Where to start your first AI project in an SMB: a 3-step path

Where do you start with AI in your SMB? Not with the tool. BCG's 10/20/70 rule splits the value of a project like this: 10% in the algorithms (the AI models that come already built), 20% in the technology and 70% in the people and in how the work changes. Read as an order of work, it leaves three steps: choose the problem, prepare the minimum you need and design how your team is going to adopt it. Gartner projects that through 2026, 60% of projects without AI-ready data will be abandoned.

Why does starting with the tool go wrong?

Because it reverses the order. The typical conversation with an SMB owner starts with "we want to implement AI", and the natural reaction is to go looking for a product: request a quote, compare options, approve a license. Three months later the tool is open in two tabs and nobody uses it.

What follows explains why that path almost always ends the same way, and how the same BCG rule offers a more reliable route.

The rule indicates where the value of an AI project is usually created; it is not a formula for allocating budget or hours.

What does the 10/20/70 rule actually say?

BCG published it in February 2026 and it became a reference quickly: 10% of the value of an AI project comes from the algorithms, 20% from the technology and the data, and 70% from the people, the processes and organizational change.

It is worth pausing on the first word, because it is the one that creates the most noise. The algorithms, in this split, are the AI models that do the task, the ones that already come built into the tools you sign up for: you do not train them or program them, you choose them and put them to work. They weigh 10% precisely for that reason, because they are the part that arrives already solved. What does not arrive solved is the rest: where your information lives and who can consult it, which is the 20% of technology and data, and how the work changes for the people who are going to use the AI, which is the 70%.

It is easy to read it backwards and conclude that you have to put 70% of the budget into training. That is not it. The rule does not say where to spend the money, it says where the value lives. And therefore, where to look when something is not working.

One figure reinforces that reading: only about 5% of organizations capture substantial financial gains from AI, and that group plans to train more than half of its workforce compared with 20% among the laggards. The distance between the two is not explained by technology budget alone.

Step 1: what problem are you going to tackle?

The pain first, the tool second. The pain that works for getting started is the one someone on your team can describe without thinking about it, with the process and the person who suffers it, and that is already costing you something concrete: hours that disappear, work that gets redone, clients who cool off because an answer arrived late. In a services company that cost shows up on many fronts. The usual candidates are meetings that produce no follow-up, sales proposals that take more time than they need to, knowledge that depends on two or three people and client follow-up that falls apart for lack of shared criteria. That list is not closed. If your worst pain is not on it, it is still a good candidate as long as the cost is real and someone can name it.

The first use case should meet three conditions: that the problem happens often, that it can be measured and that you have at least some idea of how you will know in 90 days whether it improved.

Before moving on, write in one sentence which pain you are going to tackle and why that one before the others. If that sentence does not come out, there is the pending work, and the tool can wait.

Step 2: what is the minimum you need in order to test?

With the problem clear, the technical part comes in, and here the rule is not to overdo it. Technology weighs 20%, which means two things at the same time: it is worth having the information accessible and a tool that does the job, and it is not worth setting up three months of data cleanup or buying the most sophisticated package before validating the first use case.

Gartner projects that, through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data. The useful reading is not to spend a year preparing data before touching anything. It is to verify that the key information for the problem you chose exists somewhere you can consult, even if it is imperfect.

For a first project it is usually enough to have one approved tool that covers the use case, a shared place where the information lives and a narrow workflow. When you choose, review privacy, data retention, permissions, integration and total cost with the provider you already use.

If step 2 is taking you more time than step 1, it almost always means the use case was not well defined.

Step 3: how is your team going to adopt it?

This is where the game is decided, and it is the step almost nobody plans in the detail it deserves. Designing adoption means, at a minimum, naming someone to carry the project day to day and someone in leadership to defend it when it competes with other priorities. Defining who is going to use the AI and how often. Deciding how the team will be trained, not only in how to ask the tool for things but in when it makes sense to use it and when it does not. Setting at least one indicator for days 30, 60 and 90. And reviewing results on dates set in advance, to decide whether it gets expanded, paused or redirected.

McKinsey reports that 48% of surveyed employees would use AI more with formal training, and 45% if it were built into their workflow. Training and integrating are not symbolic gestures: they are the two material conditions for people to use it.

Give this step most of your attention. The 70% describes where the value is created, not a formula for dividing up hours or budget.

Starting with the tool reverses the order and increases the risk of ending with a solution that has no real use.

What is the most common mistake?

Doing the three steps in reverse order. First the tool gets bought, then it turns out nobody defined what it was wanted for, and in the end it becomes clear that nobody is using it. Three months later the license gets canceled and the conclusion left behind is that AI is not for this company yet.

That is almost never true. What happened is that it started with the wrong step.

Which step are you standing on?

VegasiO's Discovery web takes 5 minutes, it is free and at the end it tells you which of the three you should focus on first, with a concrete recommendation for your case.