AI vendor selection must test the ability to resume operations after exit

The Digital Agency's June 2026 edition guideline organizes generative AI implementation into three categories: standard terms only, standard terms with individual contracts, and individual development with individual contracts, requiring procurement and contract checklists to align with case risks. Since the rights negotiable and information obtainable change depending on the contract type, one cannot select a method solely based on the ease of PoC. [1]

In exit design, define service termination, major failures, quality degradation, terms changes, price changes, and legal/policy changes as termination triggers. Confirm before the contract who can obtain the necessary data, settings, prompts, evaluation sets, audit logs, and model outputs, and in what format. Assets that cannot be obtained are entered into the decision-making process as dependency costs from the start. [1]

The pass condition for the exit is not whether a single file was obtained, but whether important business can be resumed in an alternative environment or manually. During selection, verify which of the settings, prompts, search materials, or evaluation history, if missing, would degrade quality or explainability, and move assets that cannot be resumed to contract negotiation or separate storage. Since obtainability varies by contract type, do not unconditionally reuse the same evaluation sheet. [1]

NIST AI 600-1 organizes generative AI risks into 12 groups, citing Value Chain and Component Integration as one. Generative AI services stack external data, foundation models, libraries, search infrastructure, plugins, and cloud, meaning management by a single contracting party cannot capture the impact of upstream changes or stoppages. [2]

The evaluation sheet should include the upstream model and version, data provider, hosting region, key subcontractors, replaceable components, change notification deadlines, and known single points of failure. Confirm not only the presence of model cards or certifications but also the track record of failure/vulnerability disclosure, remediation time, scope of third-party audits, and evidence that can be explained to downstream customers. [2]

Before determining quality degradation as the contractor's sole non-performance, trace alternative causes such as upstream model changes, search components, cloud failures, and internal settings. Assign to each part of the value chain the person receiving change notifications, the existence of alternatives, and the scope of obtainable evidence. [2] If a single point of failure with unidentified causes remains, narrow the scope of adoption for important business, even if certifications or average performance are good.

The UK government's AI procurement guidelines present Top 10 Considerations, treating the avoidance of black boxes and vendor lock-in independently in item 8. This approach involves considering explainability, open standards, royalty-free licenses, and conditions allowing other suppliers to continue development, and reflecting them in procurement documents. [3]

Lock-in is not just API compatibility. It is divided into data lock-in where conversation history or evaluation results cannot be exported, functional lock-in where unique prompt functions cannot be reproduced, human lock-in where operational procedures and staff skills depend on specific screens, and contractual lock-in where responsibility boundaries or deletion certificates cannot be transferred. Estimate alternatives and migration costs for each. [2] [3]

Government procurement materials and NIST's generative AI profile indicate that obtainable information and controls differ according to contract type and value chain. [1] [2] When all requirements cannot be obtained via contract, do not silently accept gaps; instead, create a concession ledger that pairs each conceded condition, remaining business impact, and compensating control one-to-one.

In the concession ledger, individually record unobtained change notifications, audit information, return formats, and termination support, and determine if they can be compensated by internal storage, reduced scope of use, maintaining manual processes, or short review periods. Gaps that cannot be compensated are not offset by price or average performance but remain as reasons for not adopting them in important business. Also record the failure conditions of the compensating controls.

As a hypothetical example, if evaluation history cannot be obtained at contract termination, first test whether necessary items can be saved to the internal side during operation. If that is also not permitted, it is an exit gap where decision grounds cannot be reproduced in a separate environment. Make the actual ability to operate the compensating control, rather than the presence of contract clauses, the condition for accepting the concession. Also record the test date and responsible person.

Even if exported files exist, important business cannot always be reconstructed. Classify assets into four states: directly transferable to a separate environment, usable after conversion, creatable manually, or unreconstructable. For the last two states, confirm whether necessary staff skills and decision grounds remain, and assign an owner for reconstruction.

In the reconstruction exercise, resume representative business without referring to the original environment, using only the obtained assets and information permitted by the contract. If missing items are compensated by staff memory, it is not a success; re-evaluate whether that knowledge can be recorded as a migration asset. Separate whether the cause of quality difference is asset loss or functional difference, and if mixed, keep both.

When accepting an exit gap, specify the target business, acceptable quality difference, manual maintenance period, renegotiation conditions, and final approver. If upstream dependencies or authorities change and previous reconstruction becomes impossible, reopen the concession ledger before the renewal contract. Do not use the single word 'portable' as the basis for approval; judge with a time limit.

• WAIVER: Individually record unobtained rights, information, and support
• COMPENSATE: Attempt correspondence with internal storage, reduced operations, and manual work
• REBUILD: Classify assets into transferable, convertible, reconstructable, and non-reconstructable
• ACCEPT: Define exit gaps, acceptable ranges, review conditions, and approvers

Even if the proposal states 'exportable,' completeness and reusability are unknown. During selection, output test settings, prompts, search documents, logs, and evaluation results, and import them into a separate environment. Measure missing items, conversion work, downtime, quality differences, and the ability to revert to manual work, scoring the results alongside price and performance.

Contracts should include data ownership, secondary use, change notifications, SLA, accident cooperation, audit rights, termination support, return formats, deletion certificates, support periods, and responsibility boundaries according to purpose. As indicated by NIST, prepare decommissioning and fallback in peacetime, and redo exit exercises not only at fixed intervals such as annually but also whenever major model changes or dependency changes occur. [2]

In the exit exercise, score not just export completion but also missing items, settings requiring conversion, work until resumption, quality differences with the old environment, and collection of deletion certificates. If the ability to resume has declined compared to the previous time, major upstream dependencies have changed, or necessary migration authorities have been lost, stop approval for the renewal contract. The procurement manager and business continuity manager sign the results separately. [2] [3]

• List functions that cannot be reproduced with alternative models
• Extract all data in machine-readable format and verify counts and missing items
• Actually run manual/degraded operations when the vendor stops
• Determine retention period, deletion method, and deletion certificate in the contract after termination
• Procurement Manager: Confirm assets and termination support obtainable via contract
• Technical Manager: Reuse settings, data, and evaluations in a separate environment
• Business Continuity Manager: Judge resumption in an alternative environment or manually