1. Define one task and the human decision
Write down the input, expected output and the action that follows. "Sort incoming enquiries into service categories for a person to review" is a workable starting point. "Automate our business" is not a specification. Identify which decisions must remain with a person, such as committing to a price, changing a booking or responding to a sensitive complaint.
Use a synthetic enquiry during planning: a potential customer asks about a service in Hindi and requests a callback. The pilot may suggest a category and prepare a draft response. A staff member confirms the service details before any message is sent. This example does not establish that a particular product already offers the workflow.
2. Map the data before choosing integrations
List where enquiries arrive, what information they contain, who can access them and where any model request will be sent. Remove identifiers that are unnecessary for the task. A routing test usually does not need a real customer's full message history, payment information or identity documents.
Keep development examples separate from live business records. Decide who may connect accounts and how access will be removed when a team member leaves. Record the systems involved so the business owner can understand the arrangement without tracing code. Treat applicable privacy and contractual requirements as a separate review, not something an AI tool automatically solves.
3. Check language quality with a real reviewer
Prepare synthetic examples in the languages your customers use. Include short messages, unclear requests and mixed Hindi-English phrasing where relevant. Ask a fluent reviewer to assess meaning, tone and missing context. A model's supported-language list is not evidence that every answer is suitable for your customers.
For each example, decide what a correct result looks like before evaluating the output. Record incorrect routing, invented facts and inappropriate promises. Keep the examples consistent when comparing configurations so the decision is based on the same work.
4. Agree on costs and acceptance criteria
Create a simple worksheet with separate rows for model usage, hosting, messaging or integration charges, maintenance and applicable taxes. Add a volume assumption and spending limit. Do not compare a model's token price with the total operating cost of the system.
Acceptance criteria should describe observable behavior: the correct category is proposed, unsupported requests are escalated and a failed provider call does not send an incorrect message. Set a manual fallback and a named person to handle it. Avoid promising a percentage saving before a representative pilot has been measured.
5. Pilot, review and hand over
Run the first tests in an isolated environment. Review failures before expanding the task or connecting more systems. Approve a limited production rollout only after the owner understands what the system does, what it cannot do and how to pause it.
Before handover, confirm the operating owner, access inventory, support contact, spending controls, recovery procedure and known limitations. Keep an ordinary manual route available when automation is unavailable. Review the workflow again when business rules, providers or language requirements change.
This checklist is planning guidance, not a compliance certification, security assessment or guarantee of savings. See how an engagement works, explore custom AI solutions, or discuss your workflow. For team review steps, read the social content approval guide.