A chatbot answers. An agent acts.
That's the difference in one sentence. A chatbot receives a question and returns an answer. An agent receives a goal, decides which steps are needed, uses tools (your CRM, your calendar, your email, your database) and executes those steps until the goal is met. And if something goes off script, it asks or stops.
Concrete example. A customer writes on WhatsApp: "I need to move my Thursday appointment to next week in the afternoon."
- A chatbot replies: "To change your appointment, call 800 000 0000."
- An agent looks up the customer's appointment in the calendar, checks afternoon slots the following week, proposes two options, confirms the chosen one, updates the calendar, sends the confirmation and notifies the professional. No human involved. In 40 seconds.
This isn't the future. I'm deploying it right now in clinics, professional firms and service companies.
What agents do well today
I've tested many things. These are the ones that work reliably in small and mid-sized companies:
Appointment and booking management. Book, reschedule, cancel, remind. With access to your real calendar. The most mature use case and the fastest return.
Lead qualification. When an inquiry comes in, the agent asks the questions your salesperson would ask on the first call (what they need, when, what budget), logs it in the CRM and alerts the right person with a summary. The salesperson joins the conversation already knowing who they're talking to.
Sales follow-up. Quotes sent with no reply, customers who haven't bought in months, renewals coming due. The agent spots the case, drafts the right message in your company's tone and, depending on how you configure it, sends it or leaves it ready for approval.
Document processing. Invoices, delivery notes, contracts, forms. The agent reads, extracts, validates against your data, flags inconsistencies and loads into the system. Anything doubtful gets marked for review.
Research and preparation. Before a meeting with a customer, the agent pulls together their history, recent purchases, open issues and recent news about their company. Delivered as a one-page summary.
First-level support. Answers from your documentation, resolves the routine, escalates the complex with all the context collected. Your team handles only what deserves human attention.
Where not to put an agent (yet)
Same criteria as any automation, but stricter because the agent acts:
Nothing that moves money without approval. An agent can prepare a payment, a refund, a discount. A person executes it with one click. Always.
Nothing with legal consequences. Signing, accepting terms, committing to contractual deadlines. The agent prepares; someone accountable decides.
Nothing irreversible without confirmation. Deleting data, cancelling orders, closing customer accounts. If it can't be undone, it goes through a person.
Poorly defined processes. An agent amplifies what you give it. If the process has exceptions your own team can't resolve, neither can the agent.
A good agent knows when it doesn't know. Part of designing it is defining precisely in which situations it stops and alerts a person. An agent that never asks is a badly configured agent.
What an AI agent costs for a small business
Indicative figures from what I deploy:
WhatsApp appointment agent connected to your calendar: €2,500 to €4,500 to implement. Monthly usage (AI models and messaging): €40 to €120 depending on volume.
Lead qualification and sales follow-up agent connected to your CRM: €4,000 to €8,000. Monthly: €60 to €200.
Document processing agent loading into accounting or ERP: €4,000 to €9,000 depending on document variety. Monthly: €50 to €150.
Add maintenance on top: models improve, your processes change, new cases appear. Budget 10 to 15 percent of the implementation per year, or a monthly plan that covers it.
For perspective: an appointment agent that prevents 15 missed appointments a month in a clinic with an average ticket of €80 recovers €1,200 a month. It pays for itself in three months and keeps working after that.
How one is built (version for non-technical readers)
Four pieces:
-
The language model. Claude, GPT or another. It's the "brain" that understands and reasons. Chosen according to the case, the language, the cost and the privacy guarantees.
-
The instructions. A document defining who the agent is, what it can do, what it can't, how it speaks, when it escalates. This is what takes the longest to get right and what separates a useful agent from an embarrassing one.
-
The tools. The connections to your calendar, your CRM, your email, your database. Each tool has specific permissions: read yes, write here yes, delete no.
-
The supervision. A log of everything it does, alerts when something falls outside the norm, and a person who reviews the first few hundred interactions.
You don't need to understand the technical side. You need to understand that points 2 and 4 are where the result is decided, and that someone experienced should design them with you.
A case: professional firm with 12 people
Problem: every week 60 to 90 new inquiries came in via web, email and WhatsApp. One person reviewed them, replied, qualified and routed them. First reply took one to two days. They were losing clients to slowness.
Solution: an agent that replies in under a minute, asks three qualifying questions, logs to the CRM, assigns to the right lawyer by subject and books the first call in their calendar. Anything that doesn't fit a category goes to the person with a summary.
Result after three months: first reply time from 36 hours to 1 minute. 30 percent more first meetings booked. The person who used to route inquiries now handles follow-up with existing clients, which was what nobody was doing.
Cost: €5,500 implementation and about €90 a month. Recovered before the second month.
Where to start
If you want to try an agent in your company, my recommendation is always the same:
- Start with appointments or lead qualification. They're the most proven cases, with the least risk and the fastest return.
- Run it in supervised mode for the first month: the agent proposes, a person approves. You see how it responds and adjust the instructions.
- Measure from day one: response time, conversations resolved without a human, appointments booked, errors.
- After a month without surprises, give it more autonomy.
That's what I do in the AI process automation service, and also within the monthly partnership for companies that want to add agents progressively.
In short
AI agents are already useful for small companies, as long as they're applied to clear processes, with limited permissions and supervision at the start. They're not magic and they don't replace your team's judgment. They replace the repetitive work that currently stops your team from using that judgment.
If you have a specific process in mind and want to know whether an agent would solve it, book a call. In 15 minutes I'll tell you whether it's feasible, what it would cost and what return to expect.
