BI Copilot

Vital signs

Why it's on the table

On the table, BI Copilot (Bi) is seat 46 of 58, in the Data & Analytics family. It is an experimental element — promising, volatile, and worth a contained experiment rather than a commitment. Budget curiosity, not dependence. It is optional: plenty of companies run without it — until a specific trigger (scale, regulation, cost, or customers) makes it essential for them. It sits in the mid price band — a real line item that should earn its keep visibly.

BI Copilot: the top 5 — v2026.Q3

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

1. 1OmniOmni Analytics (Colin Zima, ex-Looker)

   No public price list — sales-quoted annual contracts (verified Aug 2026). AI chat, modeling agent and MCP server included; LLM tokens not separately metered.

   Best for Startups that want Looker-grade governed metrics without Looker's cost or rebuild tax, with conversational access bolted to the same model.

   Built by the people who built Looker, on the correct premise for this category: the chat is only as good as the semantic model, so Omni ships a modeling agent to build the model and an AI chat that queries through it, plus an MCP server so Claude, ChatGPT and Cursor hit the same governed logic. ICONIQ led a $120M Series C at $1.5B on Apr 23, 2026 against 4x YoY revenue — up from a $650M mark in March 2025 — with BambooHR (100,000+ end users), Checkr and Cribl as reference accounts.

   Watch Zero published pricing means you cannot budget without a sales call, and the entry point is not seed-stage cheap. The modeling agent shortens the semantic-layer build but does not remove it — you are still doing the actual work, just faster. No independent accuracy benchmark for Omni's AI exists as of Aug 2026.

   $120M Series C at $1.5B led by ICONIQ, plus $30M employee tender; 4x YoY revenue growth (Apr 23, 2026) [src] · MCP server exposes the governed semantic model to Claude, ChatGPT and Cursor; agent skills for model building, querying and content management (Aug 2026) [src] · Valuation up from $650M (Mar 2025) to $1.5B (Apr 2026) [src]
   Site →Docs →

2. 2HexHex Technologies

   Community free (5 projects) · Professional $36/editor/mo · Team $75/editor/mo · Enterprise custom. Per-seat AI credit grants on paid plans; compute add-ons $0.32–$4.06/hr. BYO API key Enterprise only.

   Best for Teams with at least one analyst — the Notebook Agent accelerates them, Threads gives everyone else a governed chat surface on the same models.

   Hex is the rare vendor shipping both halves: a Notebook Agent that writes and debugs code with dependency awareness, and Threads (Oct 1, 2025) — a step-by-step reasoning chat for non-technical users, reachable from Slack, Claude and Cursor — plus a Modeling Agent for the semantic layer underneath. It scored 80.6% on BI Bench, third overall and the highest of any purpose-built BI product tested, behind only two general coding agents. Pricing is published, which in this category is itself a differentiator.

   Watch Threads and the semantic-model agent require the $75/editor Team plan; the free and Professional tiers get the notebook agent only. Hex's centre of gravity is still the analyst notebook — element Wh territory — so buying it purely as a business-user chat box overpays. Last confirmed raise is the $70M Series C (May 2025); 2026 round listings on aggregator sites are unverified.

   80.6% accuracy on BI Bench, 3rd of 11 and top-scoring BI product (vendor-run benchmark, published 2026) [src] · Threads launched Oct 1, 2025 — agentic chat for non-technical users, Team and Enterprise plans [src] · Published pricing: Professional $36/editor/mo, Team $75/editor/mo (Aug 2026) [src]
   Site →Docs →

3. 3LightdashLightdash (Apache-2.0)

   Open source free (self-hosted) · Cloud Pro $3,000/mo flat, unlimited users, AI agents included · Enterprise custom · embedding $0.05/load after 1,000 free or $790/mo for 100k loads

   Best for Teams whose metrics already live in dbt YAML and who refuse to pay per seat to let the company read its own numbers.

   The dbt-native path: metrics and dimensions are defined once in your dbt project and the AI agents query that, with Slack delivery, memory that learns from user corrections, and — uniquely among the top five — built-in evaluations to regression-test agent accuracy as your model changes. Apache-2.0 and self-hostable (5.9k GitHub stars, Aug 2026), so the escape hatch is real. Flat pricing means adding 40 viewers costs nothing.

   Watch $3,000/mo flat is the wrong shape for a five-person startup — you self-host or you pay Looker money. AI agents are Cloud Pro or an Enterprise add-on, not in the open-source build. It scored 23.8% on BI Bench at default settings, second-lowest of eleven — a result that says more about ungroomed dbt models than about Lightdash, but the dependency is the whole point of the category.

   Cloud Pro $3,000/mo flat with unlimited users and AI agents included (Aug 2026) [src] · AI agents ship with built-in evaluations to validate accuracy; Data MCP available on all Cloud tiers [src] · Apache-2.0, 5.9k GitHub stars, 729 forks, actively developed (Aug 2026) [src]
   Site →Docs →

4. 4MetabaseMetabase

   OSS free (self-host) · Cloud Free $0 · Starter $100/mo (5 users, +$6/user) · Pro $575/mo (10 users, +$12/user) · Enterprise from ~$20k/yr. AI on every tier; $3.75 per 1M tokens if you don't BYO key (1M free).

   Best for Pre-Series-A teams who need the answer today and cannot justify a four-figure monthly BI line item.

   The only tool here where natural-language questions and SQL generation are on the free tier, and the 2026 releases are serious: Metabot, an MCP server, Slack integration, Data Studio for semantic-layer curation, and a dependency graph so you can verify model integrity before the AI leans on it. There is also a master kill switch for all AI — a governance feature the enterprise vendors mostly lack.

   Watch It finished last on BI Bench at 12.4%, and Metabase's own docs are candid about why: accuracy depends on field descriptions, correct semantic types and a glossary you have to write. Out of the box, pointed at raw warehouse tables, this is the tool most likely to confidently answer the wrong question. Verified-content-only mode for Metabot is Pro/Enterprise.

   AI question-asking and SQL generation available from the $0 tier upward; AI service $3.75/1M tokens or BYO key (Aug 2026) [src] · 2026 shipping: Metabot, MCP server, Slack, Data Studio semantic layer, dependency graph and diagnostics (Jun 9, 2026) [src] · 12.4% on BI Bench — lowest of 11 tools at default settings (vendor-run benchmark, 2026) [src]
   Site →Docs →

5. 5ThoughtSpot SpotterThoughtSpot

   Essentials from $25/user/mo annual (5–50 users, 25M rows) · Pro from $50/user/mo (to 1,000 users, 250M rows) · Enterprise & Embedded custom · Developer free 1 yr (10 users). LLM tokens explicitly not metered.

   Best for Rolling a search box out to hundreds of non-technical staff who will never open a BI tool, with governance and unmetered token cost.

   ThoughtSpot invented search-first BI and is the only pure-play NLQ vendor named a Leader in the Jun 29, 2026 Gartner Analytics & BI Magic Quadrant, alongside Microsoft, Salesforce Tableau, Google Looker and Qlik. The 2026 agent line went beyond chat: SpotterViz builds dashboards from language, SpotterModel builds semantic models without code, SpotterCode generates embedding code — plus industry-tuned Spotter agents. Not metering LLM tokens matters when 400 people start asking questions.

   Watch The $25 list price is the smallest part of the bill — this is an enterprise sales motion with modeling services attached, and analysts flag semantic modeling as the bottleneck the agents only partially remove. No independent accuracy data exists, no ARR is disclosed, and the persistent IPO-or-sale speculation means the strategic picture in 12 months is genuinely unknown.

   Leader in the 2026 Gartner Magic Quadrant for Analytics and BI Platforms (published Jun 29, 2026) [src] · SpotterViz, SpotterModel and SpotterCode announced Dec 10, 2025, GA early 2026 [src] · Essentials from $25/user/mo; "ThoughtSpot does not meter or charge for LLM tokens" (Aug 2026) [src]
   Site →Docs →

BI Copilot: the top 8 compared

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

How to choose your bi copilot

If you are already paying for Databricks, Microsoft Fabric, Looker or Snowflake

Use the copilot you own before buying a fifth tool. 1.5M Genie Spaces were created in 2026 alone; Fabric Copilot dropped to F2 capacity (~$262/mo) with no per-user AI fee. Buy a standalone BI copilot only after the built-in one demonstrably fails.

If you have no semantic layer and no one to build one

Do not buy conversational BI yet — buy the modeling. Standalone semantic-layer adoption tripled from ~8% to ~28% among data teams in six months (Hex, Dec 2025) precisely because chat without a model produces confident wrong answers. Omni's or ThoughtSpot's modeling agent shortens that job; nothing skips it.

If your metrics already live in dbt and you have more viewers than analysts

Lightdash — flat pricing, agents that query your existing dbt YAML, and built-in evals to catch regressions. Self-host free if $3,000/mo is out of range.

If you have analysts and the bottleneck is their throughput, not executive self-service

Hex at $36–75/editor. It out-scored every other BI product on the one public head-to-head accuracy test, and the notebook agent compounds with the people you already employ.

If an executive will make a real decision on the AI's answer

Require citation-to-source and verified-content-only modes before rollout. Only 51% of 114 data leaders trust AI-generated insights and 26% have already been burned by an inaccurate one (insightsoftware, Jun 2026); audit trails linking output to source data were the single most-requested fix (53%).

BI Copilot: the whole field

30 more tools tracked in this category, including 3 dead, renamed, or sunsetting — a reference that hides the graveyard isn't one. Verified 2026-08-06.

BI Copilot: the category in numbers

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

• BI market $41.16B (2026) → $62.38B by 2031, 8.67% CAGR; cloud segment 9.54% — Mordor frames agentic AI as 'widening BI scope from dashboarding into AI-assisted execution, creating whitespace around governed semantic layers' [src]
• Omni: $120M Series C at $1.5B led by ICONIQ, 4x YoY revenue (Apr 23, 2026) — up from a $650M valuation in Mar 2025 [src]
• Sigma: $200M ARR, 2,000+ customers, revenue doubled YoY and 1.1M new active users added (Apr 13, 2026); $80M Series E at $3B with Databricks, ServiceNow and Workday Ventures participating (May 18, 2026) [src]
• 1.5 million Databricks Genie Spaces created in 2026 alone — the built-in copilots, not the startups, own the volume [src]
• Fivetran + dbt Labs merger completed Jun 1, 2026, spanning 100,000+ data teams, with 'Agents Schema' — an open standard putting metric definitions, semantic models and lineage in plain warehouse tables as shared agent context [src]
• Trust gap: only 51% of 114 data & analytics leaders are confident in AI-generated insights, a third worry about hallucinations, and 26% have already suffered consequences from an inaccurate AI output (insightsoftware, Jun 9, 2026) [src]

BI Copilot: method & sources

Ranking criteria, in order: (1) does the product treat the semantic model as a first-class artifact rather than an afterthought; (2) evidence of answer quality, with independent evidence weighted above vendor claims; (3) whether a sub-50-person company can actually buy it; (4) escape hatches — open source, MCP, published pricing; (5) commercial durability. No affiliate consideration; picks are editorial. Benchmark conflicts, resolved: BI Bench (basedash.com/bi-bench) is the only public head-to-head test of shipping conversational-BI products, covering 11 tools on one production-grade schema — but it is run and published by Basedash, which ranks itself #1 at 92.1%, and carries no publication date. We cite it because nothing else exists, and we flag every use. Its 'default settings' methodology systematically penalizes tools whose accuracy is meant to come from a curated semantic model (Lightdash 23.8%, Snowflake Cortex 19.2%, Metabase 12.4%), so read it as out-of-the-box behaviour, not ceiling. The academic alternative, Spider 2.0, measures research agents rather than products: GPT-4o scores 86.6% on Spider 1.0 and 10.1% on Spider 2.0's enterprise schemas (1,000+ columns), and the current Spider 2.0-Lite leader sits at 76.23% — the single best quantification of why demo-quality text-to-SQL collapses on a real warehouse. Pricing caveats: Omni, Zenlytic, Sigma, TextQL and ThoughtSpot's real-world enterprise tiers are all sales-quoted; only Hex, Lightdash, Metabase, Basedash and ThoughtSpot's entry SKU publish numbers. Anyone quoting an Omni per-seat price in Aug 2026 is estimating. Adoption-stat caveat: the widely repeated '29% of employees actually use BI tools' figure traces to BARC's Nov–Dec 2021 survey of 214 data leaders, which put it at 25% and noted no growth over seven years. 2026 restatements attribute it to Gartner without a public citation. The dashboard graveyard is real, but treat the exact number as directional. Funding caveat: Hex's last confirmed raise is the $70M Series C (May 2025). Aggregator profiles listing 2026 rounds could not be verified against a primary source. Adjacent elements: warehouses, DuckDB/MotherDuck and the notebook layer are element Wh — Hex straddles both and is ranked here for its conversational agents, not its notebook. Product and behavioural analytics (PostHog, Amplitude) are element An. Agent evals and LLM observability, including evaluating your BI agent's answers over time, are element Ev. Data movement and transformation (Fivetran, dbt) sit upstream of this element entirely. Ranking criteria: verified commercial traction, independent satisfaction surveys, agent benchmarks, and founder-fit (price floor, lock-in, surfaces). Editorial, never paid — the charter. Machine-readable twin: bi.json.

Our take

The dashboard graveyard problem is real. Conversational BI works when your data model is clean — that's the actual work.

Combines with