Choose a bounded use case
Define the useful staff outcome: a draft from permitted knowledge, document-field extraction or a similar-request search. Ordinary software is preferable for exact calculations and deterministic rules. Limit topics and actions rather than judging quality from a few successful conversations.
Prepare sources and permissions
Review provenance, usage rights, currency and access. Server-side retrieval must respect each user’s document permissions. Assess personal and confidential information before sending anything to an external API. Supplier retention, processing locations and deletion terms require service-specific review.
Make review part of the interface
Show sources and distinguish quoted evidence from generated conclusions. Staff can open documents, correct drafts and report mistakes. Uncertainty needs a clear fallback. Sending messages, changing records or financial actions require agreed approval controls. Retrieved text is not authority to execute commands.
Evaluate errors explicitly
Create representative tests with ordinary, ambiguous, outdated and out-of-scope questions. Evaluate factuality, extraction completeness, grounding and access boundaries. Repeat checks after model, instruction or knowledge changes. Critical errors should be reviewed separately rather than hidden in an average score.
Limit costs and failure impact
Agree input limits, request frequency, logs and retention. Budget for API usage, retrieval, storage and human review. Preserve ordinary search or staff escalation when the model fails. Assign ownership of source updates and review misleading answers as well as technical exceptions.
Release in stages
Start with internal users and reversible tasks, then expand after error review. Handover includes source management instructions and known limits. Independent system actions require a separate risk decision.
- Source rights are confirmed.
- Cross-role documents remain inaccessible.
- Unknown questions appear in evaluation.
- People approve consequential actions.
Project questions
Can AI guarantee a correct answer?
No. Plausible language can still be wrong. Grounding checks, evaluation and human involvement are needed where mistakes have consequences.
Can internal documents be used?
Yes, with permission and an agreed processing model. Review data categories, staff access and provider terms before connecting them.
Must we train our own model?
No. Existing models with an authorised knowledge source may suit search and drafting. Custom training needs a justified task, data and maintenance resources.
Illustrative image