Service / AI strategy

Define the work before the model.

AI strategy turns an open-ended interest in artificial intelligence into a decision about a particular task, a bounded pilot and the conditions for further investment.

Abstract computational node with separated pathways representing alternative AI project choices
Choose a route by the work it must support, not by the size of the model.

1. Describe one task precisely

Start with an action rather than a department. “Help customer service” leaves too much unresolved. “Draft a reply from an approved returns policy for a person to review” identifies an input, an output and a decision-maker. Record where the information comes from, which system receives the result and what happens when information is missing.

Include the present process. A project cannot demonstrate improvement without a reasonable comparison. Describe manual effort, review requirements and common failure types. Separate avoidable rework from work that exists for a good reason, such as checking a contractual commitment before communicating it to a customer.

2. Compare AI with simpler software

Rules-based software follows explicit instructions. It is often suitable for required fields, routing by known categories and exact calculations. Predictive models learn patterns from examples and can support tasks such as classification or forecasting, provided suitable data exists. Generative AI produces new text or other content; a plausible response is not necessarily a correct one.

Use the simplest approach that meets the requirement. A database query may answer a stock question more reliably than a language model. A template may handle a standard notification. Generative AI becomes more relevant when language varies and a human can judge the result before it causes a consequential action.

3. Establish the constraints early

List restrictions before choosing a supplier. These may include document permissions, UK data protection obligations, hosting requirements, integration access and a requirement to retain human approval. Distinguish contractual commitments from internal preferences so that the selection process does not treat every request as equally binding.

Compare managed APIs, meaning application programming interfaces supplied as a service, with models operated in your own environment. Managed services reduce some infrastructure work but require scrutiny of contracts and data handling. Operating a model yourself provides different controls while adding deployment, security, capacity planning and maintenance responsibilities. Neither route removes the need to evaluate outputs.

4. Design a pilot that can fail usefully

A pilot is a limited implementation used to answer a defined question. Select representative inputs, including difficult cases, and keep a separate evaluation set that is not used to tune the system. Agree what makes an answer acceptable before seeing the results. Criteria might include source accuracy, correct escalation and whether a reviewer can complete the task without repairing substantial errors.

Do not measure only polished demonstrations. Include empty documents, conflicting policies, ambiguous requests and attempts to make the system disregard instructions. Record failures by type. A pilot that reveals unsuitable data or excessive review effort can still support a sound decision to stop or change direction.

5. Account for the full operating cost

Supplier usage charges are only part of the cost. Include integration work, access management, monitoring, document updates, evaluation and the time needed for human review. Request pricing information directly from suppliers because model charges and contract terms change. Estimate usage from your own workflow rather than borrowing another organisation’s volume.

Check the cost of unsuccessful work as well. Retries, lengthy prompts and repeated document retrieval can increase resource use. A cheaper model may require more review; a more capable one may be unnecessary for a narrow extraction task. Compare total effort and acceptable outcomes rather than a single advertised price.

6. Put the decision in writing

A strategy engagement can be scoped to produce a task shortlist, a documented comparison of implementation options, a pilot specification and a decision record. The exact deliverables belong in the project agreement. Each recommendation should identify assumptions, dependencies and an owner who can decide whether the evidence is sufficient.

For a first enquiry, bring a process description, the names of relevant systems and your main constraint. Do not send confidential records. If the main obstacle is unclear or inconsistent information, begin with data readiness. If the task is already defined, workflow automation may be the more direct next step.