1. Write down the decision or task.
Describe the input, the person responsible and the result that person needs. “Use AI in customer support” is too broad to assess. “Draft replies using approved product documentation for an employee to review” identifies both a useful output and a boundary. Record what happens when information is missing, contradictory or outside the organisation's remit.
A language model generates text from patterns learned during training and the context supplied at use time. It is not a database of guaranteed facts. If the task depends on exact calculations, fixed eligibility rules or predictable routing, ordinary software may be the more appropriate starting point. Consulting should include the option not to use AI.
2. Establish the existing baseline.
Document how the task works without a model. Identify delays, repeated copying, avoidable corrections and work that already has a dependable automated solution. Select measures tied to the actual process, such as whether a draft contains the required information or whether a record reaches the correct queue. Do not substitute the number of generated messages for a business outcome.
Separate potential savings from demonstrated savings. A faster first draft may create additional checking work. A pilot should account for review, exception handling, infrastructure and ongoing support, not just the time spent waiting for a response.
3. Check the information and its permissions.
List the documents, records and systems the project would use. Establish who owns them, whether they contain personal data and whether the intended use is compatible with existing notices and contracts. Removing names is not necessarily anonymisation: a combination of job title, location and event details can still identify someone.
For UK personal data, involve the organisation's privacy lead and consult the Information Commissioner's Office guidance on AI. Consider whether a data protection impact assessment is required. A DPIA is a documented assessment of risks to people's rights; it is not a technical performance test or a substitute for a lawful basis.
4. Compare deployment approaches.
A hosted application provides an interface and an operating environment managed by a supplier. An application programming interface, or API, allows your software to request model outputs. A locally operated model gives your organisation more direct responsibility for infrastructure, updates and access controls. None of these choices establishes privacy or reliability by itself.
Compare them against the same requirements: data location, contractual terms, retention settings, access management, integration effort and the ability to change supplier. For a hosted API, distinguish the model provider from the cloud platform and from any application that stores prompts. For local deployment, include the people and equipment needed to run it securely.
5. Define a pilot with a stopping rule.
A pilot is a limited trial used to answer a specific question. Choose representative inputs, an agreed review process and clear acceptance criteria before starting. Include awkward cases as well as routine ones: incomplete records, ambiguous requests and content that should not be processed. Keep the model away from irreversible actions during early assessment.
State what would cause the pilot to stop. Examples include exposure of restricted information, failure to preserve required fields or repeated errors that reviewers cannot readily detect. A useful pilot can conclude that the proposed approach should not proceed. That is a decision, not a failed demonstration.
6. Request a practical handover.
The consulting brief should identify the proposed workflow, the information it requires, rejected alternatives and unresolved decisions. Ask for responsibilities to be recorded: who owns the data, who approves deployment and who handles a failure. Include dependencies that could change the recommendation, such as a supplier's contract or a missing permission.
Implementation should follow an agreed scope rather than an open-ended promise of transformation. If the work proceeds, connect the brief to an evaluation plan and a support arrangement. Bring a description of your task to the first enquiry; confidential files are not necessary to establish whether a more detailed discussion is useful.