The Private Model Needs an Exit Plan

The system becomes useful, then indispensable. Prompts, memory, indexes, policies, connectors and evaluation history accumulate around one provider.

Changing the model later may mean losing the institution’s working intelligence.

Exit includes more than data

- Source and index formats.

- Persistent memory.

- Policies and tool permissions.

- Prompts and workflow logic.

- Evaluation sets.

- Audit and incident history.

Value object — The AI Exit Package

- Exportable artefact.

- Owner.

- Format and portability test.

- Vendor assistance.

- Substitute component.

- Maximum tolerable transition.

Exercise before leverage

Move one workflow to an alternate model or local environment. Identify proprietary dependencies and answer-quality changes.

A private model is controlled only if the institution can change it without losing memory, evidence or authority.

Where this breaks

Exit planning performed after dissatisfaction is too late. The provider already controls formats, connectors, memory and the staff knowledge required to leave.

The operating move

Maintain an exportable private intelligence package and run a small migration annually. Keep policy, evaluation and source ownership outside provider-specific features.

Export memory and indexes.

Document tool permissions.

Retain model-independent evaluations.

Measure substitute answer quality.

The test

Move one live workflow to a different model for a week. The time and information lost reveal the provider’s actual leverage.