01
Collect
Gather relevant records, messages, and required fields.
Risk and boundariesDiscovery guide
What Should You Not Automate With AI?
Good automation design includes deliberate boundaries. Some work should remain human; other work can be accelerated without handing the final decision to a model.
Decision snapshot
Do not give AI sole authority over high-stakes, regulated, irreversible, or poorly defined decisions. Avoid automating rare tasks with weak data and no owner. Instead, automate preparation: gather context, detect missing information, draft a recommendation, and route the case to a qualified person.
Consequence
High
A mistake can materially affect rights, safety, money, or reputation.
Reversibility
Low
The action is difficult to undo once performed.
Clarity
Weak
The team cannot define the correct process or result.
Governance
Missing
No owner, review queue, log, or stop mechanism exists.
Working framework
A safer alternative to full automation
High-risk work can often be supported without delegating the final decision.
02
Validate
Check completeness, format, permissions, and obvious conflicts.
03
Summarize
Prepare a structured brief with source references and uncertainty.
04
Recommend
Draft a proposed classification, response, or next action.
05
Approve
A qualified person reviews and takes the consequential action.
01
Keep high-stakes decisions human
AI output can vary with context, phrasing, model changes, and incomplete inputs. That variability is unacceptable when a decision materially affects employment, credit, legal rights, medical care, safety, or another regulated outcome without qualified oversight.
Automation can still reduce administrative work around the decision. It can organize evidence, surface missing information, schedule review, and record the approved outcome while leaving authority with the responsible professional.
02
Do not automate ambiguity you have not resolved
If the team cannot agree on what should happen, the system cannot be tested against a stable expectation. A model may produce a plausible answer, but plausibility is not an operating rule.
Document the normal path, assign ownership, and decide how exceptions are handled. Only then choose which steps are deterministic, AI-assisted, or human.
- One-off strategic decisions with no repeatable pattern
- Processes that change every time a different employee performs them
- Actions based on missing or unverified source data
- Customer-facing commitments with no approval or limit
- Destructive record changes without backups or reversal
03
Avoid automation with no operating owner
Every workflow changes over time because apps, credentials, fields, policies, and business rules change. A launch without an owner is a delayed failure.
The owner does not need to be a developer. They need to know the expected behavior, review alerts, approve changes, and understand when the workflow should be paused.
The question is not only “Can this be automated?” It is “Who is accountable when the context changes?”
Stop-and-review conditions
Keep a human in the loop when any of these conditions applies.
- 1The action affects a person’s rights, eligibility, employment, or safety.
- 2The system would send money, sign an agreement, or make a binding commitment.
- 3The source data cannot be verified.
- 4The action cannot be easily reversed.
- 5The workflow handles highly sensitive information without an approved data policy.
- 6The correct answer depends on expertise that is not represented in the input.
- 7The model would be allowed to choose arbitrary tools or recipients.
- 8No owner is available to review failures promptly.
- 9The task occurs too rarely to test with representative cases.
- 10The business cannot explain the decision to the affected person.
Continue from here
Choose the next useful step.
Move from understanding the problem to evaluating the business case or scoping a concrete implementation.
FAQ
Important questions, answered directly.
Does human approval remove all AI risk?
No. Reviewers can over-trust a confident draft, and sensitive data can still be mishandled earlier in the workflow. Approval is one control alongside data minimization, validation, permissions, logging, and testing.
Can AI send customer emails automatically?
It can for tightly scoped, low-risk messages with approved rules and fallbacks. Sensitive, unusual, financial, legal, or commitment-making replies should remain review-gated.
Should we avoid AI entirely for regulated work?
Not necessarily. AI can assist with administrative preparation, but the workflow should be reviewed against applicable obligations and keep qualified humans responsible for consequential decisions.
How do we add autonomy safely later?
Start with drafts and review. Measure accuracy by case type, document failure modes, narrow permissions, and remove approval only for actions that have demonstrated reliable and reversible behavior.
Bring one workflow, not a transformation pitch.
Describe the trigger, tools, repeated work, and where the handoff breaks. You will get a direct view on feasibility and the most sensible next step.
Request a workflow audit