By Mo Aidrus, co-founder & CEO, ekai, Inc.
Picture a quarterly business review that falls apart at slide four when the discussion shifts to customers. Finance reports the company has 41,200 active customers while sales reports 47,900. Then the CEO mentions asking the company’s new AI assistant the same question last night. The answer they received? 44,600, accompanied by a tidy explanation and no hesitation.
All three numbers came from the same warehouse, and all three systems worked exactly as designed, yet no one can say which definition of “active customer” is right, who decided it, or when.
Every tool in your company draws on meaning inherited from somewhere else, and almost nobody can tell you where it came from.
Generation Doesn’t Equal Meaning
A few years ago, mapping how dozens of tables connect took weeks of analyst work. Today, most capable AI systems can produce a semantic model, a metric definition, or a confident answer from raw data in seconds. That capability arrived faster than anyone expected, and the progress is real.
But while a machine can now tell you, almost perfectly, how your data is wired together, it cannot tell you what it means, and that’s why generation has become the least useful part of the problem.
Meaning Doesn’t Live in The Data
Here are two examples that most companies face today, in some form. Start with payments. At one company, a “payment” in a supplier contract means Net Cash Adjusted in the ledger, and only records marked settled count.
Another example is delivery time. For one company, “delivery time” starts when an order is marked picked, not when the truck leaves, because that is when the customer’s service-level agreement begins.
Neither rule is in a column name or can be inferred from a table. Both live in the heads of people who have worked the data for years, including the controller who remembers why the adjustment exists, and the operations lead who negotiated the SLA.
Most auto-generated semantic layers are built from past usage, such as query history and existing dashboards. As a result, they reflect what people have already asked, not what the business actually means, and they give no signal when the information they share is wrong. No amount of generation speed recovers knowledge that was never written down.
It’s All About the Context
On a set of Spider 2.0 questions, our team recently tested questions answered 90.0% correctly, ten points ahead of the best coding agent. Spider 2.0 is public, and public benchmarks are routinely absorbed into LLM training data. A company’s private data warehouse is not. On enterprise data, coding agents face an unfamiliar schema and business definitions they have never seen, so they struggle far more, while ekai’s semantic models supply exactly what each question depends on. That is where ekai’s advantage is greatest
No business should run on numbers that are wrong. Your warehouse isn’t. No model has seen your definition of an active customer or your rules for settled payments, so on private data those scores would fall further. What closes the gap isn’t a stronger model. It’s context, meaning knowing what the data actually means.
Being Wrong Got Expensive
Before the arrival of AI agents, a bad definition usually lived in one dashboard, and eventually an analyst caught it. Today an agent reuses that definition across every question, every team, and every downstream decision. And it does so at machine speed while rarely flagging uncertainty.
When nobody checks whether a generated definition is correct, teams end up working from different numbers for the same metric. This erodes trust long before anyone discovers which definition was wrong. And trust doesn’t break selectively. Once people doubt one number, trust is eroded, even if the remaining data is accurate.
A Blue Tick for Data
The fix is to make human-verified definitions visible. Think of it as a blue tick for data, showing that someone who actually knows the answer checked a metric. But a tick alone isn’t enough. For it to mean anything, it has to certify four things:
- A Designated Owner. Not “Finance,” but the name of the person in that department.
- Reconciliation. The definition was checked against what the data actually shows, not only what the expert believes.
- A Documented Scope and Exceptions. Which records count, which don’t, and why.
- A Review Date. Definitions drift as the business changes, so the tick should clearly show when someone last checked.
The blue tick also has to be machine-readable. An AI agent should answer confidently from a verified definition and say so when it is reasoning from an unverified one. That is how people and agents can trust a number on sight instead of re-litigating it with every query.
The social-media era offers a warning: a checkmark that can be bought or auto-assigned quickly stops meaning anything. Verification carries weight only because a person in the know has put their name on it.
Framed this way, verification isn’t a compliance step bolted onto the end of a data project. It is the product.
Where Chief Data Officers Should Start
- Pick 20 Metrics That Reach Your Board. Assign one named owner to each.
- Test and Then Test Again. Before evaluating any AI analytics tool, agree on a set of questions with known correct answers. Lock it in before results arrive, not after.
- Demand Provenance. Require AI agents to show which definition they used to answer a query and whether they verified it.
- Measure the Asset. Make verified definitions a KPI for the data team. In the AI era, this is the asset you can’t buy from a model provider or scrape from the web.
Go back to that quarterly business review and run it again with one change. The answer on “Active customers” includes the finance owner, a written scope detailing which customers count, and the date the information was last checked. Now, instead of debating the right figure, the team can focus on what the number means for the business.
Building got cheap. Being wrong got expensive. Being sure is now the part worth investing in.
