Softpiper Solutions

Any industry · Knowledge and data

“Talk to your data” analyst

Let business users question Postgres, BigQuery or Snowflake in plain language, with read-only access, a semantic layer, charts and an evaluation set before rollout.

  • ClaudeAnthropic
  • GeminiGoogle
  • Typical timeline: 3–6 weeks

The problem

Business users wait days for an analyst to pull simple numbers, and dashboards can't answer the follow-up question.

  • Analysts become a ticket queue for simple questions.
  • Dashboards answer yesterday's questions.
  • Ad-hoc SQL against production is risky.

How we solve it

An analyst agent that turns plain-language questions into SQL over a read-only connection, explains the answer, draws a chart and shows the query it ran, checked against a golden set of questions before launch.

It plugs into Postgres / Supabase, BigQuery, Snowflake and Slack or Microsoft Teams, so nothing you rely on today is replaced. The architecture below shows each step and which model does the work.

Architecture

From your stack, through the AI layer, to numbers you can track

Each step names the model or system doing the work. Writes to your systems wait for a person to approve them.
“Talk to your data” analyst. Your systems (Postgres / Supabase, BigQuery, Snowflake and Slack or Microsoft Teams) connect to Softpiper's AI layer, which runs these steps: Semantic layer (Prompt caching); Ask in plain language (Slack · Teams); Query safely (Claude Sonnet · read-only SQL); Explain and chart (Code execution); BigQuery-native option (Gemini in BigQuery); Evaluate (Evaluation set). Outcomes are measured as analyst ticket volume, time to answer, accuracy on the golden-question set and weekly active business users.

Your stack

  • Postgres / Supabase
  • BigQuery
  • Snowflake
  • Slack or Microsoft Teams

AI layer

Softpiper AI layer

  • Claude
  • Gemini
  1. Semantic layer, using Prompt caching

    Tables, joins and metric definitions are documented once and cached.

  2. Ask in plain language, using Slack · Teams

    Users ask in Slack, Teams or a web app.

  3. Query safely, using Claude Sonnet · read-only SQL

    SQL runs through a read-only role with row-level security and a cost ceiling.

  4. Explain and chart, using Code execution

    The answer comes with a chart and the SQL that produced it.

  5. BigQuery-native option, using Gemini in BigQuery

    Google Cloud teams can use Gemini's BigQuery integrations.

  6. Evaluate, using Evaluation set

    A golden set of real questions is scored before rollout and after every change.

Outcomes you measure

  • Analyst ticket volume
  • Time to answer
  • Accuracy on the golden-question set
  • Weekly active business users

Why this model

Claude and Gemini, each where it leads

We choose per task on capability, cost and where your data must stay. Final versions are picked at kickoff.
ClaudeAnthropic

Claude writes and explains SQL through a read-only tool or MCP server, with your semantic layer (table meanings, metric definitions) in a cached prompt, and draws charts with code execution on the Claude API.

GeminiGoogle

If your warehouse is BigQuery, Gemini's native BigQuery integrations suit teams already working in Google Cloud.

Delivery

What you get, how long it takes, and how we keep it safe

Guardrails

  • Read-only credentials with row-level security.
  • A query-cost ceiling on every request.
  • Every answer shows the SQL it ran.
  • A golden-question evaluation set must pass before rollout.

Plus the guardrails on every build: facts from your systems, approvals for writes and money, a human one step away, audit logs and evals. See all

Deliverables

  • Semantic-layer documentation
  • Analyst agent with read-only tools
  • Slack, Teams or web interface
  • Golden-question evaluation set
  • Usage and accuracy dashboard

Typical timeline

  1. Semantic layer, agent and evaluation3–6 weeks

A planning range for a first production release. We confirm scope, price and dates in a tailored proposal.

KPIs to track

  • Analyst ticket volume
  • Time to answer
  • Accuracy on the golden-question set
  • Weekly active business users

Goals we agree with you and measure from a baseline. They are what we track, not results we promise.

Related services

The services behind this playbook

More playbooks

Related playbooks

All playbooks

Pip · Softpiper's AI

Ask Pip about this playbook

Ask Pip how this would work for you

Pip is Softpiper's AI. Tell it what you run today and it will walk through how this playbook fits, what we would build first and a tailored price range. Or book a call with the team.

Book a call