How to Decide What Not to Automate with AI

Every AI program produces a list of what to automate. Few produce a list of what not to, and the omission is where trust is lost: one project at a time, boundaries that were never written down get crossed. This guide gives executives the criteria that rule work out, the work worth protecting on purpose, and how to make the decision explicit.

Why is "what not to automate" a leadership decision?

Because it defines the company's relationship with its customers and employees, and because it will otherwise be made by default. If leadership does not say which decisions stay human, project teams will decide case by case, under pressure to show results, and the boundary will move without anyone choosing to move it. FISTA's how to set an AI vision and narrative guide places boundaries alongside outcomes as one of the three things a usable vision must contain.

What criteria rule work out?

Criterion	Question to ask	Example that fails
Severe, irreversible consequence	If the agent is wrong, can we undo it, and at what cost?	Terminating a service contract; releasing funds that cannot be recalled
Unspecifiable behavior	Can experienced people write down what correct means and agree?	Judgment calls on strategic exceptions where experts disagree
Relationship value	Is the human presence the product?	Escalated complaints; key-account negotiation; crisis communication
Legal or policy reservation	Does law or company policy require a human decision?	Decisions with legal effect on individuals in regulated contexts
Weak economics	Does volume justify build and run cost, including review?	Rare, complex cases that occur a few times a year

Any one criterion can rule a process out entirely; more often, criteria carve out part of a process, and the rest is automatable. The AI use-case scoring framework covers the positive side of the decision.

What is worth protecting on purpose?

Some work should stay human even when it could be automated, because it is where the company builds what it cannot buy:

• Expertise development. If agents handle every routine case, nobody learns the craft that handles the hard ones. Keep enough routine work with people to develop the judgment the exceptions need.
• Trust moments. The complaint handled by a person who could fix it, the renewal conversation, the moment of crisis: these are where loyalty is decided.
• Accountability. Decisions the company must be able to explain and stand behind personally.
• Institutional knowledge. Processes where the people doing the work are the only record of how it really works; automate after documenting, not instead of it.

The how to redesign jobs around AI agents guide addresses how to keep the protected work in roles that remain meaningful.

Does "not automated" mean "not assisted"?

No, and the distinction is where much of the value lives. Work that stays human can still be prepared by agents: gathering context, drafting options, monitoring for signals, checking against policy, and presenting the decision to a person with everything assembled. The boundary is who decides and who acts externally. A key-account negotiation stays human; the research brief, the pricing scenarios, and the contract redline comparison can all be agent-prepared. The human-in-the-loop AI explained guide describes these assistance patterns.

How should the decision be made and recorded?

1. List candidate processes with their automatable and reserved parts.
2. Apply the five criteria to each part, with the business owner and, where relevant, legal and risk.
3. Classify each decision as permanent (legal and policy lines), conditional (revisit when evidence or economics change), or protected on purpose (with the reason).
4. Write it down in the AI policy or vision, with review dates for conditional decisions.
5. Enforce it in the platform, so agents lack the permissions to cross the lines.

The AI policy template provides a structure for the written record, and the how much autonomy should AI agents have guide covers the lines that hold regardless of evidence.

What are the common mistakes?

Automating a process because it is possible rather than because it is wise; automating before the policy is written, so the agent encodes whatever the first engineer assumed; treating the boundary as a technical limitation to be overcome rather than a decision; and never revisiting conditional decisions, so the company leaves value on the table when the evidence changes.

What should executives ask?

• What have we decided not to automate, and where is it written?
• Which of those decisions are permanent, and which should be revisited?
• What work are we protecting on purpose, and why?
• Where are agents assisting protected work without deciding?
• Has any project crossed a boundary we never wrote down?

How can FISTA Solutions help?

FISTA Solutions works with executive teams through its AI enablement practice to map processes into automatable, assisted, and reserved parts, and builds AI agents whose permissions enforce those boundaries by design. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries.

To draw the boundaries before the next agent project draws them for you, talk to FISTA on WhatsApp, or read when to use AI agents for the positive case.