"We want to add AI." That is often the first sentence in a meeting. The second, more honest one: "but we don't want it sending just anything to our clients".
Both are right. The latest generation of AI models can now read an email, extract the useful information, draft a follow-up or summarise a month of activity. But an SME lives on its reputation. A badly worded message sent to a hundred clients costs more than the hours it was supposed to save.
The answer comes down to one simple rule, which we apply on every project: AI prepares, a human approves.
What an AI agent is (and is not)
In an SME, an AI agent is not a chatbot that "decides". It is an automated task that:
- is triggered by an event or on a schedule (an incoming email, an offer with no reply, the end of a job, Monday morning);
- reads the relevant data in your tools (CRM, ERP, inbox, calendar);
- produces something: a record update, a draft email, an alert, a report;
- and, depending on the level of risk, applies it on its own or submits it for approval.
In one project we delivered for a technical SME in French-speaking Switzerland, around forty agents run in parallel, for roughly 64,000 automatic executions a month. None of them is allowed to write directly to a client without a framework defined in advance.
Six automations that genuinely work
1. Automatic inbox reading
This is often the most profitable automation. Sales staff exchange emails with clients but forget to update the CRM. An agent reads new messages, recognises those relating to an open deal (agreement, postponement, question, refusal) and updates the record. Observed delay: between 3 and 14 minutes after the email arrives. The salesperson has nothing to do, and management sees an up-to-date pipeline.
2. Signature and payment reminders
An offer that never comes back, a document waiting to be signed, an overdue invoice: these are reminders everyone hates sending. An agent schedules them at fixed intervals (for example D+5, D+12, D+26 for a signature) and, for payments, across several levels of firmness. The wording is approved once, upfront, by management. The salesperson receives a copy and can pause the reminder in one click.
3. Client review requests
At the end of a job, the client is happy. That is the moment to ask for a review, and precisely the moment nobody thinks of it. An agent triggers the request when the file moves to "completed", then follows up at spaced intervals, never pushing beyond a set number of attempts. It stops as soon as the review is published or the client unsubscribes.
4. Nurturing during office hours
Not every prospect is ready to sign. An agent sends useful content (not promotions) to those who have gone cold, only during office hours, with a gradual increase in volume to protect the sending address's reputation. In a first test of 103 emails, the click-through rate was around 17 %. The sample is small and we quote it with caution, but it shows that a useful, well-timed message gets read.
5. Reports that write themselves
Monthly lost-deals report, lead attribution by marketing channel, real campaign profitability: an agent gathers the data, and an AI model writes the summary in plain language, with the main causes and possible actions. Management reads a clear summary instead of exporting spreadsheets.
6. Agents that monitor the system itself
The most useful ones are sometimes invisible: regular comparison of the CRM with the ERP to detect discrepancies, detection of "silent" offers that have stopped moving, a continuous-improvement agent that proposes three to five changes each week ranked by impact, or a "dead weight" agent that flags what is no longer used and can be removed.
The lesson of ignored alerts
Here is the mistake we made, and that many make.
At first, every important event triggered an alert: unhandled lead, offer without a reply, stalled file. Logical. Except that after a few weeks, we measured how recipients actually reacted: fewer than 7 % of alerts led to any action.
The problem was not the quality of the alerts. It was their number. One more notification in a busy day is not information; it is noise.
The solution was to replace individual alerts with a capped daily digest: a single message, sent at a fixed time, with the few truly priority actions, ranked. Not twenty lines. Three to five. And for each one, a button to act directly from the email.
The moral: automating detection is easy. Automating attention is impossible. You have to design for the person receiving, not for the system sending.
How to stay in control
A few principles we apply systematically:
- Three levels of autonomy. Internal, reversible tasks (updating a record, filing a document) apply on their own. Client communications follow a template approved in advance. Everything else is a proposal a human approves in one click.
- A log of every execution. Each agent leaves a trace: what it read, what it did, what it proposed. If something goes wrong, you can trace the chain back quickly.
- Circuit breakers. Rate limiting, a daily sending cap, instant deactivation of one agent without touching the others.
- Role-based permissions. When a team member queries the CRM through their AI assistant, they only see what their role allows, and any write remains a proposal.
- Measured supervision. In the case mentioned above, the automations save around 67 hours of work a month, and supervising them takes about 4. That ratio is the right indicator: if supervision time climbs, an agent is badly designed.
Where to start
Do not start by "putting AI everywhere". Start by listing the repetitive tasks that cost you the most time and follow clear rules. You almost always find the same ones: updating the CRM after an exchange, follow-ups, review requests, monthly reports. They are also the ones where the risk is easiest to contain.
To see how these agents work alongside a CRM, a client portal and a field app, visit our AI automations page or the full client case. And if you are still unsure about the underlying tool, our article Bespoke CRM, SaaS or a layer on top of your ERP will help you decide.