Skip to content
    All guides

    Keep ownership of the work

    How to avoid tying your business knowledge to one AI supplier

    AI vendor lock-in arises when changing a supplier would require rebuilding important parts of your information or workflow. Keep the business knowledge, rules and tests understandable outside the supplier’s product. Then test a small export and replacement before depending on the system for essential work.

    The starting point

    Preserve business records, rules and tests in usable forms, and treat a supplier change as a tested implementation rather than a guaranteed simple swap.

    Identify what the supplier holds for you

    List source documents, structured records, identifiers, permissions, prompts, workflow rules and evaluation examples. Add logs or review decisions needed to explain past work. Establish which are your records, where they live and what an export contains.

    A download of original documents may omit the relationships and reviewed decisions added inside the system. Those can be the parts that explain how your business works. Keep them visible in the ownership inventory.

    The GOV.UK purchasing-strategy guidance includes ownership of code and business rules in technology buying decisions. For a private service firm, use that as a prompt to clarify ownership in the actual arrangement.

    Export meaning as well as files

    What to preserve
    ItemInformation needed for reuse
    Business recordsStable identifiers, client identity and source origin.
    ProvenanceValidity start, validity end and collection date as separate fields.
    RelationshipsThe linked records and the meaning of each relationship.
    PermissionsWho may use which information, with a trusted identity mapping.
    Workflow rulesInputs, permitted actions, approval and exception behaviour.
    TestsExpected results, supporting evidence and recorded limitations.

    Said describes provenance as what makes the knowledge build useful: an agent needs the evidence and its dates, and a human needs to inspect it. Losing those fields during export can leave the text readable while the original review rules are no longer possible.

    Prefer documented formats your team can inspect and validate. A JSON or CSV file is only useful if the fields, identifiers and meanings are explained. Keep an example that another person can reconstruct.

    Keep the supplier connection separate from business rules

    An adapter can translate a request into a supplier’s API format while the business rules remain in your own workflow. The knowledge-base sample illustrates that separation with deterministic checks and an optional local model adapter.

    That design makes a change easier to reason about, but it does not establish that models are interchangeable. Different models may follow instructions differently, use different context limits or produce different tool calls. Retest the behaviour you depend on.

    Keep supplier-specific features in the dependency inventory. Document where the workflow relies on a hosted search index, proprietary automation or a particular output format. Decide which dependencies are acceptable for the value they provide.

    Test a small export and replacement

    Export one approved set of synthetic or authorised records. Reconstruct the chosen question in a separate test environment, then check the identifiers, sources, dates and access rules. Record what did not transfer.

    Run the same evaluation examples after a model or supplier change. Inspect normal answers, refusals, conflicting evidence and actions. Include the time and cost of rebuilding connectors or indexes in the decision.

    Check current export, deletion, access and termination terms directly with the supplier’s documentation and your agreement. This guide provides questions to ask; it does not assert a specific provider’s contractual terms.

    Keep a record of what can move

    Worked example

    Moving a structured reporting answer

    Invented teaching example. No client outcome is claimed.

    PRESERVE

    Invented reporting facts, client IDs, source origin, quotations and separate validity and collection dates.

    RECONSTRUCT

    Load the records into a test application and apply the same client and date checks.

    RETEST

    Current answer, expired record, conflicting sources and wrong-client request.

    LIMIT

    A successful structured-data export does not prove that an entire production system can be moved without additional work.

    Further reading

    Written by Said Lopez, founder of LeverageAI and author of Lean AI. Our guides draw on the method in the book and our work with clients.

    The next conversation

    What could your
    business do next?

    Bring your bigger goal to the AI Strategy Workshop. We help your team choose where to start, compare the likely returns and see how the first project supports what comes next.

    Discuss your AI workshop

    Start with a free 15-minute fit call. We discuss pricing there.