Artificial Intelligence and Data
Intelligent data analysis
Here “intelligent analysis” does not mean the model inventing a pretty trend from a spreadsheet attached to an email: it means a manager can ask in natural language and receive a number, chart, or text produced by a controlled query against the database or BI cube you already audit. Viscale builds the bridge between a language model and tabular data with the same permissions as the rest of the system—SQL generated only on an allowed schema, row limits, and an explanation of what was summed. Where it fits, we blend classic statistics (moving averages, seasonality) with plain-language summaries for people who do not live in Excel.
We first align business vocabulary: “net revenue” is the same formula in chat and in the official report. Then we map which tables or metrics each role may query—the intern does not see competitor cost. The interface can be portal chat, a Slack bot, or a layer on top of Looker/Metabase you already use; the point is a single source of truth.
Cases we build
“How did last week’s sales go by region?”
Table plus paragraph; numbers match the same filter as the official dashboard.
Sharp conversion drop alert
Rules detect; AI summarizes known factors (paused campaign, holiday).
Churn in board-friendly language
Cohort and rate with explicit definitions; no unnecessary statistical jargon.
Targets versus actuals
Pulls the versioned targets spreadsheet with revenue from the warehouse.
Critical stock question
SKUs below minimum with average supplier lead time when the data exists.
Investor-ready summary
CFO-picked KPIs; short copy with auditable numbers behind it.
Marketing funnel exploration
Impression to purchase by channel; highlights where the drop beats history.
Data quality before a campaign
AI reports invalid or duplicate email percentage with anonymized examples.
Internal benchmark across stores
Ranking on the same criteria; no apples-to-oranges without a warning.
Voice question on mobile
For a store manager: “today’s revenue so far” with optional spoken summary.
To avoid surprises, every “important” answer can show a summarized query or a link to drill-down in the dashboard. We detect anomalies with explicit rules (standard deviation, comparison with the same weekday last year) and use AI mainly to narrate what the chart already shows—handy for the Monday leadership meeting. Model costs stay controlled because we do not ship half a million raw rows into the prompt; we aggregate first.
We also deliver a pack of “saved questions” finance or sales validated—shortcuts so everyone does not rediscover the wheel. When data is dirty, we prefer to say “we cannot answer well until X is fixed” instead of confabulating.
Portfolio of Intelligent data analysis
Deliverables
Production experience
Chat or layer attached to the agreed BI stack.
Data and metrics map
Tables, joins, and definitions used in answers.
Dictionary of approved questions
Shortcuts with SQL or equivalent versioned logic.
Validation report
Sample questions checked against official numbers.
Access policy
Roles and examples of what each role sees.
Versioned code or configuration
Repository with a deploy pipeline.
Cost and latency monitor
Dashboard or export for FinOps.
Runbook
Model provider down, warehouse slow, schema changed.
Automated tests
Minimum CI suite so “numbers that match” never silently breaks.
Handoff session
The data team owns glossary maintenance.
Good-question guide
For business users to get value without extra risk.
Source backlog
Next tables or cubes to connect, prioritized.
Execution methodology
-
Metric and source inventory
Where truth lives and who may see each number.
-
Business glossary
Definitions aligned with finance and product.
-
Permission model
Same or stricter than current BI.
-
Safe query layer
SQL or API generated only on approved schema with row caps.
-
Question prototype
Validated list with answers checked manually.
-
Narrative and visuals
Answer format: table, simple chart, or prose.
-
Optional anomaly detection
Rules plus assisted explanation without mixing cause and correlation.
-
Channel integration
Web, Slack, Teams, or an extension on the existing dashboard.
-
Regression tests
Golden questions on every model or schema change.
-
Training and governance
How to propose a new safe question and who approves.
-
Evolution plan
Next sources or languages in the right order.
