Ask your data questions in plain language. Get the right answer back.
OFFERINGS:
AI ASSISTED ANALYTICS
WHY:
The holy grail for data teams: Let anyone ask a question, and the data answers.
Organisations sit on rich data, but getting answers still depends on a handful of data engineers and BI specialists, manual reports, and long turnaround times.
With Natural Language Querying (NLQ) maturing rapidly in platforms like Fabric (Data Agents) and Databricks (Genie Space), the timing was right to test whether AI could close that gap.
IN BRIEF
Why:
Every organisation has data. Few can answer questions without a BI specialist in the loop. Ad-hoc requests from leadership, partners, and external stakeholders pile up. Most employees have no way to explore the data themselves.
You can turn on Fabric Data Agent or Databricks Genie Space today. They will give you answers. Most of them will be wrong - and they won't tell you which ones.
We tested it. Out of the box, the AI returned the wrong answer 84% of the time. Not because the model failed. Because the metadata wasn't there.
What:
We run structured, time-boxed experiments on your data. Real business questions. Progressively richer configurations - from a zero-setup baseline to a fully tuned semantic layer with domain-specific metadata and instructions.
Every response scored by an LLM judge. No subjective grading. Repeatable and systematic.
Where single-agent setups fall short, we add a second agent: one generates the query, another validates the reasoning and catches hallucinations before the answer reaches the user.
This is context engineering - configuring the system to know your data the way your best analyst does.
How:
We start narrow. Five to ten high-priority questions, selected with your stakeholders.
We test each question under controlled scenarios and trace the agent's reasoning to pinpoint systematic failure patterns. In our first experiment, we found four: time filtering, measure selection, filter inconsistency, and temporal hallucination.
Each is addressable through metadata enrichment, better instructions, and more precise question formulation.
The result is a concrete roadmap: which tables to enrich first, which instructions to write, and which questions to start with.
what we know:
1.
It works - but only with the right foundation!
We tested NLQ on real client data. Out of the box, the AI returned the wrong answer 84% of the time.
Not because the model failed, but because the metadata wasn't there. With a tuned semantic layer, domain-specific instructions, and structured validation, accuracy changed completely.
The difference isn't the technology. It's the setup.
2.
A second agent catches what the first one misses.
Single-agent NLQ hallucinates. It generates plausible SQL, returns a confident number, and gets it wrong.
Our multi-agent architecture separates query generation from answer validation. One agent writes the query, another checks the reasoning before the answer reaches the user.
Hallucinations get caught. Trust goes up.
3.
Start narrow. Prove value. Then expand.
NLQ gets more precise as it gets more specialised. The organisations that succeed don't try to make all their data queryable at once.
They pick five high-impact questions, nail those, and build from there.
For mid-sized organisations, that changes the economics entirely: a focused scope means faster results and a clearer business case.
contact us
