Run useful reference checks for an AI agency

Ask for relevant context

Find out what type of work the reference can discuss and how similar it is to your planned engagement. Was the agency advising, building, integrating, or supporting? A strategy workshop and an operational application reveal different capabilities. Respect the customer’s confidentiality and time. You need enough context to interpret the experience, not private architecture details or sensitive business records from another organization.

Explore how uncertainty was handled

Ask what changed during delivery and how the agency communicated the consequences. Did the team make dependencies visible, identify decisions early, and preserve a usable scope? A project without any reported difficulty may simply be too different or too lightly discussed to inform your selection. Listen for specific examples of working behavior rather than relying on broad satisfaction language.

Discuss the period after delivery

Ask whether the receiving team understood the system and knew where to take problems. Explore the usefulness of documentation, training, and support arrangements. NIST’s secure development framework can inform your own lifecycle questions, but a reference should be asked about what it actually observed. Do not imply that the customer is certifying engineering practices it did not inspect or that were outside its engagement.

Reference: NIST: Secure Software Development Framework

Combine references with current evidence

Record what the conversation supports and what remains unverified. Agentix should be evaluated through the same diligence as other candidates, with any relevant examples shared only where disclosure is authorized. If a reference is unavailable, decide what alternative evidence is appropriate, such as a bounded technical exercise or stronger acceptance stage. Avoid inventing certainty from a logo list or an unattributed testimonial.

Reference: Agentix (publisher): Agentix services

Common questions