Every time a company kicks off an AI project, the first thing it does is pull out the mental playbook it already knows: the one it used for the CRM, for the ERP, for any software it has implemented before. It sets a fixed budget, picks a go-live date, trains the team, and expects the system to work from day one and keep working the same way for the rest of the year. It's a reasonable instinct. It's also why so many AI projects end up disappointing the people who paid for them.
We see it constantly. Not because companies do their job badly, but because they're applying the right framework to the wrong tool.
The real difference
A CRM or an ERP is a fixed-rules system. You configure it once, train the team, and the system does the same thing every day under the same conditions. If something in the business changes, someone goes in and adjusts a rule. The system doesn't learn anything on its own, and it isn't expected to.
AI, especially agentic AI, works differently. It learns from every interaction, adjusts with new information, and needs continuous tuning to stay useful as the business changes around it. It isn't install-and-use. It's train, supervise, and maintain.
Treating both systems with the same project logic is where the problem starts.
The misleading go-live
With an ERP, go-live is the finish line. The project ends, the system runs, and the implementation team steps away.
With an AI project, go-live is barely the starting point. The model starts operating on real data, starts showing where it falls short, and that's where the real work begins: adjusting, correcting, training on what the actual business is teaching it. Companies that treat go-live as the project's closing moment are the ones who later wonder why the system "stopped working well" a few months in. It never stopped working. Nobody kept training it.
The budgeting mistake
That same misunderstanding shows up in how these projects get budgeted. An ERP gets quoted as a one-time expense: license, implementation, training, done. When a company applies that same logic to an AI project, it budgets for the initial training and forgets that the system needs ongoing maintenance to keep generating the value it promised.
It isn't about AI being more expensive. It's about the spending taking a different shape: instead of one big spike upfront and silence afterward, it calls for a smaller but steady investment over time. That difference shows up most in operations where information doesn't live in one place. When part of the business still runs on phone calls, loose spreadsheets, or informal coordination between people, an AI system doesn't just have to learn the process once, it has to keep adjusting every time that scattered reality shifts shape. Budgeting it as a one-time expense assumes the business stays still after go-live, and no business stays still.
Companies that budget AI like an ERP end up with no room for that second stage, right when the system needs it most. The result is predictable: the system performs well in the demo, loses accuracy over time, and nobody understood that losing accuracy was an expected part of the process, not a failure.
The TRAXION case
A recent case illustrates this pattern well. When Grupo TRAXION came in to modernize its mobility system, the original plan was built on that closed-project logic: one investment, one rollout, one close. The diagnosis done before touching a single line of code showed something different. What the operation needed wasn't a system that got installed and stayed fixed, but one that kept adjusting as the operation grew.
Part of that operation depended on information scattered across phone calls and manual coordination between teams, exactly the kind of reality a one-time budget can't cover. A system trained once on top of that scattered reality would have gone stale within months. That's why the project was planned differently from the start: not as an installation with a closing date, but as a capability that keeps being tuned as the operation changes.
That's where MIND came from, with 75 improvement opportunities identified and 7 teams working in parallel to address them. The system didn't stop at go-live. It kept expanding from 3 to 24 business units, and that expansion was only possible because the project was budgeted and planned as an ongoing capability, not as an expense that closes and gets forgotten.
Rethinking the starting point
The question worth asking before starting an AI project isn't how much it will cost to install. It's how much it will cost to keep it trained, tuned, and useful six months after go-live. Companies that ask themselves that question from the start are the ones who avoid the disappointment of watching a system that worked well the first month fall behind afterward.
At Creai, this is where we start: we help companies plan their AI projects with the budget and timeline that technology actually requires, not the one inherited from their last ERP.
Similar stories
.png)
