A team has built its AI work around a single provider: the instructions, the definitions in the settings, the conversation history. Then something changes. A model is replaced by a successor, the terms shift, or a new rule requires data to stay in Europe. How quickly can the team switch, and what is lost?
Dependency is a choice, not an accusation
Frontier models have a place, including in research and business development. The question is not whether an organisation is for or against a provider. It is about choosing consciously which dependencies are acceptable. There are two routes to consider.
Route one: everything through an outside provider. Instructions, context and definitions come together in its environment. That can simplify the start, but makes the consequences of changes important. Models can disappear: OpenAI retired API access to GPT-4.5-preview.1 What then happens to recorded knowledge, established workflows and the ability to reconstruct earlier answers?
Route two: an independent company brain. Meaning, relevant history and rules are maintained in a foundation that is not tied to one model. This requires explicit arrangements for access, export, maintenance and responsibility. A separate storage location is not enough: the relationships between sources, definitions and decisions must remain portable too.
The company brain keeps the evidence. The model puts it into words.
Known answers need more than a model
For a repeatable question, such as the margin of a product group over a particular period, the calculation method may already be defined. A discount rule may also be recorded in advance. That makes the required evidence known, but does not mean every model will automatically produce the right answer.
Retrieval finds relevant information. An approved deterministic calculation executes an established rule. The model can then verbalise the result. These tasks should be distinguished, with checks on source selection, calculation and wording. Missing or conflicting information should leave uncertainty visible.
A semantic layer records concepts and relationships. Equal results do not follow automatically: definitions, versions, scope, sources and permissions must be applied consistently. A margin for a closed month is not necessarily comparable with a current provisional figure.
Permission is also a human responsibility. Which information someone may access and which actions require approval must be explicitly decided and technically enforced. Model selection does not replace those decisions.
A smaller open-weight or self-hosted model can be considered for this work. Suitability must be established through evaluation on concrete tasks, including exceptions. Self-hosting or using open weights does not guarantee data sovereignty or safety.
Distinguish controlled retrieval, approved calculations and open questions. Evaluate the wording too.
Make costs and switching visible
An independent company brain needs maintenance. Hosting, monitoring, security, model testing and updates all need an owner. The argument is therefore not automatic savings, but visibility into responsibilities and freedom of choice.
Under usage-based billing, the cost estimate is usage multiplied by the applicable rate, plus any hosting, monitoring and other operational costs. Which items apply depends on the architecture and contract. With self-hosting, reserved capacity and its utilisation also matter. Higher usage can change costs, but not in the same way under every arrangement.
Portability deserves a practical test. Can definitions, access rules, relevant history and evaluation examples be transferred? Does another model produce a usable answer from the same permitted sources? The time required for changes and renewed evaluation belongs in that assessment too. An export file alone does not demonstrate that a working process can be moved.
Portability requires transferable knowledge and renewed tests of access, answers and costs.
Frequently asked questions
Five questions to ask today
- What share of the questions put to the AI has a fixed, known answer?
- Where does that answer live: in an independent foundation, or in the settings and prompts of a single provider?
- How quickly can the model be switched, and what is lost in the process?
- What happens to costs if usage doubles?
- Which data and rules must never leave the organisation's environment?
List the uncertainties and review one repeatable question. Record its source, definition, any calculation, permissions and expected result. Then test what remains intact when the model changes.
The company brain is yours. The model is a choice.
