Every week we talk to companies that already "tried AI" and got burned. Not because the technology doesn't work, but because the project was hired wrong from the start. The outcome is almost always the same: a nice pilot in a demo, zero real use six months later.
Here are the seven mistakes we see most often, and what to do instead. If you're about to hire an AI project, run through this list before signing anything.
Mistake 1: Mistaking a no-code connector for "AI integrated into your business"
Many proposals sell automation built on no-code tools (Zapier, Make and similar) wired to a language model, and call it "AI integrated into your business". That works for simple cases, but it breaks the moment your process has an exception, a specific business rule, or real volume.
What to do instead: ask explicitly whether the solution runs as code against your real system (ERP, CRM, accounting) or is a layer of connectors on top of it. Both have their place, but they don't cost the same or offer the same guarantees. We break down this difference in what integrating AI for real actually means.
Mistake 2: Starting with the AI without looking at the data
This is the most expensive mistake, and the most common. A company hires an AI project on top of an ERP full of duplicate clients, a CRM with empty fields, and spreadsheets nobody updates. AI doesn't fix that, it amplifies it. A model reasoning over dirty data gives confident, wrong answers, which is worse than no answer at all.
What to do instead: insist that the first phase be a data diagnosis, not a model demo. In a real project with a pharmacy group, unifying and cleaning up the data from several locations before applying AI was what later made it possible to reconstruct a sale the ERP was calculating wrong, and validate it down to the cent. Without that groundwork, the AI would have had nothing reliable to reason over. See the case at data platform and AI.
Mistake 3: Only asking the AI to "suggest", never to act
An assistant that just drafts a report or suggests a reply is easy to sell, but it leaves all the heavy lifting (copying, pasting, reviewing, correcting) in someone's hands. The real time savings almost always come from AI acting inside the system: creating the order, posting the invoice, booking the appointment.
What to do instead: ask exactly what action the AI executes end to end, and in which system. In an industrial distribution project, the AI doesn't just quote over WhatsApp: it validates the client, applies the correct price tier, checks stock, and creates the opportunity and the order inside the ERP, for real. See AI sales rep inside the ERP.
Mistake 4: Accepting a black box with no exit plan
If the vendor is the only one who understands how the system works, and everything lives inside their closed platform, you're buying dependency, not a solution. The day you want to switch providers, or the vendor disappears, you're left with nothing.
What to do instead: ask where the code lives, who owns the data, and what happens if you end the relationship. A properly built custom solution should be auditable, documented and, if needed, maintainable by another team.
Mistake 5: Measuring the project by hours of AI, not by business outcome
"We integrated a language model" is not an outcome, it's a means. The outcome is: invoices post themselves, quotes go out in minutes instead of hours, an error nobody was catching gets caught. If the vendor only talks about technology and never about the business metric that changes, that's a red flag.
What to do instead: define, before you start, which specific process is going to change and how you'll measure it. In an accounting case, the metric wasn't "we have AI", it was: invoices get matched to the supplier, taxes get reconciled line by line, and the invoice posts itself in Holded, verifying the total before creating the entry. See invoices that book themselves.
Mistake 6: Ignoring where your sensitive data lives
If your business handles sensitive information (client records, internal manuals, regulated data) and the vendor's plan is to send everything to a generic cloud model with no further thought, there's an open question about data sovereignty and compliance.
What to do instead: ask which model is used, where it's hosted, and what options exist to keep the data inside your country or infrastructure if your sector requires it. That's exactly what we explored in a sovereign AI project for the food industry, with models hosted in Spain.
Mistake 7: Never asking what happens when the AI gets it wrong
No system is right 100% of the time. The question that really separates a serious project from an improvised one is: what happens when it fails? Does it stop and flag it, or does it plow ahead as if nothing happened?
What to do instead: require built-in verification. In the self-posting invoices case, if the total doesn't match, the system deletes the invoice instead of posting it incorrectly. That discipline, verifying before accepting a result as good, is the difference between trusting the AI and having to double-check everything it does.
Checklist before you sign
| Question | Why it matters |
|---|---|
| Does it run as code on my real system, or is it a layer of connectors? | Determines cost, robustness and long-term maintenance |
| Does phase one include cleaning up my data? | Without clean data, AI amplifies the mess |
| What action does the AI execute end to end, in which system? | That's where the real time savings live |
| Who owns the code and the data? | Avoids being locked into one vendor forever |
| Which business metric will change, and how is it measured? | Without this, you won't know if the project worked |
| Where is the model hosted, and my data with it? | Relevant if you handle sensitive or regulated information |
| What happens when the AI gets it wrong? | A system with no verification can't be trusted |
The underlying principle
None of these mistakes are exclusive to a bad vendor, they often happen because the company hiring doesn't know what to ask. The principle we follow at AutoBoost is simple: AI where it adds value, code where it's needed, and always with the data sorted out first. If you also want to know whether your company is ready to take this step, read the signs you're ready.
Weighing an AI project and want a second opinion? Tell us about your case and we'll tell you, with no strings attached, what we'd ask in your shoes.
