Most small-business owners I work with start the day the same way: they open WhatsApp and spend an hour wading through noise before they can actually think. Emails that could filter themselves, the same questions over and over, weekend leads sitting unsorted. My job is to build AI agents that handle that noise (they research, reply, route) so that when you sit down to work, what’s on your screen is the stuff that actually needs your head.

What is an AI agent?

An AI agent is a system that doesn’t just answer: it decides and acts. Unlike a chatbot, which replies to questions, an agent researches, chooses the steps to reach the goal you set, and carries out real actions —booking an appointment, routing a lead, updating your CRM— within the limits you give it, no code required.

The word “agent” is new but the idea isn’t that strange. A plain chatbot answers: someone asks about an order, it checks the table and replies “your order is on its way.” An agent, instead, acts. It sees the order is delayed, messages the courier to speed it up, warns the customer with an apology, and applies a discount if you set it up that way. It’s the difference between someone who can read a spreadsheet and someone who can solve a problem.

Anthropic (the company behind Claude, the model I use most) describes it as a system where the AI chooses its own steps to reach the goal you gave it. You define the outcome and the limits; the agent decides the how within that.

It gets confused with three different things. The chatbot I just mentioned. Classic automation, the Zapier or Make kind, which runs from A to B along a fixed path and breaks the moment something odd shows up —and in any business, something odd always shows up. And RPA, those programs that mimic clicks on screens and fall over when the site changes its layout. Agents are different because they understand context, decide, and adapt.

How it works under the hood

Inside an agent there’s a language model acting as the brain (almost always Claude), a set of tools it uses to act (send a WhatsApp, book an appointment, log a lead in your CRM, read a sheet) and a memory that lets it recall the conversation, the customer, and your business rules. When a task comes in, it thinks, acts, looks at what happened, and decides the next step. If something falls outside the script, it escalates to a human. Usually that human is you.

I orchestrate all of it with n8n, a visual platform where flows are built by connecting blocks. It’s not lines of code, it looks more like a diagram, and it plugs into almost everything you already use: WhatsApp, Gmail, Calendar, Sheets, Airtable, HubSpot, whatever. I don’t ask anyone to switch tools. The agent fits the stack you already have.

One thing I learned on the job, and that Anthropic says too: the best agent is the simplest one that solves the problem. You start small, see if it works, and grow only when there’s evidence it’s worth it. Sounds obvious, but most consultancies sell the opposite path.

What it does for your day to day

The cases where an agent gives back the most time and money are always the same.

Clinics and practices are the most measurable. WhatsApp appointment confirmations, scheduling, reschedules, reminders. Automated reminders cut no-shows: in clinical studies, by roughly 10 to 25%, and more when they target the patients most likely to miss (2026 meta-analysis). Any practice with more than thirty appointments a week pays it off in a month.

Schools and educational institutions have two fronts. Parent questions, which repeat all year and eat up front-desk time. And teachers, with the part they struggle with most: lesson plans, announcements, written feedback. I come back to this in the examples.

E-commerce. My favourite number here is from Baymard Institute: 70% of carts are abandoned. A well-built agent recovers part of that: it detects the abandonment, sends a personalised message (not the generic “you left something in your cart”), answers the real question (shipping, returns, delivery date) and closes. In post-sale it takes tracking and repeat questions off your plate.

Professional services: accounting firms, law practices, consultancies, agencies. The complaint is always the same: hours lost on prospects who never buy. The agent takes each enquiry, researches the lead, qualifies it, logs it to the CRM with a summary, and routes. The hot ones land on your calendar; the warm ones go into an automatic sequence; the cold ones sit on a list for later.

Freelancers and solopreneurs. If you work alone, this is how you clone yourself. The agent researches the client before the first call, drafts the first version of the proposal from your notes, filters your inbox. McKinsey measured that people who use AI for communication and documentation save 3.4 hours a week. In a one-person business that’s nearly half a day.

Associations, chambers, NGOs. Memberships, renewals, member communication. Less flashy, but it’s where it shows the most: small structures running big operations.

What I build them with, and why

There are plenty of platforms for building agents today. I choose n8n for one concrete reason: when the project ends, you own the flow, not the platform. n8n is open source and can run on your own server, so your data doesn’t pass through an outside company, which matters if you handle medical records, student data, or sensitive client information. It charges per execution, not per individual step, so complex flows stay viable. Behind it is a European company backed by funds like Accel and Sequoia, which isn’t going to vanish tomorrow.

The brain of almost all my agents is Claude. Today it’s the model that reasons best on tasks where you have to decide several steps in a row, and it writes better than ChatGPT (an opinion, but one backed by a lot of testing). For specific cases I use GPT or Gemini.

When is n8n not the right call? If your case is three steps and two integrations, Zapier is faster to get going. But most real cases have conditional logic, sensitive data, and more than two systems. There, n8n wins hands down.

How long it takes to ship

Timelines vary by case but the ranges are fairly stable. An agent that does one thing well (WhatsApp appointment confirmations, say) takes one to two weeks. A lead-routing agent that touches a form, a CRM, and a calendar, one to three. And one that coordinates several tools with edge cases and human approval on sensitive steps, three to eight.

What’s almost never the bottleneck is the AI. What stretches projects out is getting access, cleaning dirty data, finding the odd cases that didn’t show up in the first conversation, and testing. That’s why we always start with a short diagnostic: if yours is a two-week job, there’s no sense packaging it as a two-month project.

Real examples

Bayard Writer. A school asked me for help with the writing load on its teachers —lesson plans, family announcements, student feedback— which was eating entire afternoons. But the real problem wasn’t writing fast: it was writing like the school. Every institution has a voice, a set of values, and its own ways of doing things, and a generic AI draft forces you to rewrite the whole thing so it sounds “like Bayard.”

So I didn’t build a writing assistant pointed at a school: I built the school into it. Bayard Writer has the institution’s full context loaded: its identity and values, the pedagogical approach, the tone of each kind of communication (it doesn’t write to a family the way it writes to a student), and the internal conventions for how each level and area is handled. It knows “how things are done here” as well as a teacher who’s been in the building for years.

The teacher gives it the idea in one line plus the context of the case; the assistant returns an announcement already in the school’s voice: not a draft to translate, but something ready to review and sign off in minutes. The AI doesn’t invent a new voice, it reproduces the one the school already had.

AI lead routing. Another case was for a business that got enquiries through several channels and was burning its days replying to prospects who never bought. Now the agent takes each lead, researches it with public info, qualifies it, logs it to the CRM with a summary, and routes the hot ones straight to the owner’s calendar. When they sit down to work, they already have a prioritised list instead of an unopened inbox. And speed matters: in a field study of 114 B2B companies, not one replied to a lead in under five minutes (Workato, 2024). The agent does it instantly.

For a sense of scale: Delivery Hero cut two hundred hours a month with a single n8n flow. It’s not comparable to a small business, but it illustrates the point. In smaller contexts, a good agent gives back five to fifteen hours a week according to HubSpot’s 2025 reports.

Frequently asked questions

Does this replace my people?

No, unless your people spend their day copying data from one system to another. And if that were the case, you had a problem before the AI. What the agent replaces is the repetitive part. Your team keeps doing the work that needs judgement: handling the important client, making the unusual call, dealing with the case that doesn’t fit. McKinsey found that 43% of companies don’t expect headcount reduction; in small businesses they even expect AI to help them grow.

Is it safe? What about the data?

It’s the question I get most, and it’s fair. With n8n running on your own infrastructure, the data doesn’t leave to a third-party platform. It’s what European regulators recommend: data minimisation, human approval before sensitive actions, transparency.

What if the AI gets it wrong?

It can get it wrong. Models sometimes “hallucinate,” which is the technical term for making things up. The way to handle that isn’t to expect them to be perfect, but to design the flow so the agent doesn’t make important decisions on its own. Any sensitive action (an email to a key client, a change to a critical record, a charge) goes through human approval before it runs.

Do I need to know how to code?

No. You define the problem, you validate how it works, I build and maintain it. If you later want to learn to tweak the flows yourself, I’ll show you, but it’s not required.

Do I own the solution?

Yes. n8n is open source, the flows stay documented, the data is yours, and if you ever decide to part ways with me you take it all with you. It’s the difference between renting a tool and building something of your own.

If you read this far and found yourself thinking “this could work for X,” send me a message and we’ll set up a free twenty-minute call. You can also try Kyn, the site’s assistant: it’s built with the same approach and can help you think through your first flow.