How it works

You ask.
It investigates.

picoask isn't a chart generator with a chat box. Every question runs through a planner-driven pipeline of specialist agents that form hypotheses, test them against your live data, and only then write the answer — with the evidence attached.

picoask · Stripe Payments DB
a real picoask session, replayed
The pipeline

Three stages behind every answer.

STAGE 1 · PREP

Gate the question

Before any analysis runs, picoask checks that your question is answerable: it asks a clarifying question if the ask is ambiguous, and verifies your data is fresh enough to trust — so you never get a confident answer built on a stale table.

STAGE 2 · INVESTIGATE

Plan, query, test

A planner decides which specialists the question needs: a hypothesis agent proposes competing explanations, a data-retrieval agent fetches exactly the data required, and statistical agents test each hypothesis against your data.

STAGE 3 · REPORT

Answer with evidence

The reporting stage picks the right visualization, writes a business-readable answer, and attaches everything it used — the queries, the tables, the significance tests — plus the concrete next step the finding suggests.

Evidence, not vibes

Statistics are a first-class output.

Significance testing

Every "X changed" claim carries a verdict: significant or not, with the test and p-value shown. A 4% lift that's indistinguishable from seasonality gets called exactly that.

Real computation

Every check runs through picoask's library of proven statistical algorithms — your data is fed into the algorithm and the result is computed exactly, never estimated by a language model doing arithmetic.

Correlation ≠ causation

A dedicated classifier labels each relationship it reports: correlated, plausibly causal, or confounded — so a finding never quietly overstates what the data can support.

Anomaly detection

Metrics are checked against their own history — trend, seasonality, and noise — so "this week looks weird" comes with a baseline, not a hunch.

Two-sample t-test · daily failed charges p = 0.003 · significant
May μ 5.0/day June μ 7.1/day failed charges per day →
Each dot is one day. June sits visibly right of May — and the shift is tested, not eyeballed: a two-sample t-test on the daily counts puts it beyond chance.
Onboarding

It learns your data before you ask.

Schema, described

On connect, picoask introspects your schema and writes a plain-English description of every table and column — which you review, correct in a chat, or reject before it's used.

A metric graph

It maps your business metrics and how they drive each other — revenue ← conversions ← traffic — so root-cause questions follow real relationships instead of assuming every relationship is true.

Drift, caught

Schemas change. picoask scans for changes automatically — and you can trigger a scan any time. Anything new or altered shows up with an AI-drafted description — approve it, refine it in chat, or reject it. Your project never silently rots.

Metric graph · revenue project 12 edges · 9 verified
Ad spend Signups Activation Net revenue Discount depth verified against your data refuted · excluded
Proposed from your schema at onboarding, then every edge is tested against your history. Root-cause answers follow only the edges that survive.
Always on

It keeps watching after you close the tab.

Metric watches

Put a watch on the metrics you care about and picoask monitors them continuously, notifying you when one moves in a way its history says it shouldn't.

Scheduled reports

Any answer can become a recurring report — re-run on your schedule and delivered to your email, with the same statistical scrutiny every time.

A knowledge base

Tell it what the data can't know — a pricing change, an outage, a campaign launch — and it uses that context when explaining anomalies. Add facts with /memory.

Metric watch · failed charges / day anomaly flagged · alert sent
expected band +38% above expected Jun 2 Jun 16 Jun 30
A watch learns the metric's normal band from its own history. The day reality leaves it, you get pinged — with the root-cause analysis already attached.
Security

Your data stays yours.

Read-only, enforced

picoask connects with a read-only database role and verifies it really is read-only before every query. It cannot write to, alter, or delete anything in your database.

No LLM sees your raw data

Raw rows stay inside the query and computation layer. No LLM agent gets access to your raw data — agents work from your schema, table descriptions, and computed results.

Known egress, no training

Connections come from a small, stable set of published IPs you can allowlist. Your connected data is never used to train foundation models without your explicit consent.

Your stack

Plugs into the data you already have.

Connect PostgreSQL, MySQL, Amazon Redshift, Snowflake, or BigQuery directly. Pull Meta Ads, Google Ads, and GA4 into a managed warehouse with first-party connectors. A single project can span multiple sources — and picoask proposes the cross-source joins for you.

PostgreSQL MySQL Amazon Redshift Snowflake BigQuery Meta Ads Google Ads GA4

See it run on a real question.

See how we compare