Design and pilot an AI chatbot as an accountable service rather than an open-ended demonstration.
A useful chatbot completes a bounded set of reader jobs with approved information, clear identity, observable outcomes, and a reliable route to people. Start with the workflow and risk, then decide whether generative behavior improves it.
Choose a narrow first job
List the questions or transactions that are frequent, well documented, reversible, and safe to automate. Separate answering, collecting, looking up, changing, committing, and advising. Each verb carries different authorization, data, accuracy, and escalation needs. Exclude high-impact decisions until governance and evidence justify them.
Define the operating contract
- User population, channels, languages, accessibility, and disclosure
- Approved knowledge sources, owners, refresh timing, and citation behavior
- Data collected, purpose, retention, permissions, and sensitive-data boundaries
- Allowed actions, authentication, confirmation, and transaction limits
- Uncertainty, refusal, escalation, outage, and abuse handling
- Quality metrics, sample review, incidents, changes, and rollback owner
Match automation to task risk
| Task | Safer starting design | Escalate when |
|---|---|---|
| Stable FAQ | Grounded answer with source | Source missing or conflict |
| Guided intake | Collect minimum structured facts | Sensitive or exceptional case |
| Account action | Authenticated deterministic workflow | Identity or confirmation fails |
Build the pilot around evidence
- Baseline current volume and outcomes.
- Select approved intents and explicit exclusions.
- Prepare and test the knowledge set.
- Design escalation with transcript and context.
- Run adversarial, accessibility, and failure tests.
- Launch to a bounded cohort and review sampled outcomes.
Plan for predictable failure classes
- Confident unsupported answers
- Prompt injection or untrusted retrieved content
- Sensitive data copied into logs or model inputs
- Automation trapping a user who needs a person
Operate changes as releases
Version prompts, policies, models, knowledge sources, tools, and routing. Compare success, escalation, correction, abandonment, latency, accessibility, and incident measures before and after each material change. Preserve representative transcripts according to a documented privacy and retention basis.
Continue with the next decision
Sources and further reading
Primary and contextual sources used to verify definitions or give readers a relevant next resource.
- NIST AI Risk Management Framework Primary voluntary framework for governing, mapping, measuring, and managing AI risks.