I’ve been hands-on involved in two AI complex projects already.

A couple of years ago, I was asked to help a startup that was burning $200,000 a year into an AI product. After about 3 years, the result was still a weak $2,000 MRR. Quarterly reviews were painful. People were being fired and replaced hoping for a change.

3 years after launch, the team was still trying to figure out who their customer was.

Now in the last couple of months, I'be been collaborating with a new one. I will be building the GTM strategy and team for them. And the story couldn’t be more different. 

With just a €50,000 investment, they’ve already unlocked over €1.25 million per year in salary savings for a heavy hardware manufacturing entreprise. Freed up 26 out of 33 full time engineers, and locked €120,000 ARR for the startup. 

All in just 4 months.

Here’s the fundamental reasons and differences I see between them.

1. Working side by side with the customer

In the failed project, let's call it project X, developers built in a vacuum. They didn’t really understand the workflows they were trying to “fix”. They were trying to solve a problem they never managed themselves.

In the current one, let's call it Y, it was the opposite. The team's mathematicians, developers, and the client’s own support engineers sat together from day one. They rolled up their sleeves and figured out together every step. Hypothesizing, testing, and iterating with the people who knew the pain points best. 

2. The “why” is not in the data, but in the brains of humans

In project X, as in many others I see out there, they only trained the AI model on existing data. Right or wrong.  Yes or No. 

That gave them the “what”, the problem-solution relationship, but never the “why.” Accuracy never rose above average, and annotators and human intervention was often needed to correct mistakes, because new issues without previous data could not be solved by the AI without the fundamental why.

The startup Y soon realised this, and built a process where top engineers dumped their wisdom into the system. The WHY that lives in their heads. 

This combined with some mathematical magic from the AI team, was the quality breakthrough: accuracy shot from 60–70% to over 95%. Suddenly, the model was outperforming the best engineers, at a tiny fraction of time.

3. Building trust step by step

AI adoption is not only a matter of performance. It's actually much more a matter of human trust.
Most projects try to go too fast, dreaming of replacing people before the AI is even good enough.
But if humans don't trust the AI, they never actually adopt it, nor make it better. What they get in return is 3 years gone, 600k burnt and 2k MRR.

In project Y, they started small. First, a simple chat for engineers to ask questions. To prove it could match a human expert. 

Once that quality arrived, they integrated the AI into their workflows: ERP, ticketing, spare parts ordering, but only as a co-pilot. Engineers could approve or reject every suggestion from the AI, and each rejection made the model smarter.

Two weeks ago, when the AI consistently showed that more than 90% of actions were approved, did they set it free in autonomous mode. By then, trust was already earned, not imposed.

4. Results that still shock me

Five months in, here’s what we’re seeing:

  • -99.6% time to first response (from 270 minutes to 1)
  • -66% time to resolution.
  • >95% accuracy
  • 26 engineers freed up to focus on scaling and innovation. 7 dedicated to just monitor and maintain the AI system. That's €1.3M of yearly salary savings, which now the company will use to expand into new regions.

The speed, the accuracy, the freed-up human talent. It’s shocking, especially for me.

What I’ve learned

I believe AI doesn’t fail because of the technology. It fails when it’s built in isolation, without the people who know the problem best, and without a stepped path to trust.

  • 33% of its success relies on experts putting in the right technology.
  • 33% on actually dumping the brains of the top performers in the tasks at hand, into the system.
  • and 33% on integrating AI step by step into their workflows. One step at a time, and don't move forward until quality is as good as or better than humans'.

It succeeds when it’s hands-on, collaborative, and gradual.

And when it succeeds, the impact isn’t just numbers on a dashboard. It’s a lot of people with more time to think, to innovate, to expand the business. That’s the real magic I’ve just witnessed.

👉 Here is the full AI-powered B2B technical support workflow infographic for those who want to dive into the details of project Y.