5 signs your business is ready for an AI agent

01.1. The same decision gets made hundreds of times a week

Volume is what makes automation pay. A decision made five times a month rarely justifies an agent. The build, testing and monitoring cost the same whether it handles ten cases or ten thousand.

Count it. Ask the team how many purchase orders, tickets, invoices or requests they process in a week, and how long each takes. If one person spends half their week on the same kind of decision, that is a strong candidate.

02.2. Most cases follow rules people can explain

If your team can write down "we approve when X and Y, unless Z", an agent can learn it. The rules do not have to be perfect or complete. They have to exist.

A useful exercise is to ask two people to describe how they handle the same request type, separately. If their answers match, the rules are clear. If they differ, you have found an inconsistency worth fixing before any automation. It often turns out that the written policy and the real practice have drifted apart. An agent project forces you to settle it.

Where the rules are truly subjective, such as a creative review or a sensitive complaint, keep a person in charge and let the agent prepare the case.

03.3. The data the decision needs already lives in your systems

Agents act on what they can reach. If the facts sit in someone's inbox, in a shared spreadsheet or in a person's head, fix the data path first.

Check three things:

• Can the agent read the records it needs through an API or a database, without a person exporting files?
• Are the key fields consistent, such as supplier names, product codes and customer IDs?
• Is the data fresh enough? A decision based on last month's stock figure is a bad decision, however clever the agent.

Messy data is not a reason to give up. It is a reason to budget for a data clean-up step inside the project, and to expect more exceptions at the start.

04.4. Mistakes are recoverable

Start where an error can be caught and reversed. Re-routing a ticket is recoverable. Sending money out is not. Marking an order as "needs review" is recoverable. Deleting a customer record is not.

This is the most important filter. You can improve rules and fix data over time. You cannot undo some actions. Begin with decisions where a wrong call costs a few minutes and a correction. Save irreversible actions for later, with a human approval step.

05.5. Someone owns the outcome

An agent needs an owner who reviews its decisions weekly and tunes its rules. No owner, no agent.

The owner is usually the person who runs the process today. They know what good looks like. They do not need to be technical. They need time, about an hour a week at the start, to read a sample of decisions, flag misses and approve changes to the rules.

Projects without an owner drift. Rules go stale, exceptions pile up and trust drops. Then the agent gets switched off and people say "AI does not work here". The tool was fine. Nobody was looking after it.

06.Two candidates compared

Picture two workflows in the same company.

The first is invoice matching in accounts payable. A few hundred invoices arrive each week. The rule is simple: match the invoice to the purchase order and the goods receipt, and pay if they agree within a tolerance. The data sits in the ERP. A wrong match can be reversed before payment runs. Finance has a manager who reviews exceptions. This scores five out of five.

The second is handling sensitive customer complaints. Volume is low, every case differs, the rules are a matter of judgement, and an error can damage a relationship for good. Data sits in email threads. Nobody owns the quality of decisions. This scores one out of five.

Both are real problems. Only one is a good first agent. The second may still benefit from a smaller tool, such as a summary of the case history for the person handling it. That is a different project with lower risk, and it can come later.

The point of the exercise is not to rule things out forever. It is to pick the order. Start where the answer is clear, build trust and skills, then move toward harder work with more experience behind you.

07.A quick way to size the prize

Take the hours your team spends on the workflow each week. Multiply by the share of cases that follow the rules cleanly. That share is the part an agent can plausibly handle. Be honest about the share; most teams guess too high. Sample fifty real cases and count. The result tells you whether the project is worth a scoping call.

08.What the first scoping conversation covers

Expect these questions, so you can bring answers:

• Which systems does the decision touch, and who administers them?
• What does a good decision look like, and who says so?
• What is the cost of a wrong decision, and can it be undone?
• How many cases per week, and what share are routine?
• Who will own the agent after launch?

If you can answer four of the five before the call, you are further along than most. The fifth answer is the one teams most often leave blank.

09.Score yourself

Go through the five signs and give your best candidate workflow a yes or no on each.

• High volume of the same decision.
• Rules people can explain.
• Data already in your systems.
• Recoverable mistakes.
• A named owner.

Three or more? You have a candidate. Start small, measure it honestly, and expand only when the numbers hold.

10.Common mistakes when picking the first workflow

• Choosing the most visible process, not the most suitable. The board may want to see AI in sales. The best first win is often in finance operations or support routing.
• Skipping the baseline. If you do not record time and error rates before you start, you cannot show improvement later.
• Aiming for 100% automation. The goal is to handle the routine majority well and send the rest to people with context.
• Treating it as an IT project. The operations owner has to be in the room from day one.

11.What to do next

Pick the one workflow that scores best. Write a one-page description: the inputs, the rules, the systems involved, the actions allowed and the cases that must go to a person. Pull a sample of past cases. That page and that sample are enough to get a useful scoping conversation.

If you want a second pair of eyes, Brandsmashers can review the workflow with you in a 30-minute call and tell you honestly whether an agent fits or whether a simpler fix would do.