Which Tasks Suit AI Agents? 5 Questions to Decide

Core Principle: Task Fitness = Verifiable Acceptance × Controllable Risk

The article argues that the key question is not whether an agent can do a task, but whether we can clearly judge if it did the task correctly after it finishes. If we cannot verify the result, stronger agents only amplify errors faster.

Task fitness is defined as the product of two factors: the result can be accepted (verified) and the risk can be controlled. "Can do" only means capability coverage; "fit to delegate" requires that the result can be checked, errors can be discovered, permissions can be limited, and problems can be corrected in time.

Four Dimensions of Verifiable Acceptance

True acceptance is not a glance at the result after the fact. It must be defined before the agent starts: what to check, what criteria to use, what counts as failure, and how to handle errors. The article breaks this into four requirements:

Visible result : The deliverable has a clear form — report, table, code, email draft, operation log, or completed system state.

Judgable criteria : Pass/fail conditions can be written in advance — factual sources present, fields complete, tests passing, format matching template.

Traceable evidence : The conclusion leaves a trail — citation links, calculation steps, test logs, file diffs, operation records.

Reversible errors : Mistakes can be found and corrected — re-run, restore files, revert changes, or stop before external impact.

A research report without sources, code without tests, or a system change without logs all have results but lack a reliable acceptance loop.

Four Collaboration Modes Based on Verifiability and Risk

Plotting "easy to verify" against "low risk if wrong" yields four quadrants:

A / Verifiable × Low Risk → Full delegation to agent . Examples: organizing data by fixed fields, building comparison tables from official sources, generating weekly reports from templates, modifying code and running existing tests.

B / Verifiable × High Risk → Agent executes, human gives final confirmation . Result is checkable but execution affects customers, money, permissions, or production. Examples: drafting client emails, identifying contract risks, preparing expense reimbursements, generating system change plans. Agent stops at "awaiting confirmation", not "executed".

C / Hard to Verify × Low Risk → Agent assists exploration . No single correct answer, but error cost is low. Agent expands options and provides drafts; human selects. Examples: brainstorming, headline ideas, visual direction exploration, event concepts.

D / Hard to Verify × High Risk → Do not outsource core judgment . Standards depend heavily on human values; errors are major or irreversible. Examples: medical diagnosis and treatment decisions, major investment decisions, hiring/firing, legal liability determination. Agent can organize information, surface gaps, generate question lists, but responsibility stays with humans.

The article stresses: not every task should aim for full automation. Good agent collaboration starts with choosing the right level of automation.

Redesigning Tasks to Make Them Verifiable

Task fitness is not a fixed property. The same work, redefined, can shift from "hard to judge" to "verifiable". Two examples:

Original : "Research recent Agent products." → Redesigned : Limit time range, use only official sources; build candidate list, pick three cases; each case covers updates, target users, value; deliver comparison table, narrative, and source list. Verification becomes: dates checkable, sources openable, count verifiable, structure inspectable, conclusions traceable.

Original : "Reply to this customer for me." → Redesigned : Read conversation history, draft a reply that commits to no price and no deadline; list cited facts and missing info; save as draft awaiting human confirmation before sending. The agent moves the work to a confirmation point instead of making the decision.

When a task seems unfit for direct delegation, don't immediately reach for a stronger model. First try adding evidence requirements, checkpoints, permission boundaries, and human confirmation nodes.

Quick Check: Five Questions Before Delegating

No complex scoring needed. Ask these five:

Deliverable : Will the task leave a checkable file, state, or result?

Standard : Can you write at least three explicit pass conditions before the agent starts?

Evidence : Can sources, tests, data, or logs prove the result is reliable?

Correction : If the agent errs, is the problem easy to spot, and can the result be redone or rolled back at low cost?

Permission : Can you grant only the data, tools, and operation rights strictly needed?

Quick judgment :

5 yes → Delegate fully, verify against checklist.

3–4 yes → Agent does the bulk; human checks key nodes and final result.

0–2 yes → Redesign the task first, or let agent only provide materials and alternatives.

Veto line : If the task involves irreversible operations, large sums, external commitments, personal safety, or high-risk decisions that are hard to hold accountable, keep human final confirmation regardless of the score.

Mature Agent Use Is Not About Removing Humans

It's easy to equate agent value with reducing human involvement. But mature use puts human attention where it matters most: defining success criteria, watching for anomalies, handling exceptions, confirming high-risk actions.

The article cites Anthropic ("Building effective agents", "Demystifying evals for AI agents") and OpenAI ("A practical guide to building agents") as emphasizing clear success standards, feedback loops, effective human oversight, tool boundaries, guardrails, and human-in-the-loop nodes calibrated to risk.

Don't ask "Can the agent do this?" first. Ask "After it's done, how do I prove it did it right?" If you can answer that, even complex tasks can be gradually delegated. If you can't, even simple tasks shouldn't be handed over blindly.