Engines do the maths. Agents do the judgement.
Automation, engine and agent are three different jobs. Get them confused and you either trust a chatbot with your numbers, or never trust AI with anything.
Three answers to one question
Ask a chat model to work out last month’s margin from a settlement report. Then ask again in a new chat. Then a third time. There is a fair chance you get three different answers.
That is not the model misbehaving. A language model predicts the most likely next words. That makes it brilliant at reading a messy email and poor at giving the same answer twice. A margin has one right answer. It needs a calculator, not a writer.
Three words, three jobs
Same input, same answer
- AutomationmovesAn order ships, a review request goes out, a file lands in a folder. No thinking.
- EnginecomputesMargin per product from the settlement, stock cover, VAT. Code, not a model.
A judgement call
- AgentdecidesReads the engine’s answer, works out why margin fell or what to reorder, drafts the fix.
- YouapproveYes or no. Nothing leaves without it.
The cricket version
Think of a review in Test cricket. The captain decides whether to send the decision upstairs. That is judgement: the match situation, the reviews left, how sure the bowler looks. Hawk-Eye decides whether the ball would have gone on to hit the stumps. That is maths, and nobody wants the captain’s opinion on it. The umpire gives the final decision. In your business, that is you.
The captain decides whether to review. Hawk-Eye decides whether it is out.
Football has the same split. Goal-line technology says whether the ball crossed the line. The pundit says why the defence fell apart. You would never let the pundit decide whether it was a goal.
The rule
- Agents never do the maths. They call the engine.
- Engines never make the call. They hand the answer on.
- Automations never decide. They move things.
“Can’t the model just use a calculator?” Yes. And that calculator is an engine. That is the point.
What it looks like in a real business
Most useful systems end up with the same five steps:
- An agent reads the mess in: PDFs, emails and exports turned into clean rows.
- An engine applies the rulebook: fees, costs, tax, whatever the maths is.
- Everything lands in one reconciled store.
- An agent explains what changed and drafts the action.
- A person approves it.
Where it breaks
Be honest about the weak spots. There are three.
- Wrong rules give confidently wrong numbers. An engine is only as good as its rulebook, so check its output against the source system, not against itself.
- An agent’s explanation can still mislead. That is why it drafts and a person decides.
- The boundary leaks. The moment an agent “just quickly works out” a total, you are back to three answers to one question.
Which one do you need?
Take any job from your week and ask two questions. Does a number have to be right? Build an engine. Does it need a judgement call? Put an agent on top. If the answer to both is no, it is an automation.
Most jobs need both: engine first, agent on top.
If a number can be argued with in an audit, code produces it. If it needs reading, an agent writes it.