Define the decision.
Then select the AI.

1. Describe the task without naming a model

Start with the action a person needs to complete. “Prepare a reply using an approved product guide” gives a clearer starting point than “introduce a chatbot”. Identify the input, the intended output and the person accountable for the result. Include exceptions: a missing document, an ambiguous request or information that contradicts another source.

Next, compare AI with a simpler alternative. A fixed eligibility rule usually belongs in ordinary software. A search problem may need better indexing rather than text generation. Generative AI, which produces new content from patterns learned during training, is more relevant when inputs vary and the output requires interpretation. The project brief should explain why that flexibility is needed.

2. Inspect the information the task depends on

List the systems that hold relevant material and identify their owners. For a UK organisation, this may include Microsoft SharePoint documents, a CRM export, support tickets and records containing personal data. Check whether the organisation has permission to use each source for the proposed purpose. Access to a file does not automatically justify sending it to an external model provider.

Inspect representative material for missing fields, conflicting instructions and obsolete content. Record whether documents are searchable text or scanned images. Separate a data-quality problem from a model-selection problem: replacing a model cannot resolve a policy document that has no clear owner. Where personal data is involved, include a privacy review before choosing hosting or retention settings.

3. Compare approaches against operating requirements

A hosted application programming interface, or API, provides model access through a supplier-operated service. Self-hosting an open-weight model gives the organisation responsibility for deployment and infrastructure. Open weights do not necessarily mean unrestricted licensing; review the particular model’s terms. Compare these approaches on data handling, maintenance, access control and the skills available internally.

OpenAI and Anthropic publish API documentation; Microsoft Azure provides documentation for its managed AI services. Their names alone are not selection criteria. Examine the specific service contract, deployment region, logging behaviour and access to features your task requires. Check changing commercial terms directly with the supplier. Avoid choosing solely on a public benchmark that measures a different task.

4. Define how outputs will be judged

Create an evaluation set from examples you are authorised to use. Include routine inputs, difficult exceptions and requests the system should decline. Write down what counts as an acceptable result before reviewing outputs. For a drafting assistant, criteria might include factual support, completeness, appropriate tone and whether it adds claims absent from the source.

Keep correctness separate from fluency. A polished answer can still be wrong. Ask reviewers to record the reason for rejection, not just a general preference. Compare the proposed approach with the existing process using the same inputs. Include the human review effort and the cost of handling errors when deciding whether automation is useful.

5. Agree the decision and the next responsibility

A consulting engagement can be scoped around a project brief, a data dependency list, an approach comparison and an evaluation plan. The written scope should identify which artefacts are included and which implementation work requires a separate agreement. It should also identify dependencies that may prevent a build, such as unavailable APIs or unresolved permissions.

Before proceeding, name the business owner, technical maintainer and reviewer responsible for release. Document a stopping condition as well as a success condition. A useful outcome may be a decision to improve records first, use conventional software, or limit AI to drafting. Proceeding with a model is one possible conclusion, not the purpose of the assessment.

For technical preparation, consult OpenAI’s API documentation, Anthropic’s documentation and the ICO’s AI guidance. Use supplier documentation to check product details and qualified advice for legal questions.