AI agents or workflow automation: which does your process need?

Every vendor you've spoken to uses both terms, and most use them to mean whatever they sell. That leaves you choosing between two things nobody has defined, for a process you know better than any of them.

The difference comes down to one question you can ask about each step in the process: does the input arrive in the same shape every time?

What each one does

Workflow automation

Automation runs a fixed sequence. A form is submitted, so a record is created in the CRM, a confirmation goes to the customer and a task lands with the account manager. Every input has the same fields in the same places, and every run does the same thing with them. It's fast, cheap to run and easy to test. When something arrives that doesn't fit the rule, it stops or does the wrong thing, and it can't tell which.

AI agents

An agent is given a goal and the tools to reach it, and works out the steps itself. Read this invoice, whatever its layout, and return the supplier, the amount and the purchase order it belongs to. Read this email and decide whether it's a complaint, a question or a new order. Draft a reply from these three approved documents. The input varies, so the handling has to vary with it. That flexibility is why agents are useful, and it's also the risk: an agent can be confidently wrong in a way a rule never is.

A test you can run on any process

Write the process down as a numbered list of steps, the way you'd explain it to a new starter. Then go through the list and ask of each step: could I write the rule for this on one line? "If the total is under the approval limit, approve it" is a rule. "Work out what this customer is asking for" is not.

The steps with a one-line rule are automation. The steps without one are where an agent earns its place. In most back-office processes the split is lopsided: most steps are rules, and one or two need reading or judgment. That's the shape to build, with the agent handling only its own steps and the automation carrying the rest.

If every step has a rule, you don't need an agent at all, and adding one makes the system slower, dearer to run and harder to trust. If no step has a rule, the process probably isn't ready to automate, because nobody can yet say what done correctly looks like.

Where each one fits

The same handful of processes come up in companies of a few hundred people, and they split cleanly.

• Supplier invoices. Routing, matching to the purchase order and queuing for approval are rules. Reading the invoice itself, in whatever layout the supplier used, is an agent step.
• Orders by email. Creating the order in the ERP and confirming it to the customer are rules. Reading a PDF or a free-text email to find the items, quantities and delivery date is an agent step. Processing orders from email and PDF goes through that build.
• Support tickets. Pulling the customer's order history into the ticket and sending the reply are rules. Working out what the customer wants and drafting the answer from approved information are agent steps.
• The weekly report. Almost entirely automation. The numbers come from fixed places and the layout doesn't change. An agent can add a written summary, and that summary needs checking.

One enquiry, step by step

Take a new enquiry from your website form and follow it through. The form arrives and a record is created in the CRM: a rule. The company name is checked against existing accounts, so a returning customer isn't treated as new: a rule. The free-text message is read to work out what they want, how large the job is and whether it fits the criteria sales set: an agent step, because the message could say anything. A first reply is drafted from that reading and your approved wording: an agent step. The draft and the record go to the right salesperson by region: a rule. The salesperson reads the draft, changes a line and sends it: a person.

That's six steps: three rules, two agent steps and a person at the end. The automation carries the enquiry from the form to the salesperson's screen without anyone touching it. The agent reads and drafts. The salesperson decides what goes out. If the reading is wrong, the salesperson sees it before the customer does, and the correction is recorded so the next draft improves. Nearly every good first system has this shape.

Where the person goes

The agent's step is the one most likely to be wrong, so it's the one that most needs a check. Hold to one rule: anything an agent produces that could reach a customer, a supplier or the accounts goes in front of a person before it does. The person sees the draft, the source it was built from and how sure the agent was, and approves or edits it in seconds. Routine runs take none of their time. Their attention goes to the exceptions.

That check is also what lets the system grow. Once the team has watched the agent's drafts for a while and seen how often they need a change, they can decide which kinds could go out unchecked. That decision stays with them.

Questions that cut through the pitch

When someone proposes a system, three questions tell you whether they've looked at your process or only at their product.

1. Which steps are rules, and which does the agent handle? If they can't split the process, they haven't looked at it.
2. What happens when the agent is wrong? A good answer names the person who sees it, what they see and how the correction is used.
3. Why does this step need an agent? If the only answer is that agents are more capable, a rule would have been cheaper and safer.

What to do this week

1. Pick the process you'd most like off your team's plate and write it as a numbered list of steps.
2. Mark each step as a rule or a judgment, and count both.
3. If the judgments are one or two, you have a system worth building: automation with an agent inside.
4. Write down who would check the agent's output, and what they'd need to see to check it in under a minute.

Splitting the steps this way is the first thing we do when we build a system. Bring the list and we'll go through which steps are rules, which need an agent, and whether it's worth building yet. If you're still choosing between processes, start with which AI use case to fund first.

Get the next article when it's published.