Describe the work in plain words
“We need AI” is not a project brief. “Our operations team reads incoming requests and copies the same details into a job record” is a useful start. Name the input, the decision, the destination and the person who takes responsibility. Then record what happens when the information is incomplete.
Our strategy service starts with that operational picture. The aim is a scoped recommendation, not a shopping list of models. Bring the process owner and someone who handles exceptions. Their account of missing references, duplicate requests and informal workarounds matters as much as the documented procedure.
Compare AI with a simpler answer
Use deterministic rules when the decision is explicit and stable. A required field check does not need a language model. A better form, a shared template or a conventional integration may remove the problem without adding probabilistic output. Keep that option in the comparison.
Microsoft 365 Copilot is worth evaluating when the task sits inside Microsoft 365 and an employee remains in control of the result. A custom application becomes more relevant when you need a specific approval flow, a constrained interface or connections to business systems. Check licensing, permissions and product documentation before choosing either route. Neither category automatically fits your data obligations.
Check whether the data can do the job
Inspect the material the proposed system would actually receive. Look for inconsistent terminology, scanned pages, old policy versions and fields that mean different things across departments. A demonstration built on clean examples can conceal these problems. Your pilot needs authorised examples of the awkward cases too.
Map where information originates, who may use it and where it could be sent. Customer records, employee information and commercially sensitive documents need different handling decisions. If the lawful basis, access rights or supplier terms are unclear, resolve that before uploading content. A data protection impact assessment may be needed where processing is likely to create high risks.
Define success before selecting a model
Write down what a reviewer will accept. For a draft response, that might mean factual support, the right tone and no unsupported commitments. For extracted data, it means the correct field values and a visible route for uncertain cases. A single “accuracy” score can hide the mistake that matters most.
Separate development examples from evaluation examples. Include empty inputs, contradictory documents and requests outside scope. Compare against the existing process, including review time. If a proposed assistant saves drafting effort but creates more checking work, that trade-off belongs in the decision rather than being averaged away.
Cost the whole operating process
Model usage is only one cost. Include integration work, access management, evaluation, employee training, document maintenance and incident handling. Retrieval systems may also need indexing and search infrastructure. Self-hosting shifts responsibility towards your team; it does not make operation free.
Ask suppliers how usage is measured and what happens when limits are reached. Token charges, connector licences and storage can follow different billing rules. Use the supplier’s published pricing when preparing a proposal, then document the assumptions. Avoid a precise savings forecast until you have a trustworthy baseline and a representative pilot.
Leave with a decision and an owner
An agreed strategy scope can produce a process map, a comparison of implementation options, a data-readiness assessment and a pilot brief. The brief should identify acceptance criteria, approval responsibilities, dependencies and stop conditions. You should be able to explain why the proposed work deserves to proceed.
Sometimes the right outcome is to improve the source data or change the workflow first. Sometimes an existing tool is enough. If a custom build is justified, move into workflow automation or a knowledge assistant with clear boundaries. Decide who will own the system after delivery, not after the first incident.
Bring the constraints to the first conversation
Send a short description of the task, the systems involved and the result you want. Mention procurement restrictions, accessibility needs and any requirement to keep information in a particular region. Do not send live personal data. We can agree what evidence is needed and how to exchange it before detailed assessment begins.