AI workflow design · 9 July 2026 · 8 min read
When not to use an AI agent.
An agent can plan, choose tools, and take several steps toward a goal. That flexibility is also its cost: more possible paths, more permissions, and more ways to fail. Many workflows need something smaller.
The goal is not to maximise AI. It is to improve a real workflow with the least unnecessary complexity.
Agent is an architecture choice
Do not use “agent” as a synonym for automation.
A scheduled export is automation. A validation rule is automation. A fixed sequence that retrieves a record, fills a template, and asks for approval is automation. An agent becomes relevant when the route itself must vary: it interprets a goal, selects among tools or steps, observes the result, and decides what to do next.
Rule
Best when inputs and decisions are stable, explicit, and fully testable.
Assisted workflow
Best when AI handles one bounded language or retrieval step inside a controlled sequence.
Agent
Potentially useful when legitimate cases require different paths and the choices can be constrained and evaluated.
Nine stop signs
Reasons to choose a simpler path—or no build.
One stop sign may be enough. Several together usually mean the workflow needs redesign before technology selection.
The answer is a stable rule
If the same inputs should always produce the same output, ordinary code or a workflow rule is easier to test, explain, and maintain.
The process should disappear
Do not automate an unnecessary report, duplicate approval, or broken handoff. Remove or simplify the step first.
The burden is too small
A rare task with modest effort may never recover the build, review, monitoring, and support cost. Measure before investing.
No one owns the outcome
Without an accountable owner, exceptions linger, success drifts, permissions accumulate, and changes have no decision-maker.
There is no credible test
If the team cannot describe acceptable output or assemble representative cases, it cannot know whether the agent improved the work.
The data path is not viable
Missing authority, poor source quality, unclear permissions, or no realistic sample can block the workflow before model choice matters.
Failure cannot be contained
If a wrong tool choice or action cannot be detected, paused, reversed, or escalated, autonomous execution is the wrong starting point.
The first use case is a high-impact decision
Do not begin by letting an agent autonomously make medical, hiring, fraud, legal, crisis, or similarly consequential decisions. Keep accountable human judgement at the boundary.
An existing feature already solves it
Check the current platform, integration options, and standard products. A custom agent must earn its extra operating surface.
The five-question test
Make the workflow earn complexity.
Answer these questions in order. A “no” points to the work that must happen before an agent experiment is justified.
Is it worth fixing?
There is meaningful, repeated burden or quality loss, measured against the current process.
Must the path vary?
Legitimate cases require different steps; a form, rule, or fixed integration is insufficient.
Can choices be bounded?
Tools, data, actions, spending, recipients, and stopping conditions can be constrained.
Can it be judged?
Representative normal, edge, and failure cases can be scored with useful acceptance rules.
Question 5
Can a person recover it?
A named operator can see what happened, inspect evidence, correct the output, reject or escalate a case, stop future actions, and restore a known state. If recovery exists only as “someone will check it,” the workflow is not ready.
Better alternatives
The answer is often a portfolio of small changes.
A strong design can combine process simplification, deterministic integration, a narrow AI step, and human approval. That is less dramatic than a general agent, but easier to reason about and often closer to what the work actually needs.
If inputs are inconsistent
Improve the intake form, validation, required fields, naming, or source-of-truth ownership before interpreting the mess downstream.
If language varies but the path is fixed
Use AI for classification, extraction, retrieval, or drafting inside a defined sequence, with evidence and review.
If systems do not connect
Use an API, event, scheduled transfer, or robotic automation where the mapping is stable and authorised.
If cases genuinely branch
Test a constrained agent with an allow-list of tools, limited actions, explicit budgets, and escalation states.
When an agent may be justified
Look for bounded variability.
A reasonable candidate has meaningful volume, a clear owner, accessible representative data, and cases that truly require different routes. Tool access can be narrow. Important effects remain reviewable. Success and failure are observable. The organisation can operate the queue and respond when the system abstains.
Promising shape
Research across approved sources, assemble evidence, choose a permitted analysis tool, draft an output, and ask a qualified reviewer to decide.
Dangerous shape
Read unrestricted data, decide an important case, contact an external party, change records, and continue without a clear limit or human checkpoint.
Pressure-test the idea
Use AI only where the workflow earns it.
Quantify the current burden, inspect the evidence foundations, and test one controlled working slice. A good outcome can be build, simplify, redesign, or stop.