We get calls from companies who hired an AI vendor, spent six figures, and ended up with something that doesn't work. This happens more often than you'd think. Here are the things we wish someone had told them before they signed.

"We'll handle the data" usually means they won't

The number one thing AI vendors gloss over is data preparation. They'll show you a slick demo running on a perfectly formatted dataset. Then when they get access to your actual data — with its inconsistencies, missing fields, legacy formats, and that one system that exports everything as semicolon-separated PDFs — suddenly the timeline doubles.

Ask your vendor: "What percentage of this project's budget is allocated to data preparation?" If the answer is less than 25%, they either don't understand your data or they're planning to charge you for it later as "change requests."

"Our model is 95% accurate" means nothing without context

Accuracy percentages are meaningless without knowing what they're measured against. 95% accuracy on a clean benchmark dataset is very different from 95% accuracy on your messy real-world data.

We had a client whose previous vendor claimed 98% accuracy on their document processing system. When we audited it, we found they were measuring accuracy on the easy cases only. The system was routing 40% of all documents straight to a human queue as "low confidence" — and those weren't counted in the accuracy metric. The real accuracy on all documents was closer to 58%.

Ask: "Is this accuracy measured on all inputs, including edge cases? What happens to documents the system can't process?"

Lock-in is real and expensive

Some vendors build everything on proprietary platforms. Your AI system runs on their infrastructure, uses their tools, and stores your data in their format. If you want to switch vendors — or just bring it in-house — you're starting over from scratch.

We've seen companies spend €200K+ on a system they couldn't take with them when the relationship soured.

Before signing, ask: "If we end this contract, what do we walk away with? Can we export our trained models? Our data pipelines? Our integration code?" If the answer is vague, that's your warning.

"Maintenance is minimal" is almost never true

AI systems need ongoing care. Models drift as data patterns change. APIs get deprecated. Source systems get upgraded and break integrations. Edge cases that didn't exist at launch start appearing six months in.

Budget at least 15-20% of your initial project cost per year for maintenance. If your vendor's proposal doesn't include a maintenance plan, ask why. Either they're planning to charge you separately (at higher rates), or they genuinely think the system won't need updating. Both are red flags.

The talent question nobody asks

Who actually builds your system matters. A lot of vendors sell with senior engineers and then staff the project with juniors. You meet the PhD during the sales pitch. The person actually writing your code just graduated.

Ask: "Can I meet the team that will work on my project? What's their experience with similar implementations?" And ask for references — not just company names, but actual people you can call.

What good looks like

Here's what we think a fair AI engagement should include:

  • A discovery phase where the vendor spends time understanding your actual data and processes before quoting a price
  • A realistic data preparation budget (25-40% of total project cost)
  • Clearly defined accuracy metrics measured against your data, including how edge cases are handled
  • Your code, your models — you should own everything that's built, with full access to source code and trained model weights
  • A maintenance plan with defined SLAs and costs from day one
  • Honest timelines — if someone promises an enterprise AI system in 4 weeks, walk away

The discovery phase on that list is what our fixed-price AI assessment delivers, before any implementation quote.

We don't always win deals against cheaper competitors. But the clients who come to us after a failed engagement usually say the same thing: "I wish I'd asked these questions the first time."

The full question list is in How to Choose an AI Consulting Firm in Hungary.