For fifty years, the rule was simple: garbage in, garbage out. Bad data produced output that obviously looked bad, so someone caught it and fixed it. AI broke that bargain. Feed a copilot flawed data and it returns a fluent, confident, well-structured answer that happens to be wrong, and nothing on the screen looks broken.

In this episode, Richard Muirhead makes the case that data quality is the one problem AI both makes more dangerous and is genuinely built to solve. He covers why the failure mode went invisible, how AI-assisted profiling, entity matching, observability, and cataloging actually find bad data, where the return on investment shows up, and the three traps that separate a fix that holds from a demo that doesn't.

In this episode:

- Why "garbage in, confidently out" is the new failure mode

- The numbers behind AI-readiness (Precisely 2025, Gartner)

- Four things AI-assisted data quality does that hand-written rules never could

- How to build the ROI case your finance team will actually fund

- Three traps: tooling without ownership, overconfidence, and treating it as a one-time project

Read the full whitepaper: https://themoderncdo.gumroad.com/l/s3e1-Garbage-In-Confidently-Out

Work with Richard: https://perigee.pro

The Modern CDO is written and hosted by Richard Muirhead, fractional Chief Data Officer and founder of Perigee.

Most AI governance ends at the moment the real risk begins. The day a model goes live is the day it starts to drift.

A model learns a snapshot of the world, and the moment it ships that snapshot starts to age. Accuracy and fairness degrade silently, with no alarm, because the governance was all spent at the gate and none on the operation. Watch the model for as long as it runs.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e29-no-post-deployment-monitoring

Generative AI is most dangerous at the exact moment it is most convincing.

It fails fluently. It produces a confident, polished, authoritative falsehood with no signal that it is wrong, leaks whatever an employee feeds it, and varies its output from one run to the next. Govern the real failure modes, not the ones traditional software taught you to expect.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e28-the-generative-ai-traps

A model can be 95 percent accurate and systematically unfair to the people who matter most. The aggregate number hides it.

Treated as a late-stage checklist item, bias gets discovered too late to fix, rationalized away under deployment pressure, or missed entirely. A single accuracy number is not a fairness measure. Measure by subgroup, from the start, where you can still act on what you find.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e27-bias-as-an-afterthought

The moment a real decision can be blamed on a model, you have built a machine for laundering accountability.

A model makes a consequential decision and no human owns it. The data scientist points to the business, the business points to the model, the sponsor treats the output as objective. Everyone is adjacent and nobody is accountable, which means nobody is watching when the decisions drift wrong.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e26-the-accountability-vacuum

When a model starts making bad decisions and you cannot trace what it learned from, you do not have a problem. You have an outage with no off switch.

A model runs beautifully for a year, then starts producing bad outcomes nobody can diagnose, because its lineage was never captured. An untraceable model is an unfixable model. Make lineage a deployment gate, the way tests gate code, and capture it at build time.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e25-model-risk-without-lineage

Standing up a separate AI governance function is how you rebuild data governance from scratch and get it wrong twice.

Most AI risk is data risk with higher stakes: bias from training data, privacy from data use, unexplainability from missing lineage. Wall off a separate tower and you rebuild lineage, quality, and ownership badly, then watch the two versions diverge and open a seam where the worst risks hide.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e24-ai-governance-in-a-separate-tower

Most companies wrote their AI policy before they had any idea how much AI they were running.

Writing the policy is the easy part. The AI it governs is mostly invisible: shadow models, AI features inside SaaS tools, employees pasting company data into public generative tools. You cannot govern what you cannot see. Discovery comes first, and it never stops.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e23-policy-without-an-inventory

AI governance that only knows how to say no does not make AI safer. It drives the risky work into the shadows.

A brake is not a steering wheel. A function that can only stop makes itself an obstacle, and obstacles get routed around. The risky AI work continues with no oversight at all, while the function congratulates itself on what it blocked. Build the steering wheel instead.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e22-governance-as-a-brake

MDM is never finished. It is operated. And the operating budget is the line item every program forgets.

Every MDM business case has a detailed build budget and a stewardship line that is tiny or missing. Go-live is not the end. It is the start of the longest and most important phase. The exception queue that needs human judgment does not shrink after launch. It grows.

Long-form whitepaper is at https://themoderncdo.gumroad.com/l/s2e21-underestimating-stewardship