From AI demo to useful workflow: what to decide before production

01 / Choose a task with a visible before and after

Start with a workflow people already perform. Document its trigger, inputs, steps, output and owner. A focused first project might draft an internal summary from approved documents or suggest a category for an incoming request. “Add an AI agent” leaves too many decisions open to evaluate whether the result is useful.

Record the current effort and the main source of mistakes or delay. Then decide which part requires interpretation and which part follows dependable rules. If the task is entirely predictable, a conventional integration may be easier to operate. If interpretation matters, define what the AI should produce and who can judge the quality of that output.

02 / Define the information boundary

List the sources the workflow may use, who owns them and which users may access each one. Decide how current the information needs to be and what happens when a source is missing, contradictory or out of date. A polished answer is not useful if the underlying document no longer reflects the business.

Treat content retrieved from documents, websites or messages as information to examine. It should not gain authority to change the system’s permissions or instructions. OWASP identifies prompt injection as a risk where malicious or misleading inputs influence model behavior. Limit access and available actions, and test attempts to cross those limits using realistic examples.

Reference: OWASP Gen AI Security Project: Prompt Injection ↗

03 / Build an evaluation set before polishing the demo

Collect representative examples with permission, or create clearly labeled test cases that reflect the real work. Include routine requests, ambiguous inputs, missing context and cases the workflow should decline or send for review. Write down the expected behavior and the reasons an output would be unacceptable.

Evaluate more than whether the answer sounds convincing. Check factual support, source relevance, completeness, handling of uncertainty and time to produce a usable result. For a classification task, examine the consequences of different kinds of errors. For a drafting task, measure how much human correction is still needed. Keep these cases so changes to prompts, models or data can be compared consistently.

• Typical cases drawn from the intended workflow.
• Incomplete, conflicting and out-of-scope inputs.
• Attempts to reveal information outside the user’s permissions.
• Unavailable sources and connected-service failures.
• Expected escalation behavior and examples of unacceptable output.

Explore AI evaluation, access controls and release planning in Digital Products & AI ↗

04 / Separate a suggestion from an authorized action

Drafting a reply, sending that reply and changing a customer record are different capabilities. Define them separately. A useful pilot may propose an action while a person reviews the exact recipient, content and affected record. The interface should make that distinction obvious and preserve enough context for a meaningful decision.

Choose a narrow set of permitted actions and validate their inputs in ordinary application code. Record the operation and handle retries so a temporary failure does not create duplicate work. Make it possible for an operator to pause the workflow and return to the manual process. The review design should reflect the consequences of an incorrect action, rather than the confidence of the generated wording.

05 / Plan the operating model before release

Estimate running costs from expected volume, model usage, retrieval, storage and any connected services. Name the person who receives failures and the person who approves changes. Define what is logged, how long records are kept and what information should be excluded from logs. Start with a limited release so behavior can be observed before expanding the scope.

NIST’s AI Risk Management Framework organizes ongoing work around governing, mapping, measuring and managing AI risks. For a practical project, that means keeping responsibilities, context, evaluation and improvement connected throughout its life. Revisit the evaluation set when the task or data changes. A production workflow should remain understandable to the people responsible for it after the initial demonstration is over.

Reference: NIST: AI Risk Management Framework ↗