What Actually Counts as an AI Agent

A word that's lost its meaning, and four questions that restore it.

Three vendors will pitch you an "AI agent" this quarter. There is a good chance none of them are selling the same thing.

One will be a chatbot with a better interface. One will be scripted automation that has been running for years under a different name. One might be the real thing. The pricing will not tell you which is which, and neither will the demo.

Gartner gave the practice a name: agent washing — rebranding assistants, robotic process automation, and chatbots as agentic without adding substantial agentic capability. Of the thousands of vendors claiming agentic capability, Gartner estimated only about 130 were building something that genuinely deserved the label.

Treat that number as directional rather than precise; Gartner never published the methodology behind the denominator. But the direction is not in dispute, and if you are evaluating AI right now, the direction is the part that costs you money.

The fix is not more technical knowledge. It is a clearer vocabulary.

Four levels, not one word

The useful mental model already exists, and you have it from cars. Nobody confuses lane-keeping assist with a vehicle that drives itself. Level 2 and Level 4 are different products with different risks, and everyone knows it.

Software procurement has not caught up to that discipline. Here is the same ladder, applied to AI.

Level 1 — Chatbot. Answers questions from a fixed knowledge base. It retrieves; it does not reason about your business. Good for deflecting repetitive questions. What it demands from you: accurate source content, and not much else.

Level 2 — Assistant. Does the task you ask, hands the result back, and waits. Drafting, summarizing, analysis, first-pass code. This is where the overwhelming majority of real business value sits today. What it demands from you: a human who checks the output. Every time.

Level 3 — Workflow. Multi-step automation on a predetermined path. Step one, then step two, then step three. It may call a model at each step, but the sequence was decided by a person in advance. Good for stable, repeatable processes. What it demands from you: a process that does not change often.

Level 4 — Agent. You give it a goal, not a path. It plans its own steps, selects its own tools, acts across systems, observes what happened, and adapts when it hits something unexpected. What it demands from you: scoped permissions, an audit trail, a rollback plan, and a named owner.

That fourth definition is not mine. Menlo Ventures used it to survey nearly 500 enterprise decision-makers and found that only 16% of enterprise deployments qualify as true agents — systems where the model plans, executes, observes feedback, and adapts. The rest are built around fixed sequences or routing logic wrapped around a single model call.

Sixteen percent. In a market where nearly every vendor uses the word.

The principle underneath the ladder is the one that matters: every level up adds autonomy, and every level up adds a governance requirement. They are not separable. Autonomy without the matching governance is not a more advanced deployment — it is the same deployment with the brakes removed.

Why the confusion is expensive

Forrester's mid-2026 assessment of the category was titled, pointedly, Companies Are Chasing, Few Are Catching. Roughly three-quarters of enterprise leaders report adopting agentic AI. Only a small minority have it running in meaningful production beyond what the authors call "agentish" chatbots.

That gap produces two failure modes, and they cost money in opposite directions.

The loud one: you pay Level 4 prices and carry Level 4 governance exposure for something delivering Level 2 value. The invoice says agentic. The system is a workflow with a nicer front end.

The quiet one is worse, because nobody writes it down. You refuse a Level 2 assistant across a hundred people — a deployment with almost no autonomy risk — because the board conversation about AI has been dominated by Level 4 fears that simply do not apply to it. The exposure you avoided was never there. The productivity you skipped was.

Gartner's widely quoted prediction is that more than 40% of agentic AI projects will be canceled by the end of 2027. Two things about that figure are worth knowing, and most coverage omits both.

It was published in June 2025, not 2026. A great deal of commentary drops the date and presents a prediction as a finding. And the causes Gartner named are not technical: escalating costs, unclear business value, and inadequate risk controls. Every one of those is a management failure, not a model failure.

Which means it is preventable, and preventable by you.

The gap nobody puts in the strategy deck

Deloitte surveyed 501 US leaders in spring 2026 — every one of them at an organization already piloting agentic AI, so this is not a survey of skeptics. Published this month.

Ask those leaders how ready they are, and the answers separate cleanly:

• Vision and strategy: 52% report readiness
• Technology infrastructure: 48%
• Data: 42%
• Risk, security, and governance: 39%
• Workforce: 25%
• Business processes: 21%

Strategy is where organizations feel most prepared. Process is where they feel least. That ordering is not a coincidence and it is not a scoring quirk — strategy readiness measures confidence, and process readiness measures work that has actually been done.

The rest of the survey holds the same shape. Only 15% report scaled, orchestrated multi-agent deployments, and many of those are pointed at low-risk, low-ROI applications. Seventy percent flagged concerns about their ability to govern and trust agents at all. In Deloitte's broader global research, roughly one company in five reports a mature governance model for autonomous agents.

If your AI program has stalled somewhere between an encouraging pilot and anything you would defend to an auditor, you are not behind. You are exactly where the data says most organizations are. The distance between a working pilot and a governed production system is the entire problem, and almost nobody priced it in.

Four questions before you sign

None of this requires a consultant, a maturity assessment, or a new framework. It requires four questions, asked in a room with the vendor, and the discipline to write down the answers.

1. What does this do without a human in the loop?

If the honest answer is nothing, you are buying an assistant. That is fine — assistants are where most of the value is right now. Just do not pay agent prices or build agent governance for it.

2. What can it write to, not just read from?

Read-only and write-enabled are different risk categories, not different settings. Get the list of systems and the permission scope in writing, before procurement, and confirm who can expand that scope later.

3. When it is wrong, how do I find out?

Not if. When. Ask to see the log from a failed run rather than a successful demo. A vendor who cannot produce one either has not run the system long enough or is not instrumenting it. Both are answers.

4. Who owns this in production, and on whose budget line?

A name and a number. Not a committee, not "IT," not the sponsor who championed the pilot and will have moved on by February. If neither exists yet, you are funding a demo. Demos are genuinely useful — keep the spend small and stop expecting one to become infrastructure on its own.

The question worth asking instead

Gartner's own recommendation maps almost exactly onto the ladder: use agents when a decision needs to be made, automation for routine workflows, and assistants for simple retrieval.

That is unglamorous advice, and it is right. Most organizations do not need Level 4 in one department. They need Level 2 done well across a few hundred people, with someone accountable for whether the output is any good.

So the question is not whether you are deploying agents. Half the market cannot define the word, which makes it a poor thing to measure yourself against.

The question is what level each job actually requires — and whether you are prepared to govern the level you just bought.

Sources

• Gartner, Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — press release, June 25, 2025 (a prediction, not a finding)
• Menlo Ventures, 2025: The State of Generative AI in the Enterprise — December 2025, survey of ~500 US enterprise decision-makers
• Forrester, The State of Agentic AI, 2026 — June 2026
• Deloitte, AI Agents Are Only the Beginning — August 2026, survey of 501 US senior manager to C-suite respondents, fielded April–June 2026
• Deloitte, State of AI in the Enterprise, 2026 — survey of 3,235 leaders across 24 countries