Introduction: why so many AI projects disappoint
Why do so many AI projects stall at the testing stage? As early as 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after the proof of concept, because of poor data quality, poorly managed risks, escalating costs or unclear business value. These failures are not inevitable: they stem from recurring mistakes that can be avoided if you know them. This article complements our guide AI strategy for SMEs: key steps and priorities, which lays out the order of play.
These mistakes are rarely technical. They are strategic and organizational, and often tied to a mistaken idea of what AI can deliver. Avoiding them is what allows you to move from a test to a system that runs day to day, like those presented on the diagonall AI systems page.
Here are the 5 most common mistakes and, above all, how to avoid them. Whether you are an SME owner, a digital manager or a project manager, they can save you months and a lot of money.
Mistake #1: trying to automate everything at once
The "big bang" syndrome
This is the most common and most expensive mistake. Excited by the promises of AI, some leaders want to transform all their processes in one go: a single system that would handle accounting, customer service, production and HR at the same time. This approach very often fails.
Typical scenario: an industrial SME launches a "global AI" project meant to automate about ten processes. Months later, not a single module is in production, the team is exhausted and the project is abandoned. Everything has to be started over, with a more modest approach.
Why this approach fails
A large-scale project involves dozens of connections, hundreds of edge cases and a massive change for the teams. Without prior experience, no one yet knows what works in your context. Risks add up: technical, human, financial. And if the project fails, you lose the budget, the time and your teams' trust in AI.
The solution: start with a quick win
Start with a single process that is simple and repetitive: a task that takes time, adds little human value and has clear rules. Examples by department:
- Accounting: automatic filing of supplier invoices, bank reconciliation.
- Customer service: sorting and routing incoming emails, answering frequently asked questions.
- HR: scheduling interviews, answering internal questions.
- Sales: enriching prospect records, prioritizing leads.
- Production: stock tracking, restocking alerts.
This first project should be deliverable in 4 to 8 weeks. Its success builds buy-in and provides lessons for what comes next. One successful first project is worth more than ten ambitious projects that get abandoned. To choose this first project, see Prioritizing your AI use cases: the matrix.
A word of caution for HR: a tool that sorts or scores job applications is considered high-risk under the AI Act, with obligations that apply from December 2, 2027. It is not a good first project.
Mistake #2: underestimating data quality
AI is only as good as your data
The expression "garbage in, garbage out" has never been truer. A model, however powerful, cannot make up for poor-quality data. And many SMEs discover the real state of their data only when they launch the project.
It is often the most painful moment of truth: the salesperson who promised wonders had not seen your spreadsheets full of empty fields and duplicates.
The most common problems
- Incomplete data: empty fields, partial histories.
- Inconsistent data: the same customer written five different ways, different date formats.
- Scattered data: spread across spreadsheets, CRM, ERP, emails and paper notes.
- Outdated data: contacts who have left, obsolete prices, wrong stock levels.
- Unstructured data: scanned PDFs, emails, photos of documents.
The solution: audit your data first
Before any project, carry out a data audit on the relevant scope. It answers three questions: where is my data? What is its real quality? How much effort will it take to clean it?
Include a dedicated line in the budget for data preparation. It is not a superfluous expense: this clean data will serve your future projects, as well as your analyses and day-to-day decisions. See also AI audit: assessing your company's maturity.
Mistake #3: ignoring change management
The human factor, the great forgotten
The best system in the world is useless if it isn't used. Technically successful projects fail because employees reject the tool, work around it or enter bad data into it.
Typical scenario: a sales assistance tool is rolled out in a trading SME. Six months later, only a small share of the salespeople use it. The others have found ways to do without it. The project is judged a failure, even though the technology works.
The sources of resistance
Resistance is not irrational. It is based on legitimate fears:
- Fear of losing one's job: "AI is going to replace me."
- Fear of losing one's expertise: "my know-how will no longer be needed."
- Fear of the unknown: "I don't understand how it works."
- Fear of making mistakes: "what if I get it wrong with this new tool?"
- Fear of surveillance: "AI is going to measure my performance."
The solution: get people involved from day one
Change management starts before the project:
- Explain early and clearly the "why" before the "how."
- Involve users in choosing which processes to automate.
- Train gradually, through ongoing support rather than a marathon training session.
- Showcase successes by talking about time and comfort gained.
- Rely on ambassadors: two or three convinced people who bring the others along.
Training is no longer just good practice: since July 27, 2026, Article 4 of the AI Act requires companies to take measures to support the AI literacy of their teams.
Mistake #4: choosing the wrong technology
The cutting-edge trap
The ecosystem is evolving very fast. ChatGPT, Claude, Gemini, Mistral, Llama: models and announcements keep coming. It is tempting to want the latest novelty. It is often a mistake.
The newest technology is not necessarily the best suited to your needs, and it sometimes lacks the maturity required for production use: bugs, interface changes, thin documentation.
How to choose well
The choice must start from your business need, not from the buzz of an announcement. Five criteria:
- Maturity: has the solution proven itself in production in similar contexts?
- Integration: does it connect easily to your tools (ERP, CRM, accounting)?
- Scalability: will it be able to keep up with your needs in two or three years?
- Support: clear documentation, responsive support, in French if possible?
- Total cost: beyond the license, what are the integration, training and maintenance costs?
Our recommendation for getting started
For an SME just starting out, it often makes sense to begin with no-code or low-code tools to validate the idea before investing in custom development. Platforms like Make or Zapier, connected to a language model, make it possible to test a use case in a few days. Custom development is then justified once the need is confirmed and volumes increase.
Mistake #5: not measuring return on investment
AI for AI's sake
Too many projects are launched without a measurable objective. "Our competitors have it" is not a strategy. Without indicators set in advance, you will never know whether the investment pays off, and you will not be able to justify the following projects.
The indicators that matter
For each project, define before the start:
- Time saved: how many hours per week, at what hourly cost?
- Errors avoided: what rate before, what target after, what cost per error?
- Revenue: more sales, better loyalty, higher average basket?
- Satisfaction: of employees and customers, processing times.
- Capacity: handling more volume without hiring.
Calculating the real ROI
The calculation must include all costs, not just the provider's invoice:
- development and integration (provider and in-house);
- licenses and subscriptions;
- team training;
- maintenance and upgrades;
- internal time spent (project manager, key users, IT).
A positive ROI in the first year is possible on well-scoped projects, but it is not the only criterion. Some projects create value over the long term: quality of working life, skills development, the ability to serve more customers. The full method is in AI ROI: how to calculate it.
How to avoid these mistakes: the diagonall method
Our method has four phases.
Phase 1: diagnostic (1 to 2 weeks)
Review of your processes, your data and your digital maturity. Identification of quick wins with high impact and low complexity. ROI estimate for each use case.
Phase 2: prototype (2 to 4 weeks)
Rapid development of a proof of concept on a limited but representative scope. Validation by users in real conditions. See Scoping an AI POC: data, risks, schedule.
Phase 3: deployment (4 to 8 weeks)
Gradual rollout with a pilot group, then expansion. Team training. Adjustments based on feedback from the field.
Phase 4: continuous improvement
Measurement of results against the initial objectives. Improvements based on actual use. Extension to other use cases once the first one is stable.
Conclusion
Trying to automate everything at once, neglecting data, ignoring the human factor, choosing the wrong tool, not measuring: these five mistakes explain a large share of failures. The good news is that they are all avoidable.
AI is neither a magic wand nor a threat to your company. It is a powerful tool that, when used well, can transform an SME. But its effectiveness depends entirely on how it is used.
Before launching a project, ask yourself four questions: what concrete problem do you want to solve? Is your data ready? Are your teams on board? Have you set measurable objectives? To choose the right partner, see Choosing your AI provider: the key criteria.
Frequently asked questions
What is the most common mistake in an AI project?
Trying to do too much at once. A scope that is too broad multiplies risks and delays the first visible result.
How long should a first AI project last?
Generally 4 to 8 weeks for a well-chosen first use case. Beyond that, the scope is probably too broad.
Do you need perfect data to get started?
No, but you need to know what state it is in. A quick audit on the project scope is enough to tell whether cleaning is needed before you begin.
How do you know if an AI project is a success?
By comparing the indicators set before the start (time, errors, satisfaction) with the measurements taken after deployment.



