What Happens When Your AI Agent Finds Two “Correct” Answers?

Two teams pitch the same figure for ‘active customers’ right before a leadership meeting. One dashboard displays 40,000 while the other dashboard displays 52,000. No mistakes were detected by the teams; however, the difference in the stats is due to one team or employee deploying a different data filter and the other using another join path. 

Now there's an AI agent involved, who is simply asked to 'retrieve the most recent number of active customers'. Not realising that there are two versions, it selects one, states that figure with complete confidence, and then goes about its business. In that moment, the semantic gap ceases to be merely an internal discussion and becomes a business risk.

On this episode of the Don't Panic! It's Just Data podcast, host Shubhangi Dua, Podcast Producer and B2B Tech Journalist at EM360Tech, is joined by Ananya Devraju, a Business Intelligence Developer at Premier Inn, to discuss the semantic gap and how AI is rendering an old data problem. They look at the situation in which metric definitions are spread across five different dashboards rather than kept in a single central location, and explain why an AI agent can produce a perfectly valid SQL query yet return a completely incorrect answer.

The problem, consistent even with AI in the picture, is governance. AI has just made the issue of governance more apparent. “It was never a data quality problem, but it's a governance of meaning problem, and it's been sitting there long before AI showed up. AI has just made it louder now,” Devraju tells Dua. 

'Wrong Data Is Worse Than No Data'

Devraju has worked on both sides of the data divide. First, she worked as a data scientist who used data, and then switching over to Business Intelligence (BI) and analytics engineering, where her role involved determining what a metric meant and which joins were appropriate. According to her, it was at this point that the real issue became apparent: two teams or two dashboards referring to the same metric such as revenue, occupancy, or the number of active customers were each technically correct within their own context, yet they used different date logic, different filters, or a different join path. No one had made a mistake; they just never reached an agreement on a single definition.

Devraju tells Dua that if there is no central semantic or metrics layer, then each BI tool, each analyst, and each ad hoc query ends up working out its own version of a metric using different joins, different filters, and different levels of granularity. 

In this way, an enterprise ends up with five different definitions of "active user", each of them residing in a separate dashboard, none of the definitions being incorrect by itself but all of them incompatible with one another. The responsibility then has to be shared between both sides. On the one hand, developers having to reinsert the same logic every time a new report involves the metric, and on the other hand, business users having to reverify the definition every time they access a different dashboard.

The solution that Devraju suggests is not a particular product; while she cites enteprises such as dbt Labs as examples, she emphasises that the issue is one of structure rather than one of tools. What’s needed right now is an established governed semantic or metrics layer, with named owners assigned to each domain, version control being applied, and it being treated as the one source that both manually created dashboards and AI agents draw on.

The Investment Gap Nobody's Budgeting For

Devraju cites research from dbt Labs' State of Analytics Engineering stating that 72 per cent of respondents now prioritise AI-assisted coding in their development workflows, while only 24 per cent prioritise AI-assisted pipeline management, including testing and observability; spotlighting an imbalance between acceleration and quality. 

“Everyone's essentially investing in the AI agent to write faster query, but far fewer people are investing in checking whether those answers mean what it claims to mean,” she said.

Devraju views validation as something that should be built into the pipeline from the start, rather than being added to it later on. Applying testing and observability to the actual output of the AI, not merely to the fact that the scheduled job has run successfully is essential. A refresh job may finish without any errors and yet still produce a figure that is quietly incorrect; only by monitoring the output can such a mistake be detected, not just by checking that the pipeline is up.

The other question that popped up was who is responsible for end users. “As an end user in a boardroom, who has no way to independently verify the data presented to them in the moment, the trust has to be built into the system before the answer even reaches them,” Devraju tells Dua. 

“It's not really the end user's job to verify the information at in the moment. It's the burden of leaders presenting the data to the boardroom and their team behind the report.”

Instead of having SQL queries run through side channels to produce the 'official' figures and clear, documented ownership for each key metric. Leaders should ask their data teams to provide traceability (that is, the official definition or source from which an agent drew its answer), a single, certified semantic layer that is used by both AI and humans when querying it. 

Overall, the disagreement regarding the meanings of "revenue" or "active customer" existed before AI came along. What AI does is provide a faster, smoother, and more confident answer than ever, on a large scale, however, without ever stopping to verify. 

Putting an AI agent into a warehouse that has never reached agreement on the definition of its own metrics doesn't address the fundamental issue, it only causes the incorrect answer to be presented with greater confidence.

Devraju will be presenting a talk at Big Data LDN (BDL) this year on Thursday, September 24, on When Migrations Break Your Metrics: Rebuilding Data for Commercial Analytics. The talk is scheduled for 3:20 pm to 3:50 pm in the Data Architecture Modernisation Theatre.

Takeaways

• Semantic gaps existed before AI; AI has only made them louder and faster.
• The same metric is displayed on different dashboards, each using different date logic, filters, or join paths.
• There can be five different definitions of an "active user" across the five dashboards, all of which are incompatible.
• SQL can be syntactically correct and yet provide the wrong answer to a business question.
• Validation should be placed at the definition level, not at the SQL level.
• AI agents can query the raw staging tables rather than referring to the governed marts by name alone.
• Incorrect joints or the wrong grain result in numbers that look plausible but are actually wrong.
• According to dbt Labs research, 72 per cent of peopleprioritise AI-assisted coding while 24 per cent prioritise validation.
• Incorrect data is more dangerous than having no data; it is the confident errors that pose a risk.
• Since end users are unable to check the answers at the time they occur, trust has to be established earlier on.
• Provide a semantic layer that is governed and has version control, with designated owners for each metric.
• Ad hoc SQL statements on no side-channel should generate the "official" figures that are reported.

Chapters

• 00:00 Introduction to the Semantic Gap in AI
• 02:56 Understanding Data Interpretation and Governance
• 05:56 The Importance of Centralized Definitions
• 09:05 AI Hallucinations and Data Validity
• 12:02 Navigating Governance in AI
• 14:56 The Role of AI in Data Validation
• 17:56 Final Thoughts on AI and Data Meaning