Answer
What is the difference between AI automation and an AI agent?
Automation runs a path you defined. An agent decides the path itself. Choose by whether the steps are predictable.
AI automation is a fixed sequence you designed, with a model doing steps inside it. An agent chooses its own sequence. Pick automation whenever the steps can be written down in advance; it is cheaper and more predictable.
Most things sold as AI agents are AI automations, and the mislabelling costs buyers real money because the two have different failure modes and different price curves. An automation is a route: this trigger, then this classification, then this branch, then this write. A model may be doing the classification or the drafting, but the route was drawn by a person and does not change. An agent is handed the destination and works out the route each time.
The test that decides between them is whether the number of steps is predictable. Sorting inbound enquiries into three categories is a fixed number of steps and always will be. Investigating why a particular invoice does not reconcile is not: it might take two lookups or eleven, depending on what the second one reveals. The first case is an automation and putting an agent on it buys nothing. The second case is where agents earn their complexity, and it is rarer in a small business than the market suggests.
The cost curves diverge sharply. An automation's cost per run is roughly fixed and knowable before you deploy it, because the calls it makes are the calls you wrote. An agent's cost per run is a distribution with a long tail, because the system decides how many calls to make and occasionally decides on a great many. That tail is not a defect to be engineered away; it is the price of the flexibility, and it is why agent budgets are set as caps rather than as forecasts.
Reliability diverges too, and in the direction people find surprising. Automations fail loudly and in one place — the branch was wrong, the field was empty, the interface changed — and you fix it once. Agents fail quietly and differently each time, because the route was different each time. A test suite that passes on an automation means the automation works. A test suite that passes on an agent means it worked on those inputs, which is a weaker statement than it looks.
The practical sequence for a business is therefore to automate first and agent later, for reasons that have nothing to do with technological caution. Building the automation forces you to write down the actual steps, and writing down the actual steps is most of the work either way. If you then discover the steps genuinely cannot be fixed — that the real job is investigation rather than processing — you will have the tools, the permissions and the data access already built, and the agent becomes a small addition rather than a project.
One qualifier worth stating plainly. The presence of a model inside an automation is not what makes it fragile; a classification step that is right most of the time and wrong occasionally needs a route for the wrong case, and that route is design work rather than model work. Automations that fail in production usually fail because nobody wrote the branch for the uncertain case, not because the model was insufficiently capable.
If you can draw the steps on a whiteboard, an agent will spend money rediscovering your whiteboard on every single run.
Siddharth Sharma, Context Theory
Related questions
Where does a scheduled prompt sit?
That is an automation with one model step, and it is the most under-rated shape available. A prompt that runs every morning against yesterday's data, writes a short summary somewhere a person will read it, and does nothing else has a fixed cost, an obvious failure mode and no permission surface. A large share of the value attributed to agents is actually available here.
Can an automation call an agent?
Yes, and this is usually the right architecture when you need both. The automation handles the predictable ninety per cent — trigger, classify, route, write — and hands only the exceptions to an agent with a bounded budget and a defined report format. You get predictable cost on the common path and flexibility exactly where the path could not be predicted.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Odds of qualifying a lead — replying within the first hour vs after it | 7× | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | The choice between a defined route and an agent turns on whether the number of steps is predictable in advance, because agents exist to handle open-ended problems where a fixed path cannot be hardcoded and offer nothing on problems where it can. | Anthropic's 'Building effective agents', which states that agents suit open-ended problems where the required number of steps cannot be predicted. |
| Buying behaviour | Cost per run is approximately fixed for a defined route and is a long-tailed distribution for an agent, because the number of model calls in the second case is chosen by the system rather than by the author, which is why agent budgets are set as caps rather than forecasts. | Comparing per-run token counts across repeated executions of a fixed workflow and of an agent on the same class of task. |
| Software | A passing test suite means different things for the two shapes: for a fixed route it demonstrates the route works, and for an agent it demonstrates only that those inputs produced an acceptable path, because the path itself is re-chosen on every run. | Running an agent twice on identical input and diffing the sequence of tool calls it made. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one