AI Data Quality: Moving from Useless Dashboards to Executable Rules
A 98.7 percent green data quality dashboard can measure only presence and type, not meaning. The video shows how AI can profile columns and propose executable rules with owners, and how people ratify them and failing rows are quarantined instead of silently fixed.
Hussain's video is about how AI can help an organization manage its data standards and data quality, and specifically not the dashboard version. It is about the version where somebody has to decide what valid means for 4,000 columns and then keep that decision true every day while the data changes underneath. The central claim is that the useful output is a rule, not a report.
The green dashboard illusion
Every organization already has a data quality dashboard, and the one in the video is 98.7 percent green. Nobody argues with it. But what it actually measures is whether a field is present and whether it is a string, so six spellings of one country all pass. The score is not wrong, the video says; it is measuring the wrong thing.
Why nobody writes the rules
The honest answer is to write more rules, and that is exactly why nobody does it. Of the 4,000 columns, 12 have a written rule. Getting one right, which means profiling the column, arguing the edge cases and agreeing with the owner, costs about 40 minutes. That adds up to 330 working days to write them once, and they go stale quickly afterward. A standard that lives in a policy document is a wish, while a standard that can fail a row is a rule. A rule names the allowed set, the scope it applies to, who owns it, and what happens when it breaks: block the load, or let it through with a warning.
The ask that works
The video contrasts a poor request, "clean my data," which produces a silent rewrite you cannot audit, with a better one: profile one column, propose the rule, show the evidence, name the owner, and do not change a single value. The result is six spellings ranked with counts, the proposed rule beneath them, and the number that matters most, which is that 1,284 rows would newly fail. The video calls that a negotiation opening rather than a bug report, because the real question is whether it is a data problem or a rule that is too strict.
The model proposes, people ratify
Every proposal stops at a human. In the video's example, 218 rules were ratified, 59 were bounced to an owner who knows the domain, and 41 were rejected because the model pattern-matched a rule the business does not have. Once ratified, rules run on every load. A row that fails a blocking rule never reaches the warehouse. It goes to quarantine carrying the rule that caught it and the person who owns it, so the dashboard becomes a queue somebody works.
Three ways it goes wrong
- Auto-correcting values leaves a silent edit nobody can trace.
- Every value can be valid and still meaningless; a column of perfect dates that are all the load date passes every check.
- Rules rot when the business changes and the check does not.
Key takeaways
- A high green score can measure only presence and type, not meaning.
- Writing rules by hand does not scale across thousands of columns.
- AI's contribution is proposing and maintaining rules with evidence attached, not cleaning values.
- Each rule needs an allowed set, a scope, an owner and a failure action.
- Every proposal must stop at a human owner for ratification.
- Failing rows go to quarantine with the rule and owner attached.
- The test is whether a rule can fail a row and whether that failure reaches a person.
Try it yourself
Take one table you own and ask Claude to propose, for each column, one executable expectation with a severity and an owner. Have it rank them by what a silent failure would cost and list every column where it cannot infer a rule and must ask a human. The video says to watch that last list hardest, because those columns are the ones your organization never actually decided.
Chapters
- 0:00The 98.7% green dashboard illusion: Presence vs meaning
- 0:15The 330-day manual rule writing bottleneck
- 0:30AI profiling: Proposing rules and negotiating 1,284 failing rows
- 0:50Human ratification and quarantining bad data with attached owners
- 1:10Three fatal pitfalls: Autocorrect, meaningless dates, and rule rot
Full transcript(auto-generated, with timestamps)
The 98.7% green dashboard illusion: Presence vs meaning
[0:00]Hi, I am Hussein in this video is about how I actually helps an organization manage its data standards and its data quality. Not the dashboard version. The version where somebody has to decide what valid means for 4,000 columns and then keep that decision true every
The 330-day manual rule writing bottleneck
[0:15]Single day while the data changes underneath them. Every organization already has a data quality dashboard 98.7% green. Nobody argues with it. Here's what it is actually measuring is the field present and is it a string. Six spellings of one country all pass. The
AI profiling: Proposing rules and negotiating 1,284 failing rows
[0:31]Score is not wrong. It is measuring the wrong thing. So write more rules. That is the honest answer and it is exactly why nobody does it. 4,000 columns, 12 of them have a written rule. Getting one right, profiling the column, arguing the edge cases, agreeing with the owner costs about 40 minutes. That is 330 working days to write them once. They go
Human ratification and quarantining bad data with attached owners
[0:50]Stale in 90. A standard that lives in a policy document is a wish. A standard that can fail a row is a rule. Same sentence, one difference. This one names the allowed set, the scope it applies to, who owns it, and what happens when it breaks. Block the load or let it through and warn. That is the whole art. So here's the ask that works, not clean my data. That one gets you a silent
Three fatal pitfalls: Autocorrect, meaningless dates, and rule rot
[1:10]Rewrite you cannot audit. Profile one column, propose the rule, show the evidence, name the owner, and do not change a single value. And this comes back six spellings ranked with counts, the proposed rule underneath it, and the number that matters most, 1,284 rows would newly fail. That is not a bug report. That is a negotiation opening. Is this a data problem or is my rule too strict? This is the part that gets skipped and it is the part that makes the whole thing work. Every proposal stops at a human. 218 ratified, 59 bounced to an owner who actually knows the domain. 41 rejected because the model pattern matched a rule the business does not have. The model proposes, people ratify. Then the rules
Stop being a document. They run on every load. A row that fails a blocking rule never reaches the warehouse. It goes to quarantine carrying the rule that caught it and the person who owns it. The dashboard stops being a score you look at. It becomes a queue somebody works. Three ways this goes wrong. One, let it auto correct and you have inherited a silent edit nobody can trace. Two, every value can be valid and still be meaningless. A column of perfect dates that are all the load date passes everything. Three, rules rot, the business changed, the check did not. Let us recap. The dashboard was never measuring meaning. It was measuring presence and type. Eyes real contribution is not cleaning values. It is
Proposing and maintaining the rules with the evidence attached. Every proposal still stops at a human owner. And the test for all of it, can this rule fail a row and does that failure reach a person? Your turn. Take one table you actually own and paste this. For each column propose one executable expectation with a severity and an owner. Rank them by what a silent failure would cost and list every column where you cannot infer a rule and must ask a human. Watch that last list hardest. The columns it refuses to guess on are the ones your organization never actually decided. The rule, not the report. If this changed how you will look at your next quality dashboard, subscribe. I am Hussein. Thanks for watching.
More from Humanitarians AI Fellows
2:00AI in Data Engineering: Collapsing ETL Glue Without Losing Judgment
6:34How Do You Know It Worked? Grading Claims Over Stories
2:32Agent Harness: The Loop, Tools, and Stop Conditions
3:42Planning & Decomposition: When AI Agent Plans Are Worth the Cost
2:33Why Godot is the Best Game Engine for Building Games with AI
5:16