An enterprise can select an AI supplier for a sensible reason and still accumulate dependencies it has not evaluated. A model performs well, a managed service accelerates delivery, and an existing commercial relationship makes procurement straightforward. Over time, application logic, evaluation methods, data access, and operating procedures become tied to that choice. The cost of changing suppliers grows beyond the price of replacing a model.
Dependence is not inherently a strategic failure. Suppliers provide capabilities that an organisation cannot always build economically or maintain effectively. The leadership question concerns which dependencies are acceptable, which require protection, and how the organisation will respond if price, performance, availability, or contractual conditions change.
The data-readiness article established which information supports a particular decision. This article examines where that information and the surrounding capability sit, who controls them, and what happens when a supplier relationship changes. A credible architecture needs to support today’s service and a proportionate route to future change. Promising flexibility without testing the work involved leaves the investment case incomplete.
Separate the capability from its supplier
Architecture choices need to follow the functions required by the business service. Information retrieval, document preparation, predictive assessment, decision policy, and execution are different functions, even when a supplier offers them through one product. Making those boundaries visible helps the organisation decide what to obtain externally and what it needs to control.
The lending investigation from the data-readiness article illustrates the distinction. The bank approved a bounded evaluation among existing business account holders who meet current lending criteria. Its purpose is to test whether seasonal borrowers can be served more quickly and economically, with acceptable credit and customer outcomes. Acquisition economics, assessment savings, and the limits of the repayment evidence remain part of that evaluation.
The proposed service uses a purchased language capability to extract application information and assemble material for staff review. A separate predictive component assesses cash-flow evidence, while the bank’s lending policies determine the permitted action. The supplier’s document service helps the process, but does not own the credit decision. These boundaries need to be preserved in implementation and operation, rather than existing only in a diagram.
The strategic-advantage article established the ownership principle of buying general-purpose language capabilities while retaining control of permissioned context, evaluation, decision policies, and operational integration. This lending service tests whether that boundary holds in practice. The bank retains its authoritative application records, cash-flow definitions, evaluation cases, and decision policies. Supplier outputs are recorded with the version and context needed to investigate their use. Staff review and the bank’s decision controls remain in place. This arrangement gives the team a clearer basis for assessing what would need to change if the purchased language capability were replaced.
Separation also has limits. The language service may use supplier-specific formats, retrieval features, or workflow tools. Removing it can require changes to the surrounding application, even if the business functions are conceptually distinct. Leadership needs a view of those practical dependencies before interpreting a modular design as evidence of easy substitution.
Compare the complete operating choice
An AI supplier decision should include integration, evaluation, controls, support, and change over the operating life. A low price per request can be outweighed by additional review work or difficult integration. A more expensive capability can be economical if it reduces total effort and produces acceptable outcomes. The comparison must use the bank’s own workload and operating requirements.
In the illustrative lending evaluation, the preferred document service extracts relevant fields accurately enough to support review. The alternative is 30% cheaper per request, but roughly doubles correction time per application because of errors in dates and amounts. Comparing API charges alone would favour the alternative while missing the staff effort and error-handling cost it adds.
The bank evaluates both services against the same representative applications, including difficult formats and missing information. It measures extraction quality, correction effort, processing time, and the effect on downstream assessment. The evaluation also examines access restrictions, support arrangements, and the ability to preserve the records required for oversight. The assessment includes contractual terms on document retention, use of customer information for model training, and data location, connecting them to the permissions established in the data-readiness article. Permission for the bank’s own use does not establish permission for a supplier’s use. These requirements are part of service suitability, rather than supplementary purchasing preferences.
Commercial assessment needs to include what happens as use expands. Pricing changes, minimum commitments, support charges, and usage restrictions can alter the economics. A supplier’s ability to change versions or withdraw a capability can create additional evaluation and engineering work. The bank needs to establish which of these conditions it can negotiate and which it must accept or mitigate.
The comparison does not require a prediction of every future event. It needs a few consequential scenarios that expose the investment’s sensitivity. What happens if usage grows substantially, the price rises, or an update changes extraction behaviour? What happens if the service becomes unavailable? Each scenario connects an architectural dependency to an operational and financial consequence.
Test what changing suppliers actually requires
A credible route to substitution needs evidence from the service itself. Compatibility at the interface is useful, but it does not establish equivalent behaviour. A replacement can accept the same request and return the same structure while producing different information, confidence, or errors. The receiving workflow and its controls must be tested against those differences.
The lending team runs a bounded substitution exercise alongside the already approved lending evaluation, using the same representative applications, test cases, and review process. Reusing that work limits the additional effort, which the team records separately from elapsed time. The exercise is confined to the document service, with further investment contingent on the lending evaluation. It checks that authoritative records, evaluation cases, and decision policies remain accessible when the alternative is used. The exercise also tests the application changes needed to replace the supplier and the work required to reassess the resulting service.
The interface substitution takes two days. Achieving acceptable output quality and completing the relevant evaluation takes a further four weeks. These illustrative durations expose the difference between changing a connection and changing a working capability. The bank records both efforts, together with the conditions under which the alternative is suitable.
The alternative is acceptable for a subset of standard application formats, while some complex documents still require manual preparation. The bank approves a limited substitution route for that subset and retains a tested manual process for the remaining applications. It does not describe the alternative as equivalent across the full scope. Wider substitution requires further work and an updated assessment of cost and quality.
The exercise also reveals that a supplier-specific workflow has stored evaluation review annotations in a format the bank cannot use directly. The team creates and tests an export into the bank’s application record before relying on that workflow more broadly. An exit provision has practical value only when the information and process needed to continue the service can actually be recovered.
Distinguish continuity from strategic exit
Maintaining service during an outage and changing suppliers over several weeks are different requirements. A tested alternative with a four-week transition can support an exit strategy, but it cannot provide immediate continuity. The architecture needs a separate response to interruption, based on the consequences of delay and the capacity of fallback arrangements.
For the document-preparation service, the bank can temporarily return to manual preparation within a defined workload limit. It tests whether the available staff can process applications at an acceptable service level and specifies when new intake must be restricted. Existing customer commitments determine the response. A fallback that works at low volume cannot be assumed to support the whole population indefinitely.
The operational payment service has different requirements. A delayed or unavailable result can affect a live intervention window, so its fallback needs to operate within the relevant decision time. The production article described returning to established checks where the new capability could not provide a reliable result. That response addresses immediate continuity, while any replacement of the underlying service would require separate integration and outcome evaluation.
A second supplier also needs examination of shared dependencies. Two services may rely on the same cloud, identity system, information source, or network connection. Paying for both does not establish that they will fail independently. The bank needs to identify the failure scenarios its arrangement actually protects against, with evidence that fallback can be invoked under those conditions.
Continuous duplication of the entire capability is one option, with costs and complexity that require justification. For this lending evaluation, the bank chooses a documented substitution route and a capacity-tested manual fallback rather than maintaining two complete live implementations. That choice matches the service’s delay tolerance and current scale. A materially different commitment would require reassessment.
Make retained control proportionate to the risk
Supplier governance belongs in the service’s operating arrangements and the bank’s existing third-party risk process. As the governance article argued, AI-specific evidence should strengthen established arrangements rather than create a parallel approval structure. Substitution results, data-handling terms, and fallback tests form part of that supplier assessment. The bank needs clear responsibilities for monitoring performance, evaluating supplier changes, exercising contractual rights, and deciding when a dependency has become unacceptable. The architecture team can identify options, while the business and authorised risk owners decide which consequences the bank is prepared to accept.
The linked service record captures the supplier-dependent functions, retained assets, substitution evidence, fallback capacity, and unresolved limitations. It connects those arrangements to the permitted use and value case. This prevents an exit plan from becoming a disconnected procurement document whose assumptions do not match the application in production.
The committee reviews the completed supplier assessment and confirms continued use of the preferred document service within the bounded lending evaluation. The annotation-export fix, limited substitution route, and manual fallback have been tested for that scope. Broader commitment depends on the outcome and acquisition evidence already required by the investment case, alongside maintaining and retesting these arrangements for any expanded scope or volume. The architecture assessment does not make the lending proposition viable by itself. It establishes the dependencies and conditions under which the bank is willing to test that proposition.
Operating reviews need to revisit those conditions. A change in document mix can narrow the alternative’s suitability. Higher volumes can exceed manual fallback capacity. A supplier update can alter the correction effort that underpinned the economics. Retained control matters only if the organisation can detect these changes and exercise its options when necessary.
Before making a material supplier commitment, leadership needs answers to four questions. Which capabilities and records must we retain control of? What does the complete operating choice cost under plausible changes? What evidence establishes our route to substitution or exit? How will the service continue while that route is exercised?
The bank’s substitution exercise shows why those questions matter. Replacing an interface took two days, while replacing acceptable behaviour required four more weeks and left a narrower usable scope. A credible architecture makes that distinction visible before the organisation relies on flexibility it has not demonstrated.
