The standalone Columnly web app

Your live business data,
finally answerable.

Sign up, connect HubSpot or Pipedrive or upload your files, and just ask. Columnly returns verified dashboards and briefs that stay live, re-sliceable, and refreshed daily — inside a real multi-seat workspace, kept current by Autopilot.

Start free See the trust inspector ↓

Workbench · verified dashboard

Pipeline health — this quarter

Won revenue

Deals

Avg deal size

Lost rate

Click any figure for its trace · toggle a filter to re-verify

Get in · your workspace

A real multi-seat SaaS, not a personal add-on.

Signing up creates your account, a fresh isolated organization, and makes you the Owner — with a real per-org database provisioned before the account is even written, so a workspace is never half-created.

Roles

Owner / Admin / Member

Enforced server-side on every request. Owners hold billing and ownership; members build, finalize and run Autopilot.

Invites

Branded email invites

Single-use, email-locked, 7-day links. One email belongs to exactly one workspace — enforced at send and at accept.

Credits

Seats & a credit pool

A master credit pool with per-seat grants. New teammates auto-get 500 credits; removing a member returns them to the pool.

Lifecycle

Clean ownership transfer

Hand ownership to another member; remove someone and their sessions end instantly and their email frees for re-invite.

Data in

Connect a CRM, or upload your files. One entry point.

Everything starts from a single "Add source." Connectors are read-only forever — a hard guarantee enforced by a positive scope allowlist at the provider edge. Columnly literally cannot write back to your CRM.

Connect a CRM

HubSpot & Pipedrive

HubSpot: contacts, companies, deals, tickets, line items, products, custom objects, owners, pipeline stages — and the real associations.

Pipedrive: Deals, Persons, Organizations, Activities, Products, Leads — full custom fields and the links between them.

Both at once, namespaced so their overlapping "deals" and "products" never collide.

Upload files

CSV / XLSX

Each honest worksheet becomes a queryable source. Empty tabs, pivots and legend sheets are skipped, not mis-imported.

Honest type inference: a genuinely dirty column is kept verbatim as unknown, never silently coerced. Built for 500,000 rows × 50 columns, streamed and stored as columnar Parquet in one verified commit.

It asks before it touches your data

Detect → Propose → Confirm → Clean

01 Detect

It names what's wrong and quotes your own value back — '1,000+', '58.6M', a 'Varies with device' sentinel.

02 Propose

A plain yes/no: "Clean it up" or "Use as-is." That sentinel becomes null, never 0.

03 Confirm

Accepted cleaning is verified before it's applied — if it can't be verified, it's refused, not shipped.

04 Clean

Frozen as a replayable recipe, so tomorrow's refresh re-applies the exact same decision.

Stays fresh automatically. Daily incremental refresh at 6 AM in your workspace's timezone. Failures retry, self-heal, then surface as an amber connection-health warning — a broken pipeline is loud, never silent stale data.

amber > silence

Ask

A plain-English question becomes a verified answer.

Every data question is solved twice — one model decomposes and computes it, an independent model re-derives it from scratch — and ships only when they agree. Before that, it decides which sources are actually relevant, and it discloses anything it had to leave out.

Relevance

It picks the right data

It decides which connected sources answer this question, anchors on one, and states a one-sentence reason for anything excluded.

Premise check

Right data, not just right math

A third check verifies the sources themselves were chosen correctly — and only pauses to ask when getting it wrong would change the answer.

Intent

Shaped to your question

Aggregate → a headline figure + method. Breakdown → a table rendered deterministically from computed data. A task → a confirmation.

The moat · click-to-trace

Every figure carries its own proof.

On any dashboard or brief, every number is clickable. Click it and an anchored card shows Source → Computation → Reconciliation. Try it — these are live:

This quarter you closed $4.82M in won revenue across 96 deals, at an average deal size of $50,208. Win rate held at 31% — and Columnly flags that last one, because the open deals can't honestly be counted as either yet.

The green underline means proven before you even click. The amber dashed means honestly flagged — never a silent wrong number.

Verdict at rest

See it before you click

A verified figure wears a green underline; an unreconciled one wears amber — you can see which numbers are rock-solid without opening anything.

Deterministic proofs

Checked before it ships

Before any artifact ships, automated verifiers confirm every figure cites a computed value and matches it, every chart's bars equal the stored series, and "highest/lowest" claims can't contradict their own chart.

Real choice

Refuse, don't fabricate

If an insight can't be backed, it's dropped from the artifact — not the whole board — and you get three options, any alternative itself re-verified before it's offered.

The full mechanism — the results store, server-owned figures, deterministic verification — is on the trust page →

The output · one engine, two shapes

Dashboards you monitor. Briefs you read once.

An AI composes which components to show from a library of ~23 — but it never states a number; a deterministic engine computes and verifies every value. So the composition is creative and the numbers stay bulletproof.

app.columnly.ai / dashboards / order-health

Demo-able

Verified re-sliceable filters

Filter the board and every figure recomputes and re-verifies over the slice — live, in place, no reload. Try it in the console at the top of this page.

Analyst reasoning

It leads with the meaningful metric

From a raw timestamp it derives time-of-day and day-of-week. It leads a scorecard with "delivered revenue," not a raw revenue that mixes booked and cancelled.

Taste

No two look alike

A second model adversarially grades every composition for weak stories and template-flatness, forcing a revision when it's weak.

Reliability

Autopilot keeps it current — and refuses to publish a wrong number.

Put a finalized analysis on Autopilot and it re-runs on a schedule against fresh data, no human in the loop. When you finalize, the analysis is snapshotted as an immutable record — frozen recipes, locked lineage — so it can't drift as rows change.

Never advances on a failure

Pipeline broke → loud "needs attention," last-good held. A soft figure → quiet amber, still ships. A lie → blocked entirely.

Smart & cheap

If the data hasn't actually changed, the run skips. Missed schedules collapse to a single catch-up run.

Only pinned items refresh

You pin the answers, figures and tables you care about; chat is never replayed — only the frozen computation re-runs.

Collaboration & lifecycle

Share it, keep it — and never let a number silently vanish.

Share within your team

Grant a teammate viewer or editor access to a specific analysis; revoke anytime. Sharing is member-to-member inside your org — there are no public, unauthenticated links. That's a deliberate trust posture.

Degrade visibly, never destroy

Disconnect a CRM and you choose: reconnect, retain a frozen snapshot with a stale banner, or remove the data — and even then artifacts survive as tombstones showing last-known figures. No disconnect choice ever makes a number silently disappear.

Freshness is always on-screen

Sources show "Fresh" vs "static," an as-of date, and mixed-freshness is disclosed once. Snapshots are versioned with 15-day retention and atomic swaps — a reader never sees a half-built refresh.

An analyst's working notebook

Each saved analysis holds multiple named chat threads, pins, and an inspector — column stats on the left, threads in the center, the trace card on the right.

Your live business data,finally answerable.