Automation & AIModule 4: Tools, safety and operationsLesson 15 of 15
Course progress93%
17 min lesson · Updated August 2026
What should not be automated?
Do not automate when the process is unclear, errors are high-impact or hard to detect, data or authority is missing, exceptions dominate, empathy or accountable judgment is central, or the likely value does not justify operational risk.
What you will learn
By the end, you will understand:
Recognize red flags before building
Choose manual, assisted or constrained alternatives
Define stop and rollback criteria for existing automation
Visual explainer
See the idea clearly.
✓
Candidate tasks pass through clarity, frequency, data, consequence, detectability, exceptions, accountability and value gates; safe tasks proceed, mixed tasks receive assistance, and high-risk unclear tasks stop.
Some work should remain human
Red Flag
Reason
Unclear process
Automation scales ambiguity and conflict.
Rare high-stakes decision
Little repeatability and large consequence make open automation hard to justify.
Exceptions are the work
The task depends on context, negotiation or judgment rather than a stable normal path.
No reliable source/feedback
The system cannot know or verify the correct outcome.
Empathy and trust are central
Bereavement, crisis, complaint or vulnerability may require human presence.
Unclear accountability
No one can approve, monitor or answer for the result.
High-impact decisions need stronger limits
Employment, credit, housing, education, healthcare, insurance, legal, safety and access decisions can affect rights and livelihoods. Laws vary, but a general design principle is clear: do not let opaque automation silently make or communicate consequential decisions without appropriate governance and human responsibility.
Human review must be meaningful—not a person clicking approve under impossible volume with no authority or information.
Use assistance instead of autonomy
Manual
Person performs the work with standard templates/checklists.
Assisted
System gathers evidence, calculates or drafts; person decides and communicates.
Constrained automation
System handles low-risk normal cases; exceptions and consequential actions require approval.
Autonomous
System acts within broad bounds; justified only when controls, evaluation and consequence support it.
The automation decision
01
Is the process clear?
02
Is it repeated/stable?
03
Are inputs lawful/reliable?
04
Can correctness be measured?
05
Are errors detectable/reversible?
06
Can exceptions escalate?
07
Is there an accountable owner?
08
Does total value exceed risk/cost?
Stop an existing workflow when
Unexpected harmful action
Credential/data exposure
Error rate exceeds threshold
Source becomes unavailable/stale
Policy or law changes
Human escalation fails
Cost/loop exceeds limit
Vendor behavior changes materially
Users cannot contest/correct
Owner leaves with no handover
Sunset cleanly
Disable triggers and credentials, drain or preserve queues safely, notify affected teams/users, export required records, remove unnecessary data and monitor for late events. Keep an incident/change record.
Do not leave abandoned webhooks, API keys or scheduled jobs running after a pilot ends.
Saying no can be the best design
A simpler form, clearer policy, better staffing or small deterministic rule may solve the problem more reliably than an AI agent. The goal is a better outcome, not maximum automation.
Revisit the decision if the process, data, controls or consequence changes. “Not now” is a valid evidence-based conclusion.
Real-world example
Example: bereavement complaint stays human
Example
A utility considers an AI bot to handle bereavement account closures. Because identity, legal authority, grief and unusual account relationships require care, automation only collects minimal scheduling information and surfaces policy; a trained person owns the case.
Try this
Write a no-automation case
Choose a proposed automation. Score clarity, frequency, data, consequence, detectability, reversibility, exceptions and ownership. If any critical gate fails, propose a safer manual or assisted alternative.
Common questions
Questions beginners ask.
What should not be automated?
Unclear, rare, high-impact, exception-heavy or empathy-dependent work without reliable data, verification and accountability.
Does human review always make AI safe?
No. Review must be informed, independent, resourced and empowered, and some uses remain unsuitable.
What is meaningful human review?
A qualified person has enough context, time and authority to challenge or change the automated result.
When is assistance better than autonomy?
When automation can gather or draft safely but the final judgment or communication needs a person.
Should every business use an AI agent?
No. Rules, better processes or manual work may be safer and more valuable.
What is reversibility?
The ability to undo or compensate for an action without unacceptable harm.
When should a workflow be stopped?
When safety, privacy, accuracy, cost, source, policy, escalation or ownership crosses a defined threshold.
How should a pilot be closed?
Disable triggers/credentials, handle queued work, export/delete data appropriately and document the result.
