Do You Really Need an Agent? A 6-Question Reality Check

The Challenge

The most expensive AI project this year was never built — it was the one that shouldn't have been built, but got built anyway.

We keep meeting teams that decree 'go all-in on Agents' and then spend three months assembling what they call an 'agent system': quotes auto-sent, contracts auto-read, weekly reports auto-written. On the surface it looks smart. But when we ask one question — 'how many of these actions are exactly the same every single time?' — the answer is usually 'pretty much all of them.' Same tools. Fixed steps. That isn't an agent. It's a nicely packaged workflow.

The real challenge isn't building agents. It's knowing when you don't need one.

Methodology

We distilled the judgment into a 6-question self-check. It's sharp because it forces honesty about what's actually repeating:

1) Is your agent's first action always the same? → If yes, you've built a workflow. 2) Do you call the same tools every time? → If yes, you've built an API wrapper. 3) Does the team need to plan which tools to use? → If yes, you probably don't need an agent. 4) Can the execution steps be determined in advance? → If yes, prefer a workflow. 5) Does the task need to handle unexpected intermediate results? → Only then is it agent territory.

Count your 'greens.' The more greens, the more you should fall back to a workflow or API wrapper. Only when #5 lights up does the agent actually take the stage.

Our Approach

We place any agent candidate on a scale of autonomy — three tiers:

Tier 1 · In-process autonomy: deterministic judgment hardened into a process; the process runs itself, human is outside the loop. In practice this is the workflow / API-wrapper tier — no agent needed. Tier 2 · Collaborative autonomy: semi-deterministic judgment shared between human and machine; the agent proposes, the human confirms. Tier 3 · Outside-loop autonomy: the human exits the execution loop; the agent runs autonomously.

Fuse the 'whether' (the 6-question check) with the 'how' (the three tiers) and you get a decision map: high determinism + fixed steps → workflow / API wrapper, step back directly; semi-deterministic + human-machine co-ownership → Tier 2; needs to handle unexpected results + controllable value → Tier 3; high value + heavy responsibility + no precedent → human-in-the-loop, always. The one who signs must be human.

The Outcome

Back to that team at the opening. The 'agent' they spent three months building? After the self-check: 4 of 6 questions green.

That wasn't an agent. It was a workflow they should have finished in three days.

But the boss now tells everyone: 'We've gone full Agent.'

The most expensive AI project was never built. It's the one that shouldn't have been — yet no one dared say 'let's step back and think first.'

A sobering truth we see repeatedly: most companies clamoring for Agents could cover 60% of their cases with a simple API wrapper. We don't not need Agents — we just don't need that many of them.

Key Results

Key Learnings

The framework isn't about rejecting agents. It's about matching the tool to the repeating pattern. Most 'agent' initiatives fail the moment you ask whether the action is identical every time — because the honest answer reveals a workflow wearing an agent costume.

Agents earn their place only where unexpected intermediate results must be handled, value is controllable, and errors are reversible and verifiable. Everything else is a workflow or an API wrapper wearing a badge.

Who This Is For

This 6-question reality check is for teams that: are evaluating whether to build an Agent but aren't sure it's warranted; have already decreed 'go all-in on Agents' and want a sanity check; run repetitive, tool-driven processes they assume need an agent; or want to redirect AI budget from hype-driven builds toward the highest-ROI, lowest-risk automation.