Product and revenue questions without an analyst
Ask a question in Arabic or English and get the number, with the query behind it.
What this fixes
Product, growth and finance each keep their own definition of an active user, so every review starts with a disagreement about the number.
A natural-language layer over a governed semantic model: the question is translated into a query against certified metrics, the answer shows the SQL and the source, and permissions follow the user. The governance work is the product, without a semantic layer, natural-language analytics produces confident, inconsistent numbers.
One certified metric layer
Self-service for product teams
Cohort and funnel questions
Governed access
Where it earns its place
How we build it
Foundation check
Before any buildWe inspect the systems, the data behind them and the quality of the records this capability will depend on, then write down what has to be fixed before anything is built.
Narrow first release
First releaseOne workflow, one team, real traffic, measured against the baseline we captured before we started.
Controls and evaluation
Runs in parallelGuardrails, approval gates, evaluation sets, logging and the evidence pack your auditor and your regulator will ask for.
Scale and hand over
After sign-offAdditional workflows, sites or languages, then runbooks and training so your team can operate it without us.
What has to be right first
- A modelled warehouse or lakehouse, not raw source tables.
- Certified metric definitions with named owners.
- Row- and column-level security already enforced.
- A bilingual business glossary.
Guardrails and KPIs
Every action is logged and replayable
Inputs, retrieved context, model version, output and the human decision are written to an immutable log, so an auditor can reconstruct any single interaction months later.
Data stays where your regulator requires
In-country, in-region or on-premise deployment, with the model portfolio right-sized to the workload instead of routing everything to a frontier endpoint abroad.
Accuracy is monitored, not assumed
A held-out evaluation set runs on every model or prompt change, with drift alerts and a documented rollback path when quality moves.
A human owns every consequential decision
Agents draft, route and execute reversible steps. Anything that moves money, changes a record of consequence or affects a person's rights stops at a named human approver with the reasoning attached.
KPIs it moves
- • Questions answered without an analyst
- • Query accuracy against certified metrics
- • Time from question to answer
- • Reporting backlog
- • Adoption by business unit
Integrations
Proof in this sector
Proof in this sector
Anonymised clientsProduct analytics and usage data foundation
Enterprise software provider in EuropeSupport deflection and engineering-signal automation
Technology platform operating globally