AI Projects I Turn Down

Every consultant says they’re honest

Few prove it. Here’s my proof: a list of projects I’ve turned down.

Not because I don’t want the work. Because the projects shouldn’t have been AI projects. Building them would have wasted the client’s money and damaged my reputation when they failed.

Saying no is the most honest thing a consultant can do. If a vendor has never turned down a project, they don’t have standards. They have a quota.

Project 1: The chatbot nobody asked for

A company wanted an AI chatbot for their website. “Our competitor has one.” I asked about their support metrics: ticket volume, response time, NPS. All healthy. Support tickets were manageable. Response times were under 2 hours. Customer satisfaction was high.

The chatbot would have cost €60K+ to build and would have solved a problem that didn’t exist. It might have even degraded the customer experience - replacing a fast human response with a slower AI one.

What I recommended: Improve self-service documentation. Add a search function to their help center. Cost: €5K. Impact: same as a chatbot, without the AI complexity.

Project 2: Custom model for 500 records

A company wanted to train a custom model to classify documents. They had 500 labeled records. That’s not enough for training a custom model - you need thousands, ideally tens of thousands.

The vendor who referred them had quoted €200K for a custom model. It would have been less accurate than a simple API call with good prompting, because 500 records can’t teach a model enough to generalize.

What I recommended: Use an API with carefully engineered prompts. Cost: €2K/month. Accuracy: higher than the custom model would have been. Timeline: 2 weeks vs 6 months.

Project 3: AI for a process that wasn’t defined

A company wanted AI to “optimize their operations.” I asked: “What operations? What’s the process?” They didn’t have a documented process. Different teams did things differently. There was no standard workflow to optimize.

AI can’t optimize what you haven’t defined. It would have learned from inconsistent, ad-hoc processes and produced inconsistent, ad-hoc recommendations.

What I recommended: Map the process first. Document the workflow. Standardize across teams. Then look at where AI could help. Cost: €10K for process mapping. The AI project was premature.

Project 4: The FOMO project

A board mandated an “AI initiative.” No problem, no metrics, no data budget. Just “we need to be doing AI.” The CEO was under pressure to show progress.

There was no business case. No specific problem to solve. No data to work with. No success criteria. The project would have been a €150K press release.

What I said: “I won’t take this on as an AI project. I’ll do an audit that tells your board what’s possible, what it costs, and what we’d need. If the audit identifies a real opportunity, we build. If not, you’ve spent a fraction of the budget and you have a data-driven answer for your board.”

They didn’t want an audit. They wanted a story. I’m not in the story business.

Project 5: The vendor rescue

A company had a failed AI project - wrong architecture, no data pipeline, no monitoring, no documentation. The vendor had left. The system was a black box that nobody understood and that didn’t work.

They wanted me to “fix it.” I audited it. The architecture was fundamentally wrong - it couldn’t be patched. Fixing it would cost more than starting over, and the result would be a compromised system built on a bad foundation.

What I said: “Start over. The foundation is wrong. I can build it correctly for less than it would cost to fix this, and you’ll have a system your team can maintain.”

They didn’t like hearing it. But it was honest. And they came back six months later - after trying to fix it with another vendor and failing again.

The pattern

Every turned-down project has the same problem: AI before problem definition, data, or success criteria. AI is the last step, not the first.

The correct order:

1. Define the problem (with a number)
2. Assess your data
3. Evaluate whether AI is the right tool
4. Define success criteria
5. THEN build AI

When companies start at step 5, they skip 1-4 and wonder why AI fails. It didn’t fail - it was never the right starting point.

What I say yes to

• Specific problems with measurable impact. “We spend 200 hours/month on manual review” not “we need AI.”
• Companies willing to audit first. If you won’t let me diagnose before building, I can’t guarantee results.
• Teams that want to own the system. If you want permanent dependency, hire a different consultant.
• Problems where the math works. If the problem costs more than the AI solution, it’s worth building.
• Use cases where AI is genuinely the best tool. Not where rules work, not where a spreadsheet works, not where a process fix works.

FAQ

How often do you turn down projects? About 30%. Not because I don’t want the work - because 30% of AI projects shouldn’t be AI projects. The companies that hire me appreciate this. The companies that don’t… hire someone else and learn the hard way.

Does that mean you might turn us down? Maybe. Wouldn’t you rather know before spending €100K? An audit tells us both whether AI is right for you. If it is, we build. If it isn’t, you’ve spent a fraction of the cost and you know why.