Operating design

Automation, AI assistance or human judgement?

AI WorkBook editorial team22 July 20263 min read

A governed workflow does not use AI for every step. It assigns predictable work, interpretive work and consequential decisions to the approach best suited to each.

Start by decomposing the work

A process described as “review the case” may contain many different activities: checking required fields, finding relevant records, comparing evidence with criteria, identifying gaps, preparing a recommendation and making a final decision. Treating that entire bundle as one AI task hides important differences.

Map the workflow in small enough steps that the team can ask what kind of behaviour each step requires. The answer may change when the consequence, uncertainty or available context changes.

Reliable automation: when consistency is the point

Rules and deterministic automation suit work where the same valid input should produce the same result. Examples include validation, calculation, due-date generation, routing, status changes and permission checks.

Do not use a language model merely because it is available. If a rule can be stated and tested, ordinary automation is usually easier to explain and verify. This also creates stable boundaries around the more interpretive parts of the workflow.

Selective AI assistance: when interpretation helps

AI can be useful when the work involves language, variation or comparison: extracting relevant points from a document, classifying a request, summarising evidence, comparing a response with stated criteria, explaining gaps or preparing a recommendation.

The word “assistance” matters. The agent should have a defined role, approved context, permitted tools and an expected handoff. Its contribution should remain visible enough for a person to review, particularly when the source material is incomplete or ambiguous.

Human judgement: when accountability cannot be delegated

People should remain responsible where a decision changes rights, accepts risk, resolves a material exception, weighs competing interests or requires professional judgement. Human involvement should be designed into the route, not added as a vague instruction to “check the AI”.

A reviewer needs the relevant sources, the AI contribution, the stated criteria and a clear way to approve, reject or adjust the recommendation. The decision and rationale should remain connected to the work.

Use four questions for each step

• Predictability: Should the same valid input always produce the same action?
• Interpretation: Does the step require understanding variable language or comparing contextual evidence?
• Consequence: Who is affected if the step is wrong, and who is accountable?
• Reversibility: Can an error be corrected easily, or would it create a material commitment?

High predictability points towards rules. Meaningful interpretation may justify AI assistance. High consequence or low reversibility strengthens the case for explicit human judgement.

A supplier-quality example

In an 8D-style workflow, automation might check that required sections are present and route the case by status. AI might compare a supplier response with agreed criteria and explain possible gaps. A quality professional still owns severity, findings, acceptance and escalation.

The value comes from the combination. The workflow makes routine controls dependable, gives people useful interpretive support and keeps the consequential decision with the accountable role.

Review the allocation when the work changes

The original assignment is not permanent. New policies, recurring exceptions or evidence from operation may show that a rule should change, an AI task should narrow or a decision needs a different reviewer. Changes should be proposed, tested and approved rather than silently learned into live work.