“Agent” has become the word every software vendor wants on their product page. For a business owner, the more useful question is simpler. For a given job, do you want something that follows the same steps every time, or something that can use judgement?
Those are two different tools. Choosing the right one for each step of a process is what separates automation that quietly saves hours from automation that creates a new set of problems.
Workflows: same input, same result
An automated workflow follows rules you define. When a purchase order is approved, create the bill in your accounting system. Every night, pull stock levels from inventory and update the reorder report. When a job is marked complete, send the invoice.
Workflows are predictable, cheap to run, easy to test and easy to audit. If something goes wrong, you can see exactly which step failed and why.
Their weakness is messy input. A workflow does nothing useful, or fails, when the input doesn’t fit the rules. A supplier changes their invoice layout. A customer emails a free-text order. Someone types a product code slightly differently.
Agents: judgement on messy input
An AI agent uses a language model to decide what to do next, often by calling tools along the way. It might read an email, work out what the sender wants, look up the customer, check stock and draft a reply.
Agents are good at variety: free-text emails, documents in different formats, questions phrased a hundred different ways. They can summarise, classify and draft in ways rules never could.
Their weaknesses are the flip side. They’re less predictable. They can be confidently wrong. They cost more per run and are harder to test thoroughly. And once an agent can take actions, security becomes a real concern, because content it reads can try to steer it (see our explainer on prompt injection).
The decision table
| If the job… | Use |
|---|---|
| Follows clear rules every time | A workflow |
| Starts from structured data in one of your systems | A workflow |
| Must give exactly the same result for the same input, such as money or compliance | A workflow |
| Involves reading free text or documents in varying formats | An AI step, usually inside a workflow |
| Needs a judgement call, a summary or a draft | An agent, with a person reviewing |
| Moves money, changes master data or contacts customers | Either, but with human approval |
Usually the answer is both
Most real processes mix the two, and the best designs use each where it’s strongest.
Take supplier invoices arriving by email. A workflow picks up the email and the attachment. An AI step reads the invoice, whatever its layout, and pulls out the supplier, invoice number, lines and totals. The workflow then checks those details against the purchase order and the supplier record. Anything that matches goes through. Anything that doesn’t, such as a different bank account, a price change or a total that doesn’t add up, goes to a person. Approved bills are written to the accounting system by the workflow, not the AI.
The AI does the reading. The rules do the checking. A person handles the exceptions. That’s how we build document capture for supplier invoices, and the same pattern works for orders, timesheets, delivery dockets and statements.
There’s a cost angle too. Every AI step costs money each time it runs and needs watching as models change. Rule-based steps cost almost nothing to run and behave the same next year as they do today. Keeping AI to the steps that genuinely need it keeps running costs and maintenance down.
Where to put the person
Human approval isn’t a sign the automation failed. It’s what makes it safe to automate the rest. Put a person in the loop:
- Before money moves. Payments, credits and refunds.
- Before anything goes to a customer, at least until the output has earned trust.
- Before master data changes. Customer, supplier, product and price records, and above all supplier bank details.
- When the input doesn’t look right, or the AI signals low confidence.
- Before anything that can’t be undone.
It’s the same rule we apply to the AI agents we’re building for our own delivery work. Agents can draft and check, but anything that touches master data or database tables comes to a person, and no agent can deploy a change on its own.
A quick test before you build anything
- Write the process down step by step, including the exceptions people handle without thinking.
- Mark each step as rule or judgement.
- Automate the rule steps first. They’re cheap, reliable and often deliver most of the time saving on their own.
- Add AI only to the judgement steps, with a fallback to a person when it’s unsure.
- Decide where approval sits and what gets logged at each step.
- Run it alongside the manual process for a few weeks and compare results before switching over.
Done this way, you get the reliability of workflows where it matters and the flexibility of AI where it helps, without handing your business processes to something you can’t fully predict.
Where to start
Pick one process that eats hours every week and run the six-step test on it. You’ll usually find that most of it is rules, with one or two genuinely messy steps where AI earns its place. Our integration and automation and AI and agents work covers both halves. Get in touch and we’ll help you map it out.