Conversational Analytics

Questions a leadership team asks every week - answered while the meeting is still running:

• How has monthly revenue developed this year?
• Which product categories grew fastest last quarter - and which are declining?
• What share of first-time buyers come back for a second purchase?
• Which countries and cities generate the most revenue?
• Where in the purchase funnel do we lose the most visitors?
• Which traffic sources bring the most valuable customers?

Every answer comes with the key numbers, a chart where a chart helps, and the exact database query used - so an analyst can verify any figure before it goes into a board deck. Nothing is a black box.

The uncomfortable truth about "chat with your data" tools: out of the box, they are confidently wrong. When we started, our assistant answered only 1 question in 10 correctly. By the end it answered 9 in 10 - using exactly the same AI model. The entire improvement came from the foundations underneath it:

• One source of truth for definitions - "revenue", "active customer" or "conversion rate" mean one agreed thing, written into a business glossary the assistant is required to follow, so its numbers match your finance reports
• Automated data quality - every table is checked daily; when the source data broke on day one of operations, the quality gate caught it before a single wrong number could reach anyone
• The platform profiles its own data, automatically - a scheduled step measures what is actually in each table: the real date range covered, the range each figure falls in, the exact spelling of every category. The assistant is told those facts before it answers, so "last quarter" resolves against the dates your data genuinely holds rather than an assumption. This one habit caught a subtle trap in our own source data, where records existed a few days into the future and would have quietly inflated every recent-period total
• Privacy by construction - personal data (emails, names, addresses) is masked at the database level for the assistant itself; there is no personal data for it to leak, even if asked directly
• Strict guardrails - the assistant can only read approved, curated tables, with hard limits on query cost; it cannot modify data or touch anything raw
• Measured, not assumed - accuracy is scored against banks of verified business questions, including a held-out set the system was never tuned on, where it still answered 85% correctly

That is the real lesson of this project, and the reason it is in our portfolio: the return on AI analytics comes from data governance, not from buying a bigger model. Governance work is unglamorous, but it is where the accuracy lives - and it pays off for every future use of your data, not just this one.

There are two routes to conversational analytics on Google Cloud, and instead of guessing, we built both on the same data and evaluated them head-to-head:

• Google's ready-made service (Conversational Analytics API) - integration measured in days, strong accuracy out of the box. The fastest way to validate whether your organisation will actually use this
• A custom-built assistant - the one in the video above. Roughly 40% faster answers, reliable charts with almost every response, identical answers to identical questions (auditors appreciate this), and full control over guardrails, definitions and the user experience

On independent test questions they scored within two points of each other - the difference is not accuracy but control, speed, and user experience. Our recommendation to clients is usually staged: start with the managed service to prove demand at minimal cost, then graduate to a custom assistant where the business case is proven. Because both sit on the same governed foundation, nothing is thrown away in the transition - and the foundation is where most of the value lives anyway.

Deliberately little. The whole platform - daily data pipeline, quality scans, governance, and the assistant - runs for roughly the price of a coffee per month at this scale. Answering a question costs fractions of a cent. The assistant scales to zero when nobody is using it, and hard budget alerts and per-query cost caps are built in, not bolted on.

The same cost discipline carries into client work: we design for your data volumes, with spending guardrails in the infrastructure itself - so an enthusiastic Monday morning of questions never becomes a surprising invoice.

The platform runs entirely on Google Cloud, defined end to end as code (Terraform), with separate development and production environments designed in from day one:

• BigQuery warehouse with layered modelling (raw → staging → curated star schema), transformed by Dataform on a daily schedule
• Dataplex governance: automated data-quality and profiling scans, column-level PII masking with policy tags, data lineage, and a business glossary
• Scheduled profiling scans compute the shape of every table - minimum and maximum of each numeric column, the value domain of each categorical one, and the partition date range actually present - and export the results back into BigQuery, where the assistant reads them at query time
• The assistant itself: a Gemini-powered agent on Cloud Run with guarded SQL execution, cost caps and chart generation - evaluated head-to-head against Google's managed Conversational Analytics API

A detail we are particularly proud of: the business glossary and data-profiling facts from the governance layer are fed to the assistant as grounding context - so the same investment that keeps the data trustworthy is what makes the AI accurate.

Those profile facts are what turn a plausible query into a correct one. A model with only a schema in front of it knows a column is called order_date; it does not know the table stops at last Tuesday, or that the category is spelled Women rather than Womens. Handing it measured facts instead of letting it guess removed an entire class of quietly wrong answers - and because the scans are scheduled, those facts stay true as the data moves, with no one maintaining a list by hand.

Everything shown here transfers directly: retail and e-commerce data is where we demonstrate it, but the same approach fits any company whose answers live in a database - SaaS metrics, finance, logistics, hospitality.

A typical engagement starts small: a free consultation, then a discovery phase where we assess your data foundations and agree the questions the assistant must answer correctly. You see measured accuracy on your own questions before committing to a full rollout. The full engineering write-up and evaluation methodology behind this project are available on request.