Most organisations do not need an AI strategy as a poster. They need a way to put a model into a job that already consumes time, without losing control of the data or the result. That is a workflow problem first and a model problem second.
Pick a task, not a capability
Useful starting points are narrow. Summarise this type of inquiry. Extract these fields from this type of document. Draft a reply that a person will edit. Retrieve the policy that matches this question. Each of those has an input, an output, and someone who currently does the work by hand.
Decide what the model is allowed to see
Before any prototype is connected to production data, write down the information the task requires. Then remove everything else. Customer identifiers, credentials, and unrelated documents do not belong in a prompt because they are interesting. They belong only if the task cannot be completed without them.
Keep a person on the path that matters
If the output can change a customer record, a payment, or a public reply, a review step is not a lack of ambition. It is the difference between an assistant and an unattended process. Many teams discover that the review interface is the product: a person can confirm, correct, or reject a result in seconds if the extraction is structured well.
- ▸Write the result to a holding state, not straight into the system of record.
- ▸Show the source and the proposed values together.
- ▸Record who accepted a change.
- ▸Let low-confidence results skip the fast path automatically.
Introducing AI this way is slower than launching a chatbot. It is also how the tool becomes part of operations instead of a demonstration that staff walk around.
