Hakkoda and IBM / August 2026
Plain version.
They have a solid setup: 9 semantic views feeding their AI, 100+ defined business terms, a clean data model, good dashboards. The definitions are written down.
But the definitions live in three places, Collibra, GitHub, and Snowflake, and nothing keeps them in sync. Someone changes the data, the definitions don't follow, and nobody finds out. Five of nine objects are already out of step.
Everything below fixes that one thing.
Git is the only thing here that remembers. Every change gets a timestamp, an author, and an undo button. If someone asks why this number changed last March, Git can answer. Collibra can't.
Git also runs the tests. Collibra just stores text. If you want to know whether a definition actually works, you need something that executes it.
Don't move this. It's already right.
Right now nobody has said which system is the real one. That's not a technical problem. It's just an unmade decision. Pick one.
Collibra should win because business people and auditors have to be able to read it, and they're not going to read code files.
Collibra says what things mean. GitHub says how they're built. Both are true, and they don't overlap.
The split above only works if Collibra never falls behind. The second it does, it's just a stale copy, which is exactly the setup they have now, and the worst of the four options.
The automatic push to Collibra has never been tested. The Brazil pilot skipped it. That's the thing to fix first.
Today they check data quality after it loads, on a schedule. So bad data reaches the AI first and someone notices the next morning. That's backwards.
Check at the moment data changes, not hours later.
Collibra defines the rule. Snowflake runs it. Grafana shows it. Each tool does the one thing it's good at.
Do what the AI actually needs first, not what's easiest to type in.
| 01 | Business terms and allowed values | The AI can't understand a word nobody defined. |
| 02 | Owners | No owner means nobody can approve a change. |
| 03 | Regulatory tags | Biggest risk, especially for GxP and GDPR data. |
| 04 | Lineage | Useful, but only after the above. |
| 05 | Policies | Last. |
The current checklist grades the data. It says nothing about whether the AI using that data gives good answers. Two different questions, so track them separately.
Add four things.
Pfizer already answered some findings with "we don't need those columns" and "cardinality is set, it just doesn't show in the file." Both fair. Don't argue.
Say instead: the specific columns were never the point. The point is nothing tells you when things drift apart. Argue for the alarm, not the individual findings.
| 01 | Who owns the meaning of words? | Every new AI use case makes this messier. Decide now. |
| 02 | When do we check quality? | A check after the fact is a report, not a control. |
| 03 | What does ready mean? | Not a one time score. It expires every time the data changes. |
Semantic problems don't stay in the semantic layer. They show up as wrong answers, slow answers, and higher cost, on every single query.
If they only do one thing
Everything else gets easier after that.