Don't just document your AI exit plan. Test it.
A fallback model existing on paper does not prove the application will still work.
ForgeStack takes one production AI workload and tests its credible exit options against the behaviour that matters to the application: tool calls, structured outputs, retrieval, permissions, refusals, context handling, latency and cost. The drill identifies what survives the switch, what breaks and what cannot be ported.
Who it is for
Operational-resilience, risk and technology leaders at UK banks, building societies, insurers, payment and e-money firms and other regulated teams that run production applications on external model providers. The FCA's good and poor practice paper of 19 October 2026 lists "Single-platform dependence without exit options" among its areas for improvement and names, as good practice, firms that "developed and tested contingency plans in case of unexpected changes to their AI provider arrangements".
One workload. Five to eight working days.
The engagement includes:
- dependency mapping for the workload: model, embeddings, tools, memory, identities and data flows
- a production-derived test set, built under the customer's data-protection arrangements
- testing against the current provider and credible alternative endpoints, including a second host of the same model, a second vendor and a private open-weight fallback
- deterministic checks wherever possible: schema validity, tool-call equivalence, refusal set, context fit, latency and cost per request, retrieval recall where embeddings move, a permissions checklist
- a failure register by endpoint
- a degraded-mode statement for what does not port: agent memory, provider-side reasoning state and user-delegated permissions
- a remediation list
- a dated exit-drill record for operational-resilience and governance review
The drill is performed in the customer's environment, with the customer's keys, where practical. There is no ForgeStack platform to install.
Fixed-scope paid engagement
Price is confirmed during scoping. The drill is not a pilot and not a free scan.
After the drill
Where the drill finds portability failures, remediation and ongoing re-testing can be scoped separately from the findings. ForgeStack does not sell an AI continuity software platform; repeated paid customer work decides what, if anything, should eventually be automated.