We Estimated the Real Compliance Cost of Upcoming AI Regulation

A client asked us for a real number on what complying with upcoming AI regulation would cost them, not a vague directional answer. Doing that exercise honestly meant separating two very different categories of cost that most vendor pitches lump together into one scary total, because keeping them separate is what made the number actually useful for planning.

The cheap part: documentation and process

Most of what emerging AI regulation actually asks for — documented model purpose, a record of what data trained it, a basic impact assessment before deploying a high-stakes decision system, a process for handling appeals or corrections — is work that's genuinely cheap if you build the habit of doing it as you go. For this client, retrofitting documentation on their existing models took a few weeks of one person's time, not a major budget line.

Most AI regulation compliance cost isn't about building new capability. It's about writing down capability you already have but never documented.

The expensive part: the structural gap

This client's genuinely expensive gap was explainability infrastructure — several of their highest-stakes models were black-box ensembles with no existing explanation layer, built before anyone was thinking about disclosure requirements. Retrofitting real explainability onto those specific models, not the documentation around them, was the line item that actually moved the total estimate from manageable to significant, and it's the part that needed a real multi-month engineering investment rather than a documentation sprint.

• Documentation and process gaps: cheap to close, mostly a matter of discipline and a few weeks of focused work.
• Explainability retrofits on existing black-box models: the genuinely expensive category, requiring real engineering time.
• Starting documentation habits now, regardless of final regulatory shape, avoids a future scramble under deadline pressure.
• The explainability gap is worth addressing proactively precisely because it's expensive — waiting makes the eventual retrofit more expensive, not less.

The honest total estimate wasn't the scary number the client braced for, but it also wasn't nothing — and the value of separating the two categories was that it told them exactly where to actually spend money now versus where they could reasonably wait for more regulatory clarity before committing engineering time.