AI roadmap for SMBs: the first 90 days

In many SMBs, the question is no longer whether to look at AI. The real difficulty starts right after: what to do now, in what order, and without scattering in every direction?

This is where many companies lose time. They read, test, open accounts, watch demos, spot interesting ideas, but with no clear line. The result: a few trials, few decisions, and not much that holds in daily practice.

For an SMB, the first 90 days are not about changing everything. They are about laying a clean foundation: understanding where to start, testing few things but the right ones, avoiding the obvious mistakes, and deciding what comes next with a bit of perspective.

What an SMB should have after 90 days

After three months, an SMB does not need to have rolled out AI everywhere.

It does, however, gain a lot from having secured four things:

• a sharper sense of its priorities
• one or two use cases tested for real
• minimal rules on what is allowed and what is not
• enough of a base to decide what comes next

That is already a lot. And it is far more useful than a string of scattered trials or a tool bought too fast.

Days 1 to 30: clarify the need and choose a starting point

The first month is there to avoid the most common mistake: starting with a tool instead of starting with a problem.

1. Start from what wastes time

Begin by looking at the tasks that take too long, that come back often, or that move slowly.

In many SMBs, the first topics are often the same:

• summarizing meetings or documents
• drafting a first version of an email or a note
• finding and organizing information
• answering repetitive requests
• writing down internal procedures

So the starting point is not: which tool should we test? The starting point is: what is worth improving?

2. Choose a narrow scope

The first month should not be about launching the topic across the whole company. It should be about choosing a simple testing ground.

A good starting point often looks like this:

• one targeted team or function
• one specific use case
• a short test
• a few simple criteria to judge the result

In other words, a small, well-framed trial beats a launch that is too broad from the start.

3. Look at the data question right away

Even before testing a tool, you need to know what can go into it and what must never go into it.

This is often where companies move too fast. They test first, then ask afterward which data went through. It is better to do the reverse.

From the first month, you need to separate:

• content with no particular risk
• content to anonymize
• data to exclude entirely

This initial sorting prevents many mistakes down the line.

4. Choose a tool that is simple to try

The first tool does not need to be the most complete on the market. Above all, it should fit the intended use case, make sense to the people who will use it, and meet minimal security requirements.

The right instinct is not to look for "the best AI tool". The right instinct is to look for a tool that fits the chosen need.

5. Define in advance what will count as a good result

Before launching a test, you need to know what you are going to observe.

For example:

• are we saving time
• is the output good enough
• is the tool easy to use
• does it need too much rereading
• do the people involved want to keep going?

Without this, you quickly end up with vague impressions like "that was interesting", with no idea of what to decide next.

Days 31 to 60: test, train a little, set a few rules

The second month is there to move from talk to a real trial.

1. Launch a short test

A 30-day test is often enough to learn a lot. No need for a big setup. What matters most is a readable frame:

• one use case
• a few people
• one tool
• a duration
• points to observe

The goal is not to prove that AI will change everything. The goal is to see whether one specific use case really holds in daily practice.

2. Give the people testing some guideposts

Even for a simple test, you need to keep everyone from discovering the topic alone in their corner.

This does not mean launching a big training program. But at a minimum, you have to explain:

• what the tool is for
• in which cases to use it
• what you should not hand it
• what needs to be reread
• who to ask a question

This minimum changes a lot of things. It prevents guesswork, shaky use, and bad habits picked up too early.

3. Set minimal rules

After a month of testing, it is already useful to have a few simple rules.

For example:

• which tools are allowed
• which data is excluded
• which content must be reread
• who decides in case of doubt
• in which cases the decision stays entirely human

No need for a big document at this stage. A few clear rules are often enough to secure the first use cases.

4. See what really gets in the way

The second month is not about convincing yourself that "AI works". It is about seeing what gets stuck in real life.

For example:

• a tool that is too complicated
• answers that are too inconsistent
• a team that does not quite know how to use it
• a poorly chosen use case
• rules that are still too vague

This feedback is useful. It keeps you from pursuing a test just because it looked promising at the start.

Days 61 to 90: sort things out and decide what comes next

The third month is there to get out of the fog. This is the moment when the company should be able to say: we keep it, we adjust it, or we stop.

1. Make a simple review

At the end of the 90 days, you should be able to answer a few very practical questions:

• what genuinely helped
• what took too much time for too little result
• what was easy to adopt
• what demanded too much oversight
• what needs to be framed better

This review does not need to be long. Above all, it needs to be honest.

2. Do not think in all or nothing

After 90 days, there are not just two options: "we roll it out everywhere" or "we drop it".

In practice, an SMB can:

• keep a test running because it seems useful
• expand a use case that works well
• stop a use case that does not bring enough
• switch tools
• keep the topic on a narrow scope for a while longer

It is often this ability to sort things out that makes the difference.

3. Prepare a realistic next step

Once the first cycle is over, the SMB can start preparing what comes next:

• the next use cases to test
• the people to train further
• the rules to refine
• the tools to confirm or replace
• an acceptable rollout pace

The idea is not to rush. The idea is to keep what really holds in daily practice.

What a good roadmap looks like after 90 days

A good roadmap after 90 days is not a grand theoretical plan. It is often something very simple:

• a clearly stated need
• one or two use cases tried for real
• a first selection of tools
• minimal rules already in place
• people who know a little better what to do
• a clear decision on what comes next

In other words, the company has not "finished with AI". But it has stopped circling around it.

The most common mistakes during the first 90 days

Trying to go too fast

The first mistake is wanting to show too quickly that the company "is making progress on AI". You then pile up tests, with no clear order.

Starting with the tools

A tool should not be the starting point. As long as the need stays vague, the choice of tool holds little value.

Forgetting the data question

Many companies look at this point too late. Yet it is part of the initial framing.

Confusing interest with real use

A team can find the topic interesting without that turning into lasting use.

Not giving enough guideposts

With no simple guidelines, no examples, and no explicit human validation, mistakes quickly become hard to spot.

What to take away

For an SMB, the first 90 days with AI should not be thought of as a race. They should serve to lay the foundations.

In the right order:

• clarify the need
• choose a small scope
• protect sensitive data
• test one or two use cases
• give teams a few guideposts
• set simple rules
• decide what comes next based on what genuinely worked

It is this progression that lets you make progress without chasing a trend, without needless noise, and without putting the company at risk.

FAQ

Do you have to choose a tool in the first month?

Not necessarily. The first month mostly serves to clarify the need, the scope, the data involved, and the success criteria. The choice of tool comes after.

How many use cases should you test in 90 days?

In most SMBs, one or two well-chosen use cases are enough. Beyond that, attention scatters.

When should you train the teams?

As soon as a real test begins. Even light framing beats a use discovered alone, with no clear rule.

Do you need an AI charter after 90 days?

Not necessarily a very complete version. Simple rules, though, are already useful: approved tools, excluded data, human rereading, a point of contact in case of doubt.

Sources and references

To write this article, we drew on reference resources around AI adoption in companies, choosing solutions, protecting data, and upskilling teams:

• CNIL, Using generative AI in small and medium businesses: practical recommendations on use cases, the precautions to take, and the general framework. Read
• France Num, Choosing among artificial intelligence solutions: a practical guide for starting from the need, comparing solutions, and looking at costs, data, and skills. Read
• European Commission, AI Literacy: Questions & Answers: details on the upskilling requirement tied to the AI Act. Read
• Bpifrance Le Big média, How to integrate AI into your company? 8 key steps to succeed: guideposts on the gradual integration of AI in companies. Read