Neither open-weight nor closed AI models are safer by default. Closed models concentrate risk in one supplier you cannot inspect; open-weight models move the risk in-house, onto your own security, hosting and support. Boards should ask which risks the business is better placed to carry, and require vendors to disclose which models sit behind their products.
This question is now being debated openly by AI investors and technology leaders. That debate is mostly about industry and national policy. This article is AIPEX’s own view of the narrower question a board faces: what the choice means for supplier risk.
What is the difference, in board terms?
A closed model is used as a service. The provider keeps the model, runs it, updates it and sets the price and terms. You send data in and receive answers back. An open-weight model is one whose trained parameters are published, so a business or its vendor can download it, run it on infrastructure it chooses and adapt it.
The labels hide a lot of variation. Open-weight licences differ, some with restrictions on commercial use or on the size of the user. Closed services differ on data retention, training on customer data and where processing takes place. The practical question is not “open or closed?” but “who controls this model, where does our data go, and what happens if the arrangement ends?”
Where does the risk sit in each model?
| Risk | Closed model, used as a service | Open-weight model, run by you or your vendor |
|---|---|---|
| Data exposure | Data leaves your environment; you rely on the provider’s terms and controls | Data can stay in your environment; you own the controls and any failure of them |
| Change without notice | The provider can update or retire the model on its own timetable | You decide when to change version, but must track fixes yourself |
| Price and terms | Exposed to price rises and usage limits you cannot negotiate away | Costs shift to hosting, compute and specialist staff |
| Security of the model | The provider secures and patches it | You secure the model, its hosting and the software around it |
| Provenance and licence | Contract terms define your rights | The licence defines your rights; check commercial-use limits |
| Exit | Harder if prompts, workflows and data are tied to one service | Easier to move, if the team and infrastructure can follow |
Neither column is empty. The choice is about which list your organisation is equipped to manage.
Why is this a vendor question and not just a technology one?
Most businesses do not pick models directly. They buy software, and the software vendor picks the model. A single product may use a closed model for one feature and an open-weight model hosted by a third party for another, and switch between them without telling the customer. The board’s exposure therefore depends on decisions made several links down the supply chain.
This is why the choice belongs in vendor due diligence. Guidance from the UK National Cyber Security Centre and US partners on secure AI development stresses understanding and managing the supply chain of models and data behind an AI system (NCSC and partners, Guidelines for secure AI system development, November 2023). A board that has approved a vendor on the strength of its brand may, in practice, be relying on a model it has never heard of.
What should the board require from AI vendors?
- Name every model behind the product, who provides it, and whether it is closed or open-weight.
- State where each model runs, and where our data is processed and stored.
- Confirm in the contract that our data is not used to train or improve any model, including by providers behind the vendor.
- Give advance notice before switching or materially changing a model, with the right to test first.
- For open-weight models, show how they are secured, patched and monitored, and confirm the licence permits our use.
- Explain how we would exit, and in what form our data, prompts and configurations would be returned.
A vendor who answers these readily has thought about AI risk. One who cannot name its models has not, and that is itself the finding. Our article on data governance, privacy and security at board level covers the data side of these questions in more depth.
Does the choice change the organisation’s own obligations?
It can. A business that takes an open-weight model and adapts it, or puts its own name on the resulting system, may take on duties that would otherwise sit with the vendor, including under the EU AI Act where it applies, which our article on EU AI Act fines and enforcement puts in context. The same is true of data protection: running a model yourself makes you responsible for how personal data is handled inside it. That does not make open-weight models the wrong choice. It means the decision should be made knowingly, with legal advice where the use is regulated. This is governance guidance, not legal advice.
Where to start
Our briefing on AI vendor due diligence and the questions vendors hope you won’t ask sets out 24 questions in eight areas, the red flags in the answers and six contract clauses that matter most. For boards that buy AI from vendors and cannot see where the risk sits, the step on its four-step path that fits is step 2, Assess: the AI Readiness Assessment (£999 + VAT) shows where governance, data and risk controls are weak, with a 60-minute debrief. The AI Vendor Due Diligence Checklist below puts the questions in one place for procurement.
Frequently asked questions
Are open-weight AI models riskier than closed models?
Not by default. Closed models concentrate risk in a supplier you cannot inspect, including price, change and data handling. Open-weight models move risk in-house, onto your own security, hosting, patching and licence compliance. The right choice depends on which risks the business is equipped to manage.
Why should boards ask vendors which models they use?
Because the vendor, not the customer, usually chooses the model, and may use several or switch without notice. A board's real exposure depends on those choices, so vendors should name every model behind the product, where it runs and who provides it.
What should an AI vendor contract cover about models?
No training on your data, including by providers behind the vendor; notice before a model is switched or materially changed; where data is processed; licence terms for open-weight models; and exit arrangements. See AI Vendor Due Diligence →
Correct as of 28 September 2026. This content reflects the law, official guidance and deadlines as published on that date. The EU AI Act and its guidance, standards and deadlines are still being amended and clarified. If you are reading this after 28 September 2026, check for later amendments, European Commission guidance and harmonised standards before relying on it. It is general information, not legal advice, and AIPEX accepts no liability for decisions taken on it without specific professional advice.