Latest / MBA Training Data / Operationalizing the EU AI Act for data teams: a practical playbook
Transcript
- 0:00Welcome to the Leaders Insights Podcast.
- 0:08Today's episode, operationalizing the EU AI Act for data teams, a practical playbook.
- 0:13The EU AI Act's high risk rules became fully binding this month, August 2026.
- 0:18Most companies I talked to have a policy document and nothing else.
- 0:22Is a policy document worthless?
- 0:24Worse than worthless, it's a liability.
- 0:27If a regulator asks how you classify your systems and you hand them a PDF that says,
- 0:32we take AI seriously, you've just documented that you did nothing.
- 0:37The Act, the EU's law governing artificial intelligence by risk level, doesn't care about
- 0:42your intentions.
- 0:44It cares about your logs.
- 0:46So step back for me.
- 0:47Beyond this specific deadline, what's the durable lesson here that the good operators
- 0:51already understood?
- 0:53That compliance is an operational discipline, not a legal one.
- 0:57The teams that are calm right now are the ones who treated data governance, knowing
- 1:01what data you have, where it flows, and who touched it.
- 1:05As plumbing they built years ago, everyone else is trying to retrofit plumbing into a
- 1:09house that's already flooded.
- 1:11Give me the concrete version.
- 1:13What does plumbing actually mean on a Tuesday morning?
- 1:16Take a credit scoring model haunting to deciding who gets alone.
- 1:20That's explicitly high risk under the Act.
- 1:22To comply, you need lineage, a record showing this model was trained on that dataset, which
- 1:28came from these sources, cleaned by this pipeline.
- 1:31If you can't reconstruct that chain, you can't prove your training data was relevant
- 1:35and representative, which Article 10 demands.
- 1:39Most banks I've worked with couldn't tell you what version of a dataset trained
- 1:43the model running in production today.
- 1:45That sounds like an admission of failure.
- 1:47Why couldn't they?
- 1:48Because for a decade, nobody was punished for not knowing.
- 1:52Data scientists pulled a snapshot, trained something, shipped it, and moved on.
- 1:57The dataset got overwritten next quarter.
- 2:00Now the law wants you to keep records for the lifetime of the system.
- 2:04And the honest answer, at a lot of firms, is the evidence no longer exists.
- 2:09That's not a paperwork problem.
- 2:11That's an archaeology problem.
- 2:12So what does the playbook actually tell a CDO to do first, before they panic?
- 2:17Inventory.
- 2:18Boring, unglamorous, and the thing everyone skips.
- 2:22You cannot govern systems you haven't listed.
- 2:24I've seen a mid-size insurer discover 40 machine learning models in production when leadership
- 2:29thought they had 12.
- 2:31The other 28 were built by individual teams, undocumented, some making decisions about
- 2:37claims.
- 2:38You classify each one by risk tier.
- 2:41And suddenly, you know where your real exposure is instead of guessing.
- 2:4540 when they thought 12.
- 2:47Where's that number from?
- 2:49That's from my own engagement last year.
- 2:51One European insurer.
- 2:52So treat it as one data point, not a survey.
- 2:55But the pattern is everywhere.
- 2:57Calibra and OneTrust, the governance tooling vendors, both report clients routinely undercount
- 3:03their models by half.
- 3:05They sell the solution, so discount accordingly.
- 3:08But the direction is right.
- 3:09Let's poke the sore spot.
- 3:11Isn't all this just fear of fines?
- 3:137% of global revenue?
- 3:15I think the penalty caps at?
- 3:17The fines are real.
- 3:18Up to 7% of worldwide turnover for the worst breaches, which for a large bank is a career-ending
- 3:24number.
- 3:25But fear is a bad long-term motivator.
- 3:28The better argument is that the discipline pays for itself.
- 3:32When you know your data lineage, your models break less, your audits take days instead of
- 3:37months.
- 3:38And you can actually answer a customer who asks why they were denied.
- 3:41That's not compliance theater.
- 3:43That's just running a competent shop.
- 3:45So you're saying the firms treating this as a checkbox are missing the actual return?
- 3:50Exactly.
- 3:51The act is forcing people to do what they should have done anyway, know their own systems.
- 3:56The regulation is annoying, but it's pointing at a genuine weakness.
- 4:00The operators who internalized this before Brussels made them are now faster and cheaper,
- 4:05not slower.
- 4:06Give me the one thing a data leader should do the moment this episode ends.
- 4:10Pull a list of every model actually running in production right now, not the ones you
- 4:14remember, the ones deployed.
- 4:16If that list doesn't exist, that's your first project.
- 4:19You can't govern, defend, or improve anything you can't even count.
- 4:23That's your five.
- 4:24The full CDO track, structure no fluff, is at MBA-training.com.