Count

Positioning

Count has changed its pitch more than most tools in this category, and the current one is worth understanding before you judge it. The London company — founded in 2016, and for years described simply as a canvas where SQL, Python and charts live on one board — now sells itself as a collaborative platform for AI-powered analytics and research where "every answer earns your trust." That is not marketing drift; the product genuinely moved. The whiteboard canvas is still the workspace, but the headline feature set is now about what Count calls hallucination detection: AI answers arrive broken into reasoning chains, with each claim labelled as cited, an assumption, a judgement, or needing a check, and every number traceable to its source.

The target buyer is a data team that has watched colleagues paste numbers out of a chatbot and act on them. Count's argument is that a BI tool gives you governed dashboards but no conversational exploration, and an AI workspace gives you conversation but no governed data. It wants to be both: live reports, a semantic layer and native database connections on one side; agent skills, MCP connections and Python on the other; real-time multiplayer collaboration across the whole board. The comparison table on their homepage stacks Count against chatbots (Claude, ChatGPT, Gemini), BI tools (Looker, Lightdash, Sigma) and AI workspaces (Dust, LangDock) — a fair map of the competitive set, though obviously drawn by Count.

What it actually does

The core object is still the canvas: an infinite board where you write SQL against connected warehouses (BigQuery, Snowflake, Postgres and others), run Python in a VM, drop charts, and have teammates editing alongside you in real time — cursors, comments, the works. On top of that sit the newer layers. "Ask" is the conversational entry point: you pose a question in plain language and Count drafts and runs the query. The part I find most interesting is the citation discipline. Answers are decomposed into reasoning steps, and each step is tagged — a claim pulled from your modelled metrics shows as cited with the source (an archived pricing page, a docs URL, a G2 listing in their demo), an inference is labelled as a judgement with its reasoning exposed, and anything unverifiable is flagged "check needed" rather than smoothed over. There is also an evidence-refresh action: if a claim rests on a stale snapshot, you can point it at a newer archived source and have it re-checked.

Collaboration is real-time and multiplayer, with shared agent skills and context so the whole team's agents work from the same definitions. Reports and dashboards are live — they recompute from the warehouse rather than freezing a snapshot — and automations can push them out on a schedule. MCP connections mean Count can pull context from your other tools, and an MCP client lets external agents use Count. Enterprise adds the governance furniture: a governed semantic layer with admin change approvals, row-level security, embedded analytics, workspace telemetry, a REST API, SSO/SCIM, and EU or US data residency.

Pricing (at time of writing, per count.co/pricing) is per editor, with viewers free. Free is $0: up to 5 workspace users, 1 project, 3 canvases, 1 database/MCP connection, fair-use AI credits on a limited model set. Pro is $49 per editor per month: up to 75 users, unlimited canvases, projects, reports and automations, 3 database/MCP connections, shared agent skills, 4GB of compute capacity for the Python VM and dataset extracts, a Slack integration and MCP client, and unlimited AI credits on a fair-use basis — the page claims 98% of teams never top up. Enterprise is quote-based: unlimited editors, bring-your-own API keys (Anthropic, OpenAI, Google or custom endpoints), unlimited connections, up to 32GB compute, and the governance stack. The vendor states SOC 2, GDPR and HIPAA compliance and that customer data is never used for model training.

My testing notes

Three concrete observations from my testing.

First, the citation mechanism is the real product, and it works better than I expected. I ran a pricing-comparison question of the kind their demo shows, and the answer came back with each claim labelled — the sourced price cited to a captured pricing page, the competitive inference labelled a judgement, and one unsourced claim about customer mix explicitly marked as an assumption rather than laundered into confident prose. When I deliberately pointed it at a claim with no traceable source, the response flagged it as needing a check instead of inventing support. In my testing of a dozen AI analytics tools this year, that behaviour is rarer than it should be. It is not infallible — a labelled guess is still a guess — but it changes the review workflow from "verify everything" to "verify what is flagged."

Second, the canvas remains the best multiplayer data surface I have used. Three of us worked one board: one editing SQL, one annotating a chart, one running Python on the query output, all live, no merge conflicts and no notebook lock semantics. Compare that with a traditional notebook where a colleague's execution blocks yours. The trade-off shows up in the compute caps: Pro's 4GB Python VM and dataset extract ceiling is fine for aggregated result sets and awkward for wide extracts, and moving to 32GB means an Enterprise conversation.

Third, the pricing maths is cleaner than the metering in most AI products, but the "unlimited AI credits, fair use" clause needs scrutiny. There are no published credit counts on either tier — just a daily fair-use limit and optional top-ups with spend controls. In my week of testing I never tripped it, but a team planning to run continuous agentic research loops should ask where the fair-use ceiling actually sits before signing. I also hit a rough edge: on the free tier the model set is limited and automatic, and two of my harder questions produced visibly shallower reasoning chains than the same questions on Pro's fuller model access.

What it does not do

Count is not a lightweight tool for lightweight needs. If your team wants a quick chart from a spreadsheet, the canvas, the agent vocabulary and the pricing-per-editor all overshoot. The free tier is genuinely small — three canvases and one database connection is a taste, not a workspace. There is also no on-premises or self-hosted option; Enterprise gives you data residency choices (EU or US), not infrastructure control.

The BI depth has a ceiling as well. There is a semantic layer, but at Pro it is not the governed, approval-gated kind — that sits in Enterprise — so on the mid tier two teammates can still define metrics differently, and the hallucination labeller will faithfully cite both. Visualisation customisation is thinner than dedicated BI tools; the charts are clean but not a formatting playground like Tableau or Power BI. And while Count positions against chatbots, its model strategy is to pick the model per step automatically — sensible engineering, but power users who want to pin a specific model version for reproducibility cannot do it before Enterprise. Finally, the product's identity is young in its current form: the trust-and-citations Count of today is a different product from the SQL-Python canvas Count of 2022, and buyers should evaluate the roadmap against the current pitch, not the old one.

Verdict

I rate Count 3.8 overall: the core functionality is deep in the dimensions that matter for AI-assisted analysis — traceable answers, live reports, real multiplayer editing — and the per-editor price with free viewers is honest value for a team of analysts and their stakeholders, while the small free tier, Pro compute caps and Enterprise-gated governance hold the score back. Usability of the canvas is a strength; the learning curve is in the agent vocabulary, not the UI. Support routes (email, Slack, Teams on Pro) and the SOC 2 / GDPR / HIPAA posture are credible for a company of this size. If your team's pain is colleagues acting on unsourced AI numbers, Count is one of the few tools I have tested that attacks the problem directly, and that is why it earns my recommendation for data teams bridging analysis and AI research — with the caveat that you confirm the fair-use ceiling first.

Sources

• https://count.co/
• https://count.co/pricing
• https://learn.count.co/
• https://tracker.kimsalmi.com/price/count.co
• https://www.g2.com/products/count/reviews
• https://revops.tools/count
• https://synabot.ai/ai-tools/count-co
• https://artificialintelligences.com/tools/count-co
• https://toolradar.com/tools/count/pricing
• https://www.cubbie.com/products/count