Skip to content
shiftlabs

Guide

In an SME, AI starts with the work, not the tool: measure where one team’s time goes, then test a reversible first use case under rules written before launch. A measurement on a date fixed in advance then tells you whether to extend, adjust or stop.

  • 9 min read

Published

Contents

The short answer: start with the work, not the tool

Starting with an assistant subscription means asking every employee to work out alone what it is for. Starting with the work means taking one specific task: who does it, how often, how long it takes, which data it uses. The tool is chosen afterwards, for that task.

The first barrier companies report is a question of method. In 2025, among French enterprises with 10 or more employees that do not use AI, 71% say they see no use for it and 54% cite a lack of expertise (Insee). Across the EU, among enterprises that considered AI but did not adopt it, lack of relevant expertise comes first, at 71% (Eurostat). Both barriers call for the same answer: tie AI to a real, measured task rather than to a list of features.

The five steps in this guide follow that order and the tools on this site: the express diagnostic to place your organisation, the simulator to put a figure on a first use case, then, if you want measurement done in the field, the Pilot.

Step 1: place your organisation on four tiers (express diagnostic)

Before choosing a use case, place your organisation. The express diagnostic asks twelve questions across six dimensions (strategy, data, skills, processes, tools, risks); each answer maps to a tier. It takes three minutes, gives an indicative result, and nothing is sent.

The tier is not a grade: it tells you what a sensible first step looks like. This guide is written for organisations at tiers 1 and 2.

Step 1: place your organisation on four tiers (express diagnostic)
TierWhat the scale saysSensible first step
1. Ad hoc useAI remains occasional, individual and ungoverned.List the uses that already exist, write a one-page usage charter, then pick one team to observe.
2. Fragmented initiativesTools help, but remain local and unshared.Pick one process, measure it, and test a first use case there with a named owner.
3. Integrated AIAI supports core processes, with human review.Tie each use case in production to an indicator tracked by the executive committee.
4. AI-native operating modelAI executes at scale; humans validate and arbitrate.Outside the scope of this guide.
Scroll the table sideways.

Step 2: look at where the time goes, one team, two weeks

Pick one team, not the whole company: accounts payable, sales administration, customer service. For two weeks, record what it actually does, not what the job descriptions say. France Num’s guide to generative AI for small businesses starts in the same place: review your activities and identify the repetitive, time-consuming tasks.

For each task, record:

  • what triggers it (an email, an order, a deadline) and what it produces;
  • its volume: how many times a week;
  • the time per occurrence, timed on a few real cases rather than estimated from memory;
  • the data it uses and the system it sits in (ERP, email, spreadsheet);
  • what happens when there is an error, and who catches it.

Step 3: choose a first use case that is reversible and measurable

From that record, keep a single task. The first use case does not have to be the most profitable: it has to be something you can stop without damage and measure without argument.

  • Reversible: if the tool stops tomorrow, the team picks the task up as before, with no data lost and no customer affected.
  • Measurable: time per occurrence and volume were recorded in step 2, so they can be compared afterwards.
  • Frequent: a daily task produces enough occurrences to measure quickly.
  • Reviewed by a person: the output is checked before it reaches a customer or a system.
  • Low-stakes: according to the CNIL, the French data protection authority, helping employees draft content carries less risk than a tool that helps make decisions about customers or candidates. Start on the less exposed side.
  • Owned: a volunteer in the team is accountable for it and has time for it; France Num calls this person an “AI lead”.

Step 4: set the rules before rollout: data, employees, transparency

These rules fit on one page. Written before launch, they spare you negotiating them along the way.

Step 4: set the rules before rollout: data, employees, transparency
TopicRule to write downWhat the sources say
Data enteredThe approved tool, and the list of data that must not be entered (personal data, trade secrets).The CNIL recommends a charter listing permitted and prohibited uses, and never sharing confidential information with a consumer service.
Personal dataWho signs off the use before launch: the data protection officer, if you have one.The CNIL recommends involving the DPO and, where appropriate, carrying out a data protection impact assessment (DPIA).
UsersShort training before use: how the tool works, its limits, how to check outputs.According to the CNIL, the organisation is generally liable for misuse by its staff; it recommends training users and encouraging them to check outputs.
Employees and their representativesWhat is observed and why (describing tasks, not rating people), and what the freed time will be used for. From 50 employees in France, the works council question.In French companies with at least 50 employees, the works council (CSE) must be informed and consulted on “the introduction of new technologies” (French Labour Code, articles L2312-1 and L2312-8).
TransparencyWhen a customer or candidate is dealing with an AI system, they know it.According to the CNIL, the EU AI Act requires people to be informed that they are interacting with an AI system, and published texts generated this way to be disclosed.
Scroll the table sideways.

Step 5: measure, then decide to extend or stop

Before launch, write down three things: the indicators, the measurement dates and the threshold below which you stop. A decision taken after the fact, with no written threshold, never gets settled.

  • Time per occurrence, compared with the step 2 baseline.
  • Volume handled, and the share of outputs corrected or redone by hand.
  • Actual use: how many people in the team use it each week.
  • The team’s view, gathered on fixed dates; the CNIL recommends regularly collecting users’ feedback on the system’s usefulness and limits.
  • Actual costs: subscription, usage charges, review time.

What not to do first

What not to do first
MistakeWhyInstead
Buying licences for everyoneUse stays individual and invisible: nobody can say what has changed.One tool, one team, one measured task.
Starting with the most ambitious caseIt spans several teams and systems: measurement comes late and stopping is expensive.A reversible, frequent use case, reviewed by a person.
Launching without a baselineWith no time recorded before, no gain can be proved after.Two weeks of recording in the team concerned.
Letting everyone use the consumer tool of their choiceCompany data gets entered with no rules; the CNIL advises against sharing confidential information there.A tool chosen for the task and a one-page charter.
Saying nothing about what the freed time is forThe team fills the silence with its own assumptions, and the freed time dissolves.Decide it before launch, and say so.
Handing AI a decision about people from the outsetIt is the most exposed use; the CNIL expects a higher standard for it.Help with drafting or sorting, reviewed by a person.
Scroll the table sideways.

Frequently asked questions

Where should an SME start with AI?

With a task, not a tool. Place your organisation with a maturity diagnostic, record for two weeks where one team’s time goes, choose a repetitive task that is reversible and measurable, write the usage rules, then measure before deciding to extend.

Do you need a data team to get started?

Not for a first use case reviewed by a person: you need a well-described task, someone accountable for it and usage rules. Lack of expertise is still the barrier most often cited in the EU by enterprises that considered AI but did not adopt it (Eurostat, 2025); short outside support can fill that gap, provided the method stays with you.

How much does a first AI project cost an SME?

There is no reliable reference price: it depends on the tool, the volume and the internal time involved. The simulator gives an order of magnitude from your own assumptions. For measurement in the field, the SHIFTLABS Pilot starts at €12k excl. VAT, fixed fee, over two weeks, with no commitment to continue.

Do you have to consult employee representatives before introducing an AI tool?

In France, companies with at least 50 employees must inform and consult the works council (CSE) on the introduction of new technologies (Labour Code, article L2312-8). Whether your tool falls within that rule is a question for your legal adviser before rollout, not after. Elsewhere, check your local rules.

Which data should not be entered into a generative AI tool?

The CNIL recommends never sharing confidential information, such as personal data or company data covered by trade secrecy, with a consumer service, and writing the permitted and prohibited uses into a charter. With a business tool, the rules also depend on the contract and how the tool is deployed.

How long before you know whether a first use case works?

Decide that before you start. A daily task quickly produces enough occurrences to compare with the baseline; a monthly task needs several cycles. Without a written measurement date, the decision keeps slipping.

Sources

Method: external figures checked against the source cited; SHIFTLABS figures are presented as estimates or examples.

Further reading

Keep reading

And in your organisation?

In two weeks, AI agents interview your teams; the founder, a data and AI consultant, and depending on the scope a practitioner of the business function, review and sign off every conclusion. You leave with costed use cases and a roadmap.

Request a 30-minute call

All resources