What should stay human in an AI workflow today?
A practical framework for deciding which parts of an AI workflow should stay human, from irreversible actions to ambiguity, exceptions, and changing thresholds.
Niels Kaspers
TL;DR
Humans should stay on the parts of a workflow that set scope, approve irreversible actions, resolve ambiguity, handle exceptions, and decide when the rules themselves need to change.
If you want the short answer, humans should stay on the parts of the workflow where the cost of being wrong is asymmetric.
That usually means five things:
- setting the scope of the job
- approving irreversible or public actions
- resolving ambiguity the model cannot ground cleanly
- handling exceptions that fall outside the normal path
- changing the rules that govern future runs
Everything else is negotiable.
That is the practical frame I trust most.
Why this matters now
The agent conversation has shifted. The interesting question is no longer whether models can complete multi-step work at all. The harder question is which decisions still need human judgment even when the workflow is useful.
OpenAI's practical guide to building AI agents keeps treating orchestration, guardrails, and human oversight as core design work. Anthropic's evaluation guidance reinforces the same point from another angle: define success outside the model and test the workflow against real failure modes.
web.dev's Build agent-friendly websites adds a useful product-side reminder. The systems doing the work depend on clear structure, visible actions, and stable semantics before autonomy becomes trustworthy.
That matters because the boundary between automation and judgment is now a design choice, not a philosophical debate. The teams that make it explicit usually keep more speed without losing trust.
That is why "human in the loop" is too vague to be useful on its own.
The real question is where the human belongs.
1. Humans should set the scope
A workflow gets weird fastest when nobody can explain what it is supposed to own.
A human should still decide:
- what job the workflow has
- which artifact it is supposed to produce
- where it should stop on purpose
That is the same operating concern behind What is loop engineering. Good loops are narrow enough to be auditable. Bad loops keep accumulating capabilities because nobody defined the boundary clearly.
2. Humans should approve irreversible actions
This is the easiest line to draw.
Publishing, spending money, messaging a customer, changing production data, or merging code should not feel equivalent to summarizing notes or gathering sources.
That is why Human approval is the missing layer in most AI agent workflows still matters. The point is not to slow everything down. The point is to reserve attention for the actions that are expensive to unwind.
3. Humans should own ambiguity that changes the decision
Models can often summarize uncertainty.
They are much less reliable at deciding which interpretation matters when the evidence is incomplete, contradictory, or politically sensitive.
That means a human should stay involved when the workflow hits:
- conflicting source evidence
- unclear strategic tradeoffs
- weak signals dressed up as confidence
- cases where the right answer depends on context outside the workflow
This is usually where teams confuse fluency with judgment.
4. Humans should handle exceptions that rewrite the rulebook
The default path can often be automated.
The moment an exception implies the workflow rule itself should change, a human needs to step back in.
That could mean a broken data source, a new approval threshold, a changed business rule, or a new compliance concern.
If the workflow keeps adapting itself to edge cases without review, it starts collecting agent debt faster than it collects value. That is why How to audit an AI workflow before it turns into agent debt is the companion page here, not a separate topic.
5. Humans should decide when the success criteria change
The workflow can check whether it passed a rule.
A human should still decide whether the rule is still the right one.
This is where evals become practical. The metric, threshold, or approval packet that made sense a month ago may not match the current risk or business goal.
If nobody revisits that layer, the workflow keeps succeeding against an outdated definition of success.
That review layer is easy to skip because it does not feel urgent during a good run. It becomes expensive only later, when the workflow is confidently optimizing for a target nobody meant to keep.
A checklist I would use
Interactive
Human-in-the-loop checklist
Use this when an AI workflow is useful, but you need to decide where human judgment still belongs.
Completion
This is the gap between understanding the article and actually using it.
- Use this block as the practical summary, not just the article ending.
- If one item feels vague, the article probably needs sharper guidance.
- A short checklist beats a long recap when the reader needs to act.
My take
The goal is not to keep humans busy inside every workflow.
The goal is to keep humans where judgment compounds.
If the model is doing sorting, drafting, or collecting evidence, great. If it is changing the scope, rewriting the rules, or acting publicly without a meaningful review layer, that is where trust starts to break.
That is the line I would protect first.
FAQ
What should stay human in an AI workflow?
Humans should usually keep scope-setting, approval of irreversible actions, ambiguity resolution, exception handling that changes the rules, and ownership of success criteria.
Should every AI workflow have a human approval step?
Not every step needs approval. The review layer matters most when the action is public, expensive, irreversible, or hard to recover from.
Why is ambiguity a human job in AI workflows?
Because the model may present a clean answer even when the evidence is incomplete or conflicting. A human is still better at judging which uncertainty matters.
Can AI handle exceptions on its own?
It can handle known exceptions when the rule is already defined. When the exception implies the rule should change, a human should step in.
What is the risk of automating too much?
The workflow gets easier to run but harder to trust. That is usually how agent debt starts.