Optimising business processes in SMEs: specialised AI, generative AI, data and integrations

Artificial intelligence came into small businesses through writing: an email to rephrase, a text to produce, a question to ask. It still rarely runs a process. Yet optimising a process starts with moving information without re-keying, then handing what remains repetitive to the right technology. There are three levers for that, and generative AI is only one of them.

of French micro and small businesses use at least one AI tool in 2026

40%

of French micro and small businesses use at least one AI tool in 2026

13% in 2024

to generate text, voice or images

34%

to generate text, voice or images

the most common use

to automate tasks

9%

to automate tasks

one of the least common uses

AI writes more than it automates

In 2026, 40% of French micro and small businesses say they use at least one artificial intelligence tool, against 13% in 2024, according to the France Num 2026 survey (in French) by the French Directorate General for Enterprise. The rate reaches 53% for companies with 10 to 249 employees. But uses remain concentrated on content production and chat assistants.

The same survey notes that, to run their business on their data, companies rely first on their business software (ERP, CRM), and that AI is not identified there as a way of using data. In other words, the room for progress lies less in one more tool than in how existing tools talk to each other and in handing repetitive tasks to the right technology.

Three levers to optimise a process

1. Specialised AI: precise, local and compliant

A specialised AI is trained or tuned for a single task: read one type of document and extract its fields, classify files, spot an anomaly in an image, forecast a volume. It does not chat; it produces a result you can check on examples and measure.

What it does well

• Reading documents: invoices, purchase orders, notices, scanned forms, with character recognition and extraction of the useful fields.
• Sorting and classifying: routing a document or an email to the right folder or person based on its content.
• Checking and detecting: industrial vision for quality control on a line, spotting deviations in readings.
• Forecasting: order volumes, restocking needs, maintenance of equipment from its operating data.

Why run it locally

Because it is cut for one task, a specialised AI can often run on a company workstation or server. Documents do not leave your machines: no transfer outside the European Union to manage under the GDPR (Chapter V), no model provider in the chain. It also needs less computing power than a large general-purpose model, and its running cost is predictable. That is how we automated a recurring document-processing task at Techmo Hygiène: character recognition runs locally, with no third-party service. The French data protection authority, the CNIL, has published practical guidance (in French) on applying the GDPR to AI systems, useful to frame this kind of project.

Its limit: it only does what it was tuned for. It needs real examples to be set up, and human review of the cases it flags as uncertain.

2. Generative AI: routines, agents, APIs and channels

Generative AI understands and produces language. In a process, it helps where information is free text and each case is a little different: a customer request worded their own way, an exchange to summarise, a reply to prepare. It is used in four ways.

Routines

A scheduled task, always the same: every morning, summarise the requests that came in overnight and rank them by urgency; every Friday, prepare the week's report from the team's notes.

Assistants in your channels

AI placed where requests already arrive: email, Teams or Slack, the website form, WhatsApp. It answers simple questions, collects missing information and passes the rest to the right person.

Agents

An agent chains several actions with tools: read a quote request, find the customer in the CRM, check prices, prepare the quote. The more the agent acts, the more human approval must sit at the right places, especially before anything is sent to a customer.

APIs

The model called directly from your software: one step of a flow between your software (see part 4) handed to generative AI, for example pulling the useful details out of an email before entering them in the ERP.

3. Gather, clean, filter and analyse your data

Gathering, cleaning, filtering and analysing data is often the longest part of a process, and the least visible: pulling together scattered information, de-duplicating a customer file, isolating late orders, reconciling two exports, reading hundreds of reviews. Every step can be automated, with rules or with AI. According to the France Num 2026 survey (in French), only 14% of French micro and small businesses use AI to analyse or classify documents, and 11% to analyse data or forecast.

Order matters: generative AI queried on messy data gives messy answers. Start with data that is gathered, cleaned, filtered and reconciled, then add the analysis.

4. Connect your tools: SaaS integrations, APIs and connectors

Most lost time comes from information that already exists in one piece of software and that someone copies into another. The first lever is to make that software talk, without AI: when a rule is fixed, a program applies it identically, every time.

Native connectors

Many online (SaaS) tools offer ready-made integrations: your CRM to your invoicing tool, your online shop to your accounts. It is the fastest option when it exists; it rarely covers the particular cases of your business.

Integration platforms

Platforms such as Zapier, Make or n8n connect hundreds of applications through scenarios: “when a form arrives, create the contact, alert the sales rep.” They suit simple, frequent flows. Their cost often rises with the number of operations, and a poorly documented scenario soon becomes a black box.

Custom APIs

When connectors are not enough, an API or a small dedicated program makes the link: read incoming purchase orders, enter them in the ERP, check stock. That is the automation built for the distributor 3plus, which the client puts at a saving of about €1,000 a month (3plus case study).

Which lever for which task

1. 1

   Documents or images must be read in volume

   Specialised AI, run locally when the data is sensitive: extraction, sorting, checking, with a person on uncertain cases.

2. 2

   The information is free text and every case differs

   Generative AI: routines, assistants, agents or APIs, with human approval before anything that commits the company.

3. 3

   The data is scattered or messy

   Automated gathering and cleaning first, analysis second: no AI makes up for wrong data.

4. 4

   The rule is fixed and the data structured

   A program applies the rule, with no AI. Prefer it to one more subscription when the flow is specific to your business.

5. 5

   The process mixes all of this

   The most common case: generative AI reads the email, a program enters the order, a rule checks the stock. Each step gets the lever that suits it.

GDPR and the AI Act: what each lever implies

How we work

We start from the process, not the technology. In a free diagnostic, we read one task with you, step by step, and name the lever that fits each one: an integration, a specialised AI run on your premises, a framed generative AI, or nothing at all when a step can simply disappear. What we build runs on your accounts and machines, and is delivered documented. To first estimate what the task costs you, use our calculator.