AI agents10 min
AI automation for a small business: where it pays back first
AI automation for a small business: the five processes that pay back in under three months, with numbers, tools, mistakes, and what setup costs.
If you have between 3 and 40 people, AI automation pays back first on five things and with my clients it has been the same five every time. The quotes, the reminders for unpaid invoices, the appointments, the leads that have to go into the CRM and the emails that ask the same questions all day. Most of these businesses use AI already, because somebody has ChatGPT open in a tab. But nothing is connected to anything.
At my studio a first automation starts at €900 before VAT to set up, around $1,000, plus a monthly cost that follows the volume. It takes one to three weeks.
One that I did recently was for a boatyard here in La Ciotat. The owner told me how he lost a job in July 2025 because of one email. The email came from a Belgian, at 11:52 on a Saturday night. He wanted antifouling and a check of the bow thruster on his motor yacht (14 meters), before the 20th. The office answered on Monday at 10:15, and for a yard in July that is not slow, but he had already booked in Toulon.
What AI automation is, and what it isn't
Most of what you will see sold as AI automation is in fact rules. Rules are fine, and I use them every week. Zapier, Make and n8n have done this for a long time: somebody fills in your form, a contact is created, a confirmation goes out and you get a notification on your phone. If what comes in is clean it can run for years and you never think about it.
But a lot of what you receive is not clean. You get an email typed on a phone, with the length of the boat in feet in one sentence and the dates two lines further down, and your rule can do nothing with it because it does not know which part is the date. A language model does. So you use the model only for this step. It reads the message and picks out the fields, then your rule continues as before.
If your software also has to decide the order of the steps and when to ask a person, it is an agent and it is a bigger job. I explained the difference between an agent and an automation in another article. For four of the five processes below you do not need one.
Everyone has tried it, almost nobody has wired it in
58% of small businesses use generative AI, says the U.S. Chamber of Commerce in its report of August 2025, and a year earlier it was 40%. Fine. With my clients it is usually a chat window to write product descriptions. The inbox and the calendar still work like it is 2015.
FranFunnel studied 530 franchise brands in March and April 2025 and only 25% sent a personal reply in less than five minutes, and the ones that replied by email took 9.1 hours on average. And these are brands where one signed franchise brings in $250,000 or more in fees and royalties, according to the same study. So a Saturday night email that waits until Monday in a company of five people is normal. It is also where you lose work, and it is the first thing I would fix.
Five AI automation examples that pay back inside three months
I did these five myself in the last two years, for small businesses between Marseille and Toulon in the south of France. The amounts are on the invoices. The owners counted the time themselves, usually on paper next to the phone.
Quotes
Start here if you lose jobs because your quotes go out late. The owner of the boatyard wrote his quotes at the office, from memory, four days after he saw the boat. Now he talks for two minutes into a voice note on the pontoon and the automation makes a draft quote out of it in his invoicing software, with the prices of his rate sheet. He reads the drafts in the evening, he changes one or two lines and they go out the same day.
In the first three months it was 47 quotes against 29 in the same months of the year before, with the average job still around €3,100, and he signs about one in three.
If you do this, put in one rule from the first day: every line of the quote must match a row of your rate sheet. His sheet has 290 lines and in the first weeks, if the model did not find a price, it made one up that looked correct. An empty line is better than that. Now the line stays empty with a note in red and he fills it in himself.
Invoice follow-ups
If you chase your unpaid invoices yourself, you know how it goes: it is always next week. At a landscaping company near Aix it was the owner's wife who did it on Friday afternoons, for about forty business clients, with around €62,000 overdue at all times. Now the reminders on day 3 and day 12 go out on their own and she reads the one on day 25 before it is sent. When a client answers, the AI reads the answer and files it under paid, will pay or problem. Six months after we started the amount overdue was €24,000, and she gets four hours a week back.
Appointment booking
If you work with your hands all day you do not answer the phone, and when you call back at 7 p.m. one person in three has gone somewhere else. This is what happened at a two-person car-detailing shop in Aubagne. You do not need a receptionist for this: a text, a WhatsApp message or your website form can get three real slots from your calendar in two minutes and the booking is confirmed without you. For them it was €900 before VAT to set up and €16 a month, and twenty-eight more bookings in the first month.
CRM data entry
If your leads come by email from websites, somebody probably types them into the CRM when they have time. At an estate agency on the Vieux-Port it was the junior agent. The leads come from three property portals, each with its own email format, and for weekend leads it was usually Wednesday. Now they are in the CRM on Monday at 8 a.m. The automation reads the emails and creates the contact with the budget, the area, whether the financing is there and a date to call back. What you do with these leads after that is lead qualification with an AI agent, another job.
Repetitive email replies
If you rent out boats you know the questions. A boat rental company at the port in La Ciotat gets 40 to 70 emails a day in July. I read two weeks of the manager's inbox with her before doing anything, and four out of five were the same seven questions: license, deposit, fuel, return time, the mistral, children and parking. The automation answers those with a link to the right page of the company's site. For the rest it prepares a draft and she approves it on her phone. The average reply time was 11 hours before and it is 4 minutes now.
Tools, and the ones you can skip
n8n most of the time, Make if the client already pays for it, Zapier rarely, because its price goes up fast with volume. For what I describe here the three can do the same things, and I compared n8n, Make and Zapier in another post. The reading is done by a language model through an API and you pay for it by the amount of text.
You do not need a new CRM, a new invoicing tool or an all-in-one platform that you pay for per user. The five setups here run on the software the businesses already had, and two of those programs are old. One had no API at all, so we read the notification emails it sends. It is not elegant but it works. If you change your software before you automate, usually you only delay the project.
What AI automation costs to set up
At my studio a first automation starts at €900 before VAT. This is what the detailing shop paid and they had it in ten days. Usually it is one to three weeks. For that you get the scoping, the connections, the rules about what the model may send without you, two weeks where I read the logs every morning and one hour of training. If your software has no API it costs more, because we have to work around it.
After that you pay the model by volume and not by user. It was €16 a month for the detailing shop and about €100 in July for the boat rental company. For the boatyard and the landscaper, the setup had paid for itself before the end of the second month. I did not need a spreadsheet to see it.
More than 40% of agentic AI projects will be canceled by the end of 2027, Gartner predicted in June 2025, because of escalating costs and unclear business value. I think they are right, and it will be mostly projects that start from the technology. If you start from a number you already track, like the days your quotes take, after one month you know if it works.
The mistakes
At the estate agency the first meeting was in October 2025. I asked what happens to a lead when it arrives from one of the portals. The junior said he forwards them to his own inbox, the office manager prints them and puts them in a tray, while the boss was sure they went into the CRM by themselves because this is what the salesman from the portal told him in 2022. It took about one hour at the whiteboard to agree on one way of doing it, before I connected anything.
At the boatyard it was the rate sheet. There were three prices for antifouling on the same size of boat (lines added in different years, and nobody ever deleted the old ones) and the model took the oldest. Three quotes went out with it before somebody noticed, one to a man who has kept his boat at the yard since 2011. He called the office to ask if this was a joke. The daughter of the owner spent two days cleaning the sheet after that. We should have done that first.
The boat rental manager asked me in August if she could stop approving the replies, after three weeks without a mistake. My answer was no for the replies with a price in them and the replies to complaints. You can let the model answer the parking question on its own, but not the angry customer. And I still read the logs on Monday morning.
Where to start on Monday
You count for one week, on paper if you like. The quote requests you receive and the quotes you really send, the overdue invoices and when you chased them, the calls you miss, the emails you answer after 8 p.m. You start with the process where the number is the biggest and the delay the longest; for my clients it was the quotes or the emails eight times out of ten.
Then you write down in ten lines how this process goes when it goes well, because these ten lines will be the instructions for the model. For the next steps, see how to create an AI agent for your business, and what the studio has built is in the portfolio.
If you want a second look at your numbers, bring the paper to a 15-minute call.
FAQ
What is AI automation?
For me it is a normal automation with rules, plus a language model for the one step that the rules cannot do. Most of the time this step is reading free text (an email, a voice note or a scanned document) and turning it into data for the rules. On my projects it is maybe a tenth of the work.
How much does AI automation cost for a small business?
At my studio a first one starts at €900 before VAT, about $1,000, and it takes one to three weeks. After that the model costs between €16 and about €100 a month on the projects I described here, depending on how much text it reads.
Is it safe to let a model read customer emails?
Yes, if you choose where the model is hosted, you send it only what it needs and you mention it in your privacy policy. With medical or legal data I decide the hosting with the client before anything else.