Somewhere in the last two years, "AI agent" turned from a technical term into a sales word. Software that used to be called a chatbot, a workflow, or a script with an API key bolted on now gets pitched as an agent, sometimes as an "AI employee," complete with a name and a headshot. The pitch is usually some version of: this thing works like a person, except it does not sleep, does not quit, and costs a fraction of a salary.
Some of that is real. A useful category of software genuinely can plan a sequence of steps, call tools, check its own work, and handle a task from start to finish without a human directing each move. That is a meaningful capability and it did not exist in a usable form a few years ago. But the term has been stretched to cover everything from that genuine capability down to a single API call with a system prompt, and operators are being asked to pay agent prices for chatbot software more often than the marketing admits.
This is a guide to the difference: what the word technically means, what it can and cannot replace, and what to check before you sign anything.
The technical definition, stripped of the sales layer
Anthropic, one of the companies building the models underneath most of this software, draws a useful line in its own engineering guidance. They separate workflows, where the software follows a predefined sequence of steps that a person designed in advance, from agents, where the model itself decides what to do next: which tool to call, whether the result was good enough, and when the task is finished (Anthropic, "Building Effective Agents").
That distinction matters more than it sounds like it should, because most of what gets called an agent is the first kind. A tool that reads an inbound email, extracts three fields, and drops them into a spreadsheet is a workflow. It is useful, it can save real time, and it is not making decisions, it is executing a fixed path with a language model doing the reading. An agent, in the stricter sense, is the software that gets to that spreadsheet by its own route: it decides whether the email needs a reply, whether it needs another piece of information first, whether to escalate, and it can be wrong about all three in ways a fixed workflow cannot be, because a fixed workflow does not choose anything.
Neither one is better in the abstract. A workflow is predictable, cheap to run, and easy to audit, because the same input produces the same steps every time. An agent is more capable and more expensive, in compute and in the harder-to-price cost of unpredictability, because the same input can produce a different path depending on what the model decides along the way. The right choice depends on the task, and the honest sales conversation says which one you are actually buying. Most of them do not.
Why the term got stretched
Part of this is ordinary marketing. "Agent" tests better than "workflow" or "script," and "AI employee" tests better than either, because it maps onto something a buyer already understands: hiring. It is a much easier pitch to sell a replacement for a hire than a piece of software with a defined, narrower job.
Part of it is a real supply problem. Gartner, which tracks enterprise software adoption, estimated that of the thousands of vendors currently marketing agentic AI products, only around 130 have agentic capability that would hold up to scrutiny, a pattern they call "agent washing," the relabeling of existing automation as agentic without the underlying capability changing (Gartner, June 2025). The same report predicts that more than 40% of agentic AI projects will be canceled before the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls as the main causes, not model quality.
That number is worth sitting with. It is not a prediction that the technology fails. It is a prediction that most of the projects being funded right now, under the current hype, will be recognized as the wrong bet before they finish, because they were scoped against the marketing description of an agent rather than the actual one.
What an agent can genuinely replace
Set the marketing aside and the honest list of what agentic software does well is narrower than "an employee," but it is real.
Multi-step retrieval and synthesis, where the steps are not fully known in advance. Research a competitor, pull the relevant pages, cross-reference two data sources, and produce a summary. The exact sequence of pages to check is not predictable ahead of time, which is where the model deciding its own next step earns its cost over a fixed workflow.
Tool use bounded by a clear success condition. Book something, file something, update something, where the software can check afterward whether the action succeeded and retry or escalate if it did not. The clarity of the success condition is what makes this safe to hand to something that makes its own decisions. Without it, you cannot tell whether the agent did the job or just believes it did.
Long-running tasks with checkpoints. A process that takes many steps over minutes or hours, where a human can review progress at defined points rather than approving every action. This is where "does not sleep, does not need supervision every ten minutes" is a genuine advantage over a person, not a marketing line.
Notice what is missing from that list: judgment calls with no clear right answer, situations where the cost of a wrong decision is high and hard to reverse, and anything where a client or regulator expects to know a specific person made the call. Those are the situations "AI employee" gets pitched hardest for, and they are the ones it is worst suited to.
The three questions that separate a real agent pitch from a rebrand
Before buying anything sold as an agent, three questions do most of the filtering.
What does it do when it is wrong, and how do you find out? An agent that fails loudly, that surfaces an error and stops, is a manageable risk. An agent that fails by producing a plausible but incorrect result, silently, is a liability wearing a feature. Ask specifically how errors surface, not whether the demo went well.
What is the actual decision space? Ask the vendor to describe, concretely, the set of choices the software makes on its own versus the set of steps that are fixed in advance. If they cannot answer this precisely, in specifics rather than in the language of the pitch, you are very likely looking at a workflow with an agent's price tag.
Who is accountable for the outcome, and does the contract say so? A person hired for a role carries accountability that is understood by everyone around them without a document. Software does not, unless the vendor agreement says explicitly what happens when it gets something wrong, who is liable, and what the recourse is. If that language is vague or absent, price the software as a tool with a support contract, not as a hire, because that is what you are actually accountable for when it fails.
None of these questions require technical depth to ask. They require refusing to accept the framing that the software is a colleague rather than a system, because a colleague's mistakes are covered by things a system's are not: judgment, context, and a stake in the outcome.
Why this matters for the build-or-buy decision
The workflow-versus-agent distinction is not just semantics, it changes what the right build looks like. A task that is genuinely well served by a workflow, fixed steps executed reliably, is often cheaper and safer built as exactly that: predictable, auditable, and boring in the way that dependable infrastructure should be boring. Paying for agent-grade unpredictability and agent-grade cost on a task that never needed the model to make its own decisions is buying capability you do not use and cannot fully control.
The reverse mistake is just as common: forcing a genuinely open-ended task, one where the right next step depends on what happened last step, into a rigid workflow because it was cheaper to scope and easier to demo. That produces the brittle automation that breaks the moment a real input does not match the happy path it was built for.
The useful question to bring into any vendor conversation, or any internal build decision, is not "should we get an agent." It is "does this task have a decision space narrow enough that a fixed workflow handles it, or wide enough that it genuinely needs something deciding its own next step." Answer that first, honestly, and the right tool follows. Answer it by accepting whatever the pitch calls itself, and you are pricing software by its adjective instead of its job.
Frequently Asked Questions
Is an AI agent the same thing as a chatbot?
No, though a lot of software marketed as an agent is closer to a chatbot with extra steps. A chatbot answers questions in a conversation. An agent, in the stricter technical sense, plans a sequence of actions, calls tools to carry them out, checks the results, and decides what to do next without a person specifying each step in advance. Many products marketed as agents are actually fixed workflows, predefined sequences of steps with a language model reading or writing text at certain points, which is useful but is not the same capability.
What does 'agent washing' mean?
It is the practice of relabeling existing automation or chatbot software as an "AI agent" without the underlying product gaining real agentic capability, mostly to take advantage of the term's current marketing power. Gartner has estimated that only a small fraction of the vendors currently marketing agentic AI products, roughly 130 out of thousands, have capability that holds up to that description, and expects more than 40% of agentic AI projects at large organizations to be canceled by the end of 2027 for reasons including unclear business value (Gartner, June 2025).
Can an AI agent actually replace an employee?
For a narrow, well-bounded task with a clear success condition, an agent can take real work off a person's plate, sometimes all of it. It cannot take on accountability, judgment calls without a clear right answer, or the trust a client places in a specific person, because those are not functions of decision-making speed, they are functions of stake and context an agent does not have. Treat "AI employee" as marketing language for "software that can do more of this task unsupervised," not as a literal substitute for a hire, and the buying decision gets much clearer.
How do I tell if a vendor's 'agent' is real or a rebrand?
Ask them to describe the actual decision space: which choices the software makes on its own, and which steps are fixed in advance regardless of input. A vendor with a real agentic product can answer this precisely. A rebrand tends to answer in the language of the pitch rather than in specifics. Also ask what happens when it is wrong and how you would find out, since a system that fails silently with a plausible-looking result is a materially different risk than one that fails loudly and stops.
Should I build a custom workflow instead of buying an agent product?
It depends entirely on the shape of the task, not on which word sounds more advanced. If the steps are known in advance and do not change based on what happens along the way, a fixed workflow is usually cheaper, more predictable, and easier to audit, whether you build or buy it. If the right next step genuinely depends on what the software discovers as it goes, that is the case agentic software was built for, and a fixed workflow will feel rigid and break on real-world variation. Scope the task's actual decision space first, then decide what to build or buy.
This guide provides educational information based on industry research and case studies. Individual results vary by market, budget, and execution.