What an Agentic AI Vendor Is Actually Selling You

The word changed. Check whether the product did.

You have a quote on your desk with the word "agentic" in the first paragraph. Twelve months ago the same vendor, sometimes the same salesperson, sold the same module as an assistant, a copilot, or intelligent automation. The line item moved. The demo is better. Whether the thing behind the demo changed is the question you are actually being asked to decide, and nothing in the deck answers it.

Gartner has a name for the gap. In June 2025 it estimated that of the thousands of vendors marketing agentic AI, roughly 130 were building anything that met the definition. It called the rest agent washing, and it predicted that more than 40 percent of agentic AI projects would be canceled by the end of 2027. That prediction is not really about the technology. It is about buyers signing for one category and receiving another.

We sit on the buying side of that table. Here is what is actually in the box.

Three products, one word

When a vendor says agent, they mean one of three things. They are priced similarly and demoed identically, and they fail in completely different ways.

A workflow with a model in it. The sequence is fixed. Step one reads the email, step two extracts the fields, step three writes to the ERP, step four routes exceptions to a human. A language model does the reading and the extraction, which is genuinely useful, because that step used to require a person or a brittle template. But the path is a flowchart somebody drew, and the model never decides what happens next. Most of what is sold as agentic in the mid-market is this. It is not a scam. It is a good product with a new label.

A rebranded scheduler. Something that already existed, running on a trigger or a cron, now with a chat interface bolted to the front and the word agent in the release notes. The RPA vendors did this first and fastest, because they had the largest install base to relabel. If you had this product in 2023 under a different name, the useful question at renewal is what specifically was added, described in terms of behavior rather than in terms of category.

An actual agent. The system is given a goal rather than a path. It chooses which tools to call, in what order, evaluates the result, and revises when it fails. Nobody drew the flowchart because the flowchart is generated at runtime. This is the real thing, it does exist, and it is the smallest slice of what is being sold.

The tell is not in the datasheet. Ask the vendor to describe what the system does on the fourth attempt at a task that has failed three times. A fixed workflow has one answer: it stops, or it routes to a human. An agent has to have a story about how it decided to try something different. Vendors selling the first thing will answer that question with an anecdote about the demo.

Why the distinction is a budget question, not a taxonomy question

If it were only a naming argument we would not spend a paragraph on it. It changes three numbers.

Pricing moved from seats to actions, and the meter runs on the model's decisions rather than on yours. Salesforce prices Agentforce Flex Credits at $500 per 100,000 credits, with a standard action at 20 credits, roughly ten cents, and a voice action at 30. That is transparent and reasonable. It is also a fundamentally different forecasting problem than a per-user license. Under seats, you know the bill in advance and the vendor absorbs the variance. Under per-action pricing, a system that decides for itself how many steps a task requires is deciding your invoice. A genuine agent that retries intelligently is the version of this that costs the most when it is working correctly.

Failure is silent in a way flowcharts are not. A fixed workflow that breaks throws an error, and someone gets paged. A system that reasons about its next step and reasons badly produces a confident, plausible, wrong result and moves on. The control you need is not uptime monitoring. It is an intervention rate: how often a human corrected or reversed something the system did. Vendors who have real deployments can produce that number. Ask for it as a number, from a named customer, over a stated period.

And the integration surface is wider than the pilot suggests. An agent is only as capable as the tools it can call, which means the value depends on write access to systems that currently only get read from. The security review for that is not the security review for a chatbot, and it usually happens two months after the contract is signed rather than before it. This is the same mechanism that turns a quoted pilot into something else entirely, and we have itemized what that actually costs elsewhere.

What the category is genuinely good at

We are not arguing that none of it works. The 11 percent of agent pilots that reach production are reported to return well above their cost, and the shape of those wins is consistent enough to describe.

Agents earn their keep where the work is high-variance, low-consequence, and verifiable. High-variance, because if every case looks the same you needed a workflow and a workflow is cheaper and more predictable. Low-consequence, because early autonomy will be wrong sometimes and you want those instances to be recoverable. Verifiable, because the difference between a useful agent and an expensive one is whether the system can tell that it failed. Research and retrieval, triage, drafting, reconciliation against a source of truth: these have the shape. Anything that ends in a payment, a filing, or a customer commitment does not have it yet, and the vendors quietly agree, which is why the demos are always about drafting.

The inverse is the useful screen. If your process is uniform, documented, and runs the same way a thousand times a month, an agent is the wrong tool and you are paying a premium for a decision-making capability you will spend the implementation suppressing. That is the general case of a pattern we have described before: most operations problems pitched as AI problems are not AI problems. Agentic pricing just makes the mismatch more expensive.

What we would ask before signing

Not a checklist, because the sequence depends on what you run and who supports it. But four questions separate the three products faster than any demo does.

What decides the next step, and can you show me where that decision lives? A vendor selling a workflow can point at a canvas. A vendor selling an agent has to explain a runtime, and the quality of that explanation tells you most of what you need.

What is your intervention rate in production, at a named customer, over a stated period? Vagueness here is the answer.

What does a bad month cost me? Take the per-action price and the retry behavior and describe the worst plausible week. If nobody at the vendor has modeled that, you are the one who will.

And what did this product do in 2024? Not to be difficult. If the answer is a real change in architecture, they will enjoy telling you about it. If the answer is a positioning shift, that is worth knowing before you plan around a capability that does not exist yet. The agent may be new, but a lot of these agents are just their old software wearing a trench coat.

None of this means wait. The category is real, the good deployments are real, and the buyers who sat out 2025 entirely are not obviously better off. It means buy the specific thing that is in front of you rather than the category it is being sold under, which is the same discipline that applies to evaluating any AI vendor and gets harder every time the vocabulary turns over.

Where this stops being an article

We can tell you which questions decide this. We cannot tell you what your answers are, because they depend on how your exceptions actually flow, what your systems will let anything write to, and who is on the hook when a decision made at three in the morning turns out to be wrong. That is a week inside your operation, not a blog post.

If you have an agentic quote in front of you this quarter and you cannot tell which of the three products it is, bring it to a consult. We will read the contract with you and tell you plainly whether the thing you are buying is new, and whether you need it. Sometimes the answer is no, and that is a cheaper conversation to have now than after the first invoice.