Flagship engagement
Designing AI products for regulated work
AstraZeneca
A flagship engagement about making complex AI-assisted work inspectable, usable, and accountable without overstating what reached production.
- Role
- Product design and prototype implementation
Turn broad AI concepts into supervised tasks
Define the work people supervise
A regulated content-review example narrowed broad AI concepts into bounded tasks that a person could direct and review.
Stephen's contribution
Stephen contributed product design and prototype implementation, then used stakeholder demonstrations and direct feedback to test the workflow.
Keep supervision in the task model
Frame AI as bounded assistance within a task instead of an autonomous answer.
Constraint
The prototype did not prove production use, adoption, live AI, or formal acceptance.
Ask for sources and keep evidence gaps visible
Preserve what the system cannot support
The prototype simulated source-request behavior and preserved unresolved evidence in shared prototype state instead of presenting missing support as certainty.
Stephen's contribution
Stephen designed how source requests and unresolved evidence appeared in the workflow and implemented those behaviors in the prototype.
Make absence part of the evidence
Keep evidence gaps visible across the shared prototype state so reviewers could distinguish support from omission.
Constraint
The source behavior was simulated. It did not use real retrieval or create a durable audit history.
Make trust and expert judgment inspectable
Show where people evaluate the work
The interaction model gave reviewers explicit places to inspect inputs, evidence, unresolved questions, and the output before deciding what to do next.
Stephen's contribution
Stephen connected product strategy, interaction design, and prototype behavior so stakeholders could evaluate a concrete review flow.
Do not hide uncertainty behind automation
Treat AI output as material for expert review rather than an answer that bypasses judgment.
Constraint
Domain, engineering, and generation-engine work remained shared responsibilities outside Stephen's sole ownership.
Deliver reusable product language and collaboration foundations
Carry decisions into team work
The engagement produced reusable product language and prototype patterns that teams could use to discuss tasks, evidence, review, and unresolved questions consistently.
Stephen's contribution
Stephen documented the interaction model and used the prototype as a shared reference during stakeholder collaboration and handoff.
Report the handoff without overstating impact
Describe the reusable language, prototype behavior, and stakeholder demonstrations that direct evidence supports.
Constraint
Production adoption, measured impact, formal acceptance, and sole authorship of the full system remain unverified.
Outcome
What changed
Broad AI concepts became a concrete prototype for supervised tasks, visible evidence gaps, and expert review.
What reached an audience
Stakeholders saw demonstrations and provided direct feedback on the prototype direction.
What remains unverified
Production use, adoption, measured impact, live AI, real retrieval, durable audit history, and formal acceptance are not claimed here.