Describe the outcome, not the wish to “add AI”
Start with a concrete result: reduce proposal lead time, classify requests faster, lower document rework or release specialists from repetitive search. Record the baseline cycle time, volume, cost, defects, queues and manual corrections. Without a baseline, a pilot demonstrates activity rather than value.
Break the process into events, decisions and role hand-offs. Mark where a person searches for evidence, drafts content, compares options, takes a legally or financially significant decision and confirms an action in a system of record.
Give AI a bounded role
AI may extract facts, propose a category, prepare a draft, identify contradictions or invoke an approved tool. These are different roles with different risks. For each one, document inputs, permitted actions, expected output and stop conditions.
Keep people in decisions that involve authority, accountability, exceptions or material consequences. Review must be meaningful: show sources, confidence, changed fields and the reason for escalation so that a specialist can accept or reject a proposal efficiently.
Design the flow around exceptions
A poor design makes a person recheck every AI response and adds another screen. A good design completes a routine, safe case and sends uncertainty to the right role with full context. Define exception queues, response targets, substitutes and a manual fallback for model outages.
Avoid binding the workflow to one model. Prompts, knowledge bases, policies and models should be versioned and replaceable independently, while a critical action remains available through a controlled fallback.
Measure the complete process
Do not stop at answer accuracy. Track end-to-end time, completed outcomes, returns, manual edits, escalation rate, cost per useful result and user satisfaction. Check whether work is merely displaced downstream or hidden review effort grows.
Scale only after a stable observation window, edge-case testing and process-owner approval. The pilot deliverable is a revised operating procedure, a regression set, metrics and a fallback plan—not merely access to a model.
Put the process under management
Assign a business outcome owner, technical owner, data owner and explicit risk acceptance role. Maintain an inventory of AI use cases, deployed versions, permitted data, incidents and continue-or-stop decisions.
The practical readiness test is simple: an employee understands what the AI did, the organisation can reconstruct the basis of a decision, and the workflow still operates when the model is wrong or unavailable.