change carries its risk assessment and its test evidence, and the release is gated by the quality function's approval so that the system's validated state is maintained rather than re-established.
Notes régionales
- Europe et DACH
- The Basel-area companies and their suppliers buy discretion and permanence. Swissmedic's expectations govern the validated estate, the revised Federal Act on Data Protection governs patient and healthcare-professional data, and the data plane stays in Switzerland unless a written reason says otherwise. Non-disclosure comes before any detail is exchanged, and the partner is expected to be present in Basel and Zurich rather than on a call. German is a working language of the quality function and the documents are read in it.
- Moyen-Orient et Afrique du Nord
- Companies in Jordan and the wider region are building local manufacturing capacity under national procurement and national regulators, and the work has to serve that programme. Arabic labelling and patient information, regulator relations conducted in person, sovereignty over data, and technology transfer that leaves capability in the country rather than a dependency on the vendor are part of the mandate, not afterthoughts. Jordan's Personal Data Protection Law governs how personal data is handled, and partnership is measured over years.
- Amérique du Nord
- US companies buy velocity under 21 CFR Part 11 and the FDA's inspection regime. The electronic records, the electronic signatures and the audit trails have to be in place from the first release, not retrofitted before an inspection, and the validation approach has to be risk-based so that the documentation moves at the speed of the work. Where advertising to the public is permitted, the fair-balance and substantiation rules decide what may be said, and the review path is built to meet them.
Régions desservies
Europe et DACH
Siège
Turin, Italie
Mission représentative
- Intelligence artificielle9 septembre 2026Sovereign AI and Swiss data residency for financial institutionsA Swiss bank can use frontier models without sending a byte of client data abroad. It requires an architecture decision, not a vendor promise, and the revised Data Protection Act is clearer on the point than most vendors are.
- Achats9 septembre 2026Procurement's questions about AI vendors, answeredWho owns the model weights? Where is the training data? What happens to our prompts? Can we leave? The questions procurement teams ask AI vendors are the right ones. Here are the answers a serious firm should be able to give, and the ones that should end the conversation.
Les questions qu’on nous pose
How is a model validated for use in a GxP process?
The same way any computerised system is, with an approach matched to the risk. The validation plan, written in the Define phase, states the intended use, the GxP relevance, the risk assessment and the acceptance criteria. User requirements, functional and design specifications, installation, operational and performance qualification, and a traceability matrix follow, with a person's review and approval wherever the procedure requires it. For the model itself, the evaluation set is built from the company's own records and scored by its quality staff, and each version is frozen, recorded and re-qualified before it is admitted. The package belongs to the company and is written to be read by an inspector.
Do your systems meet Annex 11 and 21 CFR Part 11 on electronic records and signatures?
They are designed to. Every record a system creates or changes carries a secure, computer-generated, time-stamped audit trail that records who did what and when, and the trail cannot be edited or disabled by the user. Electronic signatures are bound to the record they sign, to the signer's identity and to the meaning of the signing, with the controls on identity and access that both frameworks require. Records remain accurate and retrievable for the retention period in a human-readable form. The company's own procedures decide how the audit trail is reviewed and how often, and the system produces the review as a report rather than a search.
Where do the data and the models run, and can they stay in Switzerland or on our premises?
The data plane is decided per system in the Define phase and recorded in the architecture register: a Swiss region, the European Union, the United States, the company's own data centre or a national cloud. Models run inside that plane; nothing that constitutes a quality record, patient data or healthcare-professional data leaves it. For an on-premises deployment the proposal states what the company provides and which model options fit that footprint, so the residency commitment is costed rather than assumed. The sub-processor list is short, named, and approved by the company before any third party touches its data.
What happens to the validated state when a model or a dependency is updated?
Nothing changes in production without change control. A model update is a change like any other: it is raised, assessed for its effect on the validated state, tested against the frozen evaluation set, approved by the quality function and recorded, with the previous version retained as the rollback. Dependency and platform updates follow the same path with a risk-based scope, so that a security patch does not demand a full re-qualification while a change to the model's behaviour does. The periodic-review schedule confirms that the system remains in its validated state between changes.
Who owns the intellectual property, the data and the validation documentation at the end?
The company does. Source code, prompts, configuration, fine-tuned weights, evaluation sets and the whole validation package are assigned to the company on payment, and the agreement says so in terms procurement can read. Altuon retains its pre-existing tooling and general methods, which are listed. At exit the company receives everything needed to operate, re-validate and hand the system to another party, with a transition period stated in the agreement, and no record is held in a form the company cannot export.
How do you keep advertising and disease-awareness work inside the rules of each market?
By building the review into the path rather than at the end of it. Every campaign is classified by jurisdiction and by audience before anything is drafted: prescription medicines are not promoted to the public in Switzerland or the European Union, disease-awareness material does not name or allude to a product where the rules forbid it, and material for healthcare professionals is gated to them. Medical, legal and regulatory reviewers approve each asset, the approval is recorded with the claim and its reference, and nothing is published without it. Where a market permits direct-to-consumer advertising, its substantiation and fair-balance requirements govern the copy.
Bring us the system your inspector will ask about.
Whether it is a document-intelligence programme, a change to a validated core, a competitive-intelligence function or a campaign under advertising rules, the Discover phase begins with your GxP obligations and ends with a validation approach your quality function can sign. Request a proposal, or book a briefing for the heads of quality, regulatory and technology together.