AI agents for small businesses: where to start, and when to go without one

By three conditions being met together: it arrives in writing, answering it requires finding information elsewhere, and a person can review the result in a few seconds. If even one is missing, the agent will produce work that no one can verify.

Let us take them one by one, because this filter determines what follows.

• It arrives in writing. An email, form, attachment, or CRM entry. A request made by phone or handled in the workshop gives an agent nothing to work with unless someone enters it afterward, and then the data entry, not the request, is the real task.
• Answering requires a search. A catalog reference, customer history, supplier price, or availability. This is where the agent saves time: it opens the five places where the information is stored while a person would open only one.
• The result can be reviewed quickly. A proposal, message, classification, or completed record: something you can tell is correct in ten seconds. A task that takes as long to verify as to perform is not worth automating.

Three tasks pass this test in most small businesses that sell through quotes: a price request that arrives by email and must be priced, a follow-up on a quote sent three weeks ago, and sorting the partner or contact inbox that no one really checks.

One condition does not appear in this list because it comes first: the agent must be able to find the correct information. If the same customer has three names in three software tools and no one knows which quote is current, the filter is useless. This is the subject of AI knowledge base: organize what your agent needs to know before changing models.

Yes. The criterion is not headcount but repetition: the task must recur often enough to take a regular share of one person's time. A few inquiries per month do not justify the project, and we will tell you before pricing anything.

You can measure it in your company without any tool: over an ordinary week, count how many times the task comes up and how many minutes it takes each time. If the total fits into one morning per month, keep your money. If it fills one day per week, there is something to address, and that figure, your own, is the basis for everything else. We never make up another one.

One point that lists of “AI use cases” leave out: a task can be highly repetitive and still cost nothing because it is done alongside something else. What really costs a small business is the task that waits: the inquiry received Friday and handled Tuesday, the quote that is never followed up. Start with that one, even if it is not the most frequent.

The agent prepares and proposes; a person approves anything that commits the company toward a customer, supplier, or employee. This boundary is written before construction, not after the first incident.

Read the third column first when comparing providers. An agency that cannot tell you what it refuses to automate has not yet told you where its system stops. For access and permissions, we wrote the questions to ask in four questions to ask before connecting an AI agent to your tools.

In four clear cases: a fixed rule is enough, the process is not described, the task carries a relationship, or the volume does not justify the project. None of these is solved by changing the model or tool.

• A fixed rule is enough. A follow-up that always goes out after the same delay, an acknowledgment, a file placed in the same location every Monday: simple automation does it, costs less, and is easier to verify. A language model adds nothing.
• No one can describe the process. If three team members give three different versions of “how we handle an inquiry,” the problem is not technical. An automated unclear process remains an unclear process, executed faster and without a witness.
• The task carries the relationship. The unhappy customer, negotiation, goodwill gesture, or announcement of a delay: these are moments when the company itself is at stake, and they cannot be delegated to a system that does not bear the consequences.
• The volume is not there. See above: repetition is the fuel. Without it, the project costs more than it returns, and maintenance rarely survives six months of indifference.

There is a fifth case, less common but real: the company that wants to “automate everything” at once. We do not take these projects. Not as a show of caution, but because a broad scope makes it impossible to observe anything, and a system whose performance cannot be determined is eventually abandoned.

With a free twenty-to-thirty-minute discovery audit on a real task, followed by a written scope before any paid work. The first system for one role is delivered in two to three weeks, including trial operation.

• The discovery audit (free, 20 to 30 minutes, by video call or on site): one task, your tools, and the decisions that must remain human. You leave with a written opinion, including if it says to build nothing.
• The written scope: what the agent may do alone, what goes through you, the authorized sources, and a fixed-price quote. When the need extends beyond the task reviewed, we propose a paid scoping engagement whose scope and price are written before work begins.
• Construction: connection to your tools in read-only mode first, rules, output template, and tests on your past cases.
• Trial operation, two weeks included in the timeline: the agent works without sending anything, we compare its proposals with what your teams actually answered, and we make corrections. You decide whether to put it into service at the end.
• Ongoing support: supervision and adjustments each month, with no commitment and one month's notice.

The scope and price are written before work begins, and no paid work starts until you have approved the quote. Hosting and third-party tools are in your name, paid directly by you, with no markup from us. See the framework and offers.

Two questions come up after the system goes live, and both have written answers: what determines the AI usage bill (pay for the task, not the token), and what to look at when the agent takes too long to respond (why your AI agent is slow).

One of the three that most often pass the test: incoming inquiries that need pricing, quote follow-ups, or written prospecting. Each has its own page because each has its own limits.

• Inquiries and quotes. The agent reads the incoming message, recognizes the customer, finds the references, and prepares the proposal with its sources; the employee approves with yes or no. This is the field of the inquiry and quote agent.
• Follow-ups. A quote that was sent and never followed up is money that has already left the company. See follow up on your quotes without spending every Friday on it.
• Prospecting. A prospecting agent applies your criteria, identifies contacts, prepares a message based on a verifiable observation, follows up, and keeps the CRM up to date.

If your need does not resemble any of the three, it probably belongs to custom automation, and there is a serious chance that a fixed rule will be enough. That is good news: it costs less.

Small and medium-sized businesses where part of the daily work arrives in writing, anywhere in France, with an on-site meeting available in Île-de-France.