The distinction in one sentence
An automation follows a path you defined. An agent decides the path each time, using a language model, inside limits you set.
That is the whole difference. Everything else follows from it: price, speed, reliability, and how you debug the thing on a Friday evening.
What an automation is good at
Deterministic work. The invoice arrives, the total is extracted, the record updates, the accountant gets a notification. Same input, same output, every time.
It runs in milliseconds and costs close to nothing per execution. When it breaks you can see exactly which step failed, which matters more than people expect.
Roughly 70 percent of what we build is this, even in projects the client describes as an AI project.
What an agent is good at
Work where the next step depends on content you cannot predict. A support message that might be a complaint, a booking or a supplier chasing payment. A contract where the renewal clause is on page four of some documents and page eleven of others.
Rules cannot cover that space without becoming a thousand line mess nobody maintains. A model reads it, decides, and hands off.
Use an agent where you cannot write the rule. Use an automation everywhere else, because it is cheaper, faster and it tells you why it failed.
The running cost nobody models
Automations cost server capacity, which is flat. Agents cost per call, and the bill scales with your volume.
Here is a real shape. A company handles 4,000 support messages a month, and each one needs classifying, summarising and routing.
Forty four dollars a month is trivial, so the agent is obviously right. Now change the design. An early version of that system asked the model to re read the full thread on every update, at eight updates per conversation.
Same volume, same model, $610 a month. Nothing was broken. The design just paid for the same reading thirty two thousand times.
The rule we apply
Let the model touch a job once, at the point of judgement. Everything before and after should be plain logic. That single principle usually cuts the model bill by 80 percent without changing the result.
Reliability, told honestly
A well built automation is right essentially always, because it does exactly what it was told. An agent is right most of the time, and the remaining slice is not random noise you can ignore.
This is why every agent we ship has a confidence threshold and a queue. Below the line it goes to a person, with the reason attached. Reviewing 6 percent of cases is a job. Checking 100 percent because you cannot tell which 6 were wrong is not.
A test you can apply in one minute
Take the process and ask: could you write the instructions on one page, such that a new starter would follow them the same way every time?
- Yes. Build an automation. Adding a model here buys you cost and uncertainty in exchange for nothing.
- Yes, but only after reading something. Automation for the flow, model for the reading step alone. This is the most common answer by a wide margin.
- No, it depends every time. Agent, with limits, a confidence threshold and a human queue.
- Nobody can write it at all. You have a process problem. Software will make it faster and no clearer.
Where the marketing gets loose
Plenty of tools sold as agents are decision trees with a chat window on top. That is not a criticism of the tool, and it is a problem when you are paying agent prices for automation work.
Two questions cut through it. Does the model choose the next step, or only fill in text at the end? What happens to the price if my volume triples?
A supplier who cannot answer the second question has not run one of these in production.
What we usually end up building
A typical sales system runs deterministic routing, deterministic reminders and deterministic reporting. The model appears in three places: reading the enquiry, drafting the reply, summarising the call afterwards.
People call the whole thing an AI system. Fine by us. The reason it stays up is the boring 70 percent underneath.
Latency, and why users notice
An automation replies in under a second. An agent thinks for two to eight seconds, sometimes longer on a large document.
Nobody minds waiting for a summary. Everybody minds waiting for a screen, so put the model where a short pause reads as thinking rather than as breakage.
Ownership, in both cases
Both run on your infrastructure, in your accounts, with your keys. Model access is a supplier relationship you hold directly, so the provider can be changed without rebuilding the system around it.
That matters more than it sounds. Model pricing and quality both moved several times over the last two years. Companies able to switch in an afternoon paid a fraction of what the locked in ones did.
Five things you can do tomorrow.
- Take your most annoying process and try to write it on one page. The attempt gives you the answer.
- For anything already using a model, count how many times it is called per item. More than twice, look for repeated reading.
- Ask your supplier what the monthly model bill looks like at three times your current volume.
- Set a confidence threshold and build the human queue before go live, not after the first bad week.
- Move any step that follows a fixed rule out of the model. It is nearly always cheaper and faster.
