The short answer
If you can draw the path on paper, you need a workflow, not an agent. A workflow executes a route you specified. An agent selects a route against a goal you specified. Most problems people bring to agents are workflow problems, and workflow tools solve them for less money and fail in ways you can predict.
What is the actual difference between a workflow and an AI agent?
The difference is where the decision lives.
In a workflow, you made every decision in advance. A form arrives, look up the domain, check the headcount, route to one of four owners, send one of three templates. The system considers no alternatives. It follows the map you drew.
In an agent, you named the destination and left the route open. Work out whether this account is worth pursuing, then do something sensible about it. The system reads what it finds, chooses the next step, and may take a different path on Tuesday than it took on Monday from an identical starting point.
Automation repeats a decision you already made. Autonomy makes one you did not.
Everything else in this argument is packaging.
How do I know which one my problem needs?
Count the branches. If you can list the situations your process meets, and the list stops at a dozen or so, you have a workflow. If the situations are open ended, if new ones turn up most weeks, and if a competent person would have to read something and form a view before acting, you have a candidate for an agent.
| Shape of the problem | What it needs | Why |
|---|---|---|
| One trigger, a handful of known steps, no judgement required | Zapier, Make, or a native integration | The path is already known. Paying a model to rediscover it every time buys variance. |
| Many branches, all of them listable, high volume | A coded workflow or an orchestration tool | Deterministic logic is testable. You can prove afterwards that it did the right thing. |
| Structured routing, but one step needs a paragraph read and understood | A workflow with a model inside one node | The judgement is a single step inside it. Keep everything around it deterministic. |
| Stable goal, open-ended situations, new cases most weeks | An agent, with a written scope and a full audit log | No fixed path exists. The value sits in the choice of next step. |
| The right action depends on reading several sources and forming a view | An agent, with a person approving anything that leaves the building | Judgement is the work. Approval is where the risk is absorbed. |
| Nobody in the business agrees what a correct output looks like | Neither. Fix the definition first | Automating an unclear process produces an unclear process running faster. |
Where does Zapier stop being enough?
Zapier and Make are very good at one thing: moving a known object from one place to another when a known condition is met. Inside that boundary they are usually the cheapest correct answer, and rebuilding their function in code is a common and expensive mistake.
They stop at two points. The first is branch count. Once the logic runs past roughly fifteen conditional paths, a visual builder becomes harder to reason about than a script, and nobody on the team can say with confidence what happens to an edge case.
The second is judgement. When a step requires reading a paragraph and deciding what it means, no arrangement of filters will get there. That is the point where a model belongs inside one node. It is rarely the point where the whole workflow should become an agent.
Why do so many products called agents turn out to be workflows?
Gartner reviewed the market in June 2025 and found that of the thousands of vendors describing themselves as agentic, roughly 130 were genuine. The same analysis expects more than 40% of agentic AI projects to be cancelled before the end of 2027.
Most of the remainder are workflows with a language model in one step and a new word on the pricing page. That is not automatically a bad product. A well-built workflow with a classification step is often exactly the right thing to buy. The failure is in the gap between what was priced and what was expected. The buyer paid for a system that would handle situations nobody wrote down, and received one that handles the situations someone did.
S&P Global surveyed more than 1,000 businesses and found 42% scrapped most of their AI initiatives during 2025, up from 17% in 2024, with 46% of proofs of concept abandoned before production.
What does it cost to get this wrong in each direction?
The two mistakes are not symmetrical.
Build an agent where a workflow would have done, and you pay per token for decisions that never vary, while introducing variance into a process that had none. You also lose the ability to explain why one record ended up where it did.
Build a workflow where an agent was needed, and you get a system that handles every case you imagined and quietly mishandles the rest.
The first failure is loud and arrives on an invoice. The second is silent and arrives in the pipeline.
Is there a middle option?
Most systems that survive in production are workflows with one or two agentic nodes inside them. The deterministic parts do the moving and the logging. The model does the reading and the classifying. A person approves anything that reaches a customer.
This shape is unfashionable because it does not demonstrate well in a sales call. It is also the shape that survives contact with a real week of data.
What is not known
There is no independent published comparison of an agent implementation and a workflow implementation of the same task, by the same team, against the same baseline. Vendor case studies are not that, and no research group has run the trial.
What does exist is evidence that self-assessment in this area is unreliable. METR's 2025 study of 16 experienced developers across 246 tasks found AI assistance made them 19% slower while they believed it had made them 20% faster. Gartner's early 2026 survey of 210 chief sales officers found that 25% report a return above 50% and 20% report a negative return worse than 50%. The same technology, applied to the same function, is producing outcomes at both ends of that range. A separate Gartner survey of 227 chief sales officers found that 31% name difficulty proving the return on AI tools as a top challenge for 2026.
The test, if you keep one line
Write the runbook by hand. If you can finish it, build the runbook and do not buy an agent. If finishing it requires the sentence "it depends what you find", you have located the part that needs autonomy. Build that part, and only that part.