Reducing operational work with AI-assisted workflows
Redesigned a manual, document-heavy operational workflow using AI assistance, deterministic validation and human review.
The problem
A high-volume operational workflow relied heavily on manual document review, interpretation and data entry. The work was repetitive, but it still required judgement when source material was incomplete, inconsistent or ambiguous.
I led the product work across discovery, workflow definition, requirements, human-review design, engineering clarification, launch and evaluation. Engineering, data and design owned their implementation craft.
Accuracy mattered more than maximum automation. Ambiguous cases still required human judgement, and failures needed recoverable behaviour rather than silent failure.
The product decision
The key decision was not whether to use AI. It was how much responsibility the system should be allowed to take.
Deterministic automation only
High control and predictable behaviour
Poor fit for variable and unstructured inputs
Fully automated AI workflow
High automation potential
Too little control when interpretation is ambiguous
AI-assisted workflow with review
Combines AI interpretation with validation and human judgement
- AI interprets
- Deterministic checks validate
- Human resolves ambiguity
- System completes the action
How I approached it
Understand
Mapped the existing workflow, repetitive work and common failure points.
Define
Set the product boundary, success criteria and where automation should stop.
Design
Defined the assisted workflow, validation principles, human-review points and recovery behaviour.
Deliver + Evaluate
Worked with engineering through requirements, clarification and iteration, then measured handling time and observed usage after introduction.
What I learned
AI automation should not automatically mean removing human control. In workflows where incorrect interpretation creates downstream cost, deciding where the system should stop and request human judgement can be as important as deciding what should be automated.
The strongest product behaviour often comes from making uncertainty explicit and recoverable rather than designing as if the system is always certain.