Most AI projects fail because they start in the wrong place. A project that begins with "let's hand all customer communication to AI" becomes, three months later, a system nobody trusts. The right start is more boring and far more profitable.

In a fifteen-person business, the first agent project should take over a job that eats hours every week, has clear rules and makes errors you can measure. Three jobs fit that description in almost every sector.

1. Answering repeat questions

Opening hours, price ranges, delivery times, return terms, appointment availability. These questions arrive every day on WhatsApp, Instagram and email, and someone types the same answer each time.

An agent can answer most of them using information you have approved. The key is that answers come from a current source, not from memory. Keep the information in one place so that when prices change, the agent knows.

How to measure it: share of conversations closed without a handover, and first response time.

2. Receiving enquiries and routing them

When a prospect fills in a form or sends a message, the first minutes are the most valuable. In many businesses that enquiry waits in an inbox for a day, sometimes forever. We wrote about how much response speed matters separately.

The agent does three things here: replies at once, clarifies the need with a few short questions, and logs everything in the CRM while alerting the right person. It does not sell. It puts a ready opportunity in front of the sales team.

How to measure it: first response time and the share of enquiries that become a call or meeting.

3. Information carried by hand between systems

If order details move from the store panel to accounting, form entries to a spreadsheet and meeting notes to a task list by hand, you lose time and accuracy there. Nobody counts it because it is invisible, but a week of logging usually produces an uncomfortable total.

An agent or plain automation takes this over. Sometimes AI is not even needed; a simple integration does it. AI earns its place when the input is messy, such as pulling order lines out of a free-text email.

How to measure it: hours saved per week and errors caused by manual entry.

The first agent project should not be exciting. It should be boring, repetitive and measurable. The excitement comes at the end of month one, when you see the savings.

Which one first?

For one week, have your team count three things: how often the same question was answered, how many enquiries waited more than an hour, and how many times information was copied by hand. The highest number is your starting point. If you want the option closest to revenue, the second job, receiving enquiries, usually pays back fastest.

What to avoid

• Presenting the agent as a human. Customers who know they are talking to AI are less disappointed.
• Letting the agent decide refunds, discounts or special prices without human approval.
• Skipping the success metric. Projects that are not measured get cancelled at the first mistake.

Frequently asked questions

Do I need a technical team for an AI agent?

Building one takes technical skill, but it does not have to sit in your team. What you really need is the person who knows the work best defining the rules: how each question is answered and when to hand over to a person. A partner can handle the technical side, as long as the settings and rules stay understandable and changeable on your side.

Will customers mind talking to AI?

Most do not, as long as they get fast, accurate answers. Discomfort usually appears in two cases: when the agent pretends to be human, and when the route to a person is closed. An agent that introduces itself and always offers "talk to a person" removes most of that risk.

How should I calculate the cost?

Compare build and monthly running costs with the hours the agent takes over and the sales it protects. A week of logging gives you the basis. For revenue-adjacent jobs like receiving enquiries, a single saved customer can cover the monthly cost.

If you want the basics first, read what an AI agent is. To build one with us, see AI agents and automation.