Disclosure: This article is published by Upsolve AI. Where our product appears alongside others, the comparison is based on publicly available information as of November 2026.

ThoughtSpot Spotter is ThoughtSpot’s AI analyst: you ask a question in plain English and it returns a chart, a table or a written analysis from your warehouse, with the logic shown. Two things catch buyers out. “Explainable” now covers three different execution paths, each leaving a different trail to check. And on the per-user Pro plan, the allowance is 25 Spotter queries per user per month, where a question with two follow-ups counts as three.

What Is ThoughtSpot Spotter?

ThoughtSpot Spotter is a conversational analytics agent built on ThoughtSpot’s semantic layer. It launched in November 2024 and now appears in Slack, Teams, mobile, Liveboards, Claude and ChatGPT, or inside your own product through the embed SDK.

Spotter is also the name of a family. SpotterModel builds data models, SpotterViz builds Liveboards and SpotterCode helps developers embed. Plans that include Spotter include all three at no extra charge, except row-based and unlimited plans, where they’re bought separately.

It no longer has to sit only on a ThoughtSpot Model. ThoughtSpot says it can build on Snowflake semantic views or dbt semantic models as well.

Spotter 1, 2 and 3 handle your data differently

“Spotter” means three experiences, and an admin picks the default. Any description of Spotter as metadata-only is describing Classic.

Spotter 3’s safeguards are real: row and column security apply first, ThoughtSpot’s LLMs don’t retain or train on the data, and you can bring your own LLM. Two details are easy to miss. Users need the Can download data privilege. And to keep a column from the LLM you hide it in the model, which also hides it from standard Search.

What happened to ThoughtSpot Sage?

ThoughtSpot Sage was the earlier natural-language search, announced on March 9, 2023 on GPT-3. ThoughtSpot’s docs now state that Spotter replaces Sage and Ask Sage. Sage clusters are upgraded automatically to Spotter 2, Embed SDK users need to update, and API users move to the Spotter APIs.

The catch: a former Sage customer lands on Spotter 2, while cross-model search, Deep analysis and connectors are Spotter 3 features. Check which version you’re on before judging the product by its marketing page.

How Does Spotter Provide Explainable Answers?

Spotter makes answers explainable by showing its logic as something you can read, and in most cases by compiling governed SQL rather than letting a model write it freely. Spotter chooses one of three paths for each question.

• Search tokens, for simpler questions. The question becomes measures, attributes and filters shown with the answer, and you can click a token to change it. This is the strongest part of Spotter’s design, since a sales manager can spot a wrong filter without reading SQL.

• AgentQL, for multi-step questions. AgentQL is a proprietary semantic query dialect for ranking, running totals and nested logic. It’s readable in the Show work panel, though that suits analysts more than business users. It has been rolling out since 26.8.0.cl.

• Python, for statistics and stitching. Forecasting, regression and answers combined from several models use LLM-written Python. ThoughtSpot’s AgentQL page treats this as the harder path to inspect, notes a row limit, and reports that about 20% of Spotter conversations include a Python step.

Deterministic doesn’t mean correct

ThoughtSpot’s headline promise is the same answer on every run, and the mechanism is real: tokens and AgentQL compile to identical governed SQL each time. The launch post put the condition plainly: “If relational search tokens are correct, SQL is 100% accurate.”

That “if” carries the weight. Reading your question into tokens is the LLM’s job, and memory and coaching are designed to change that reading over time.

Determinism tells you the same tokens produce the same SQL. It doesn’t tell you the tokens were the right reading of the question.

ThoughtSpot’s BIRD benchmark post shows the size of that gap. Spotter scored 82.8% across 1,533 questions, ahead of a frontier LLM at 78.4% and a coding agent at 72.5%. On the challenging band it scored 70.6%. So on ThoughtSpot’s own exam, about one answer in six was wrong overall, and about three in ten on hard questions.

The post publishes its method and confidence intervals, which is more than a bare percentage. But it’s vendor-run: competitors appear as categories, and it flags annotation problems in BIRD’s answer key. Also, Spotter’s product page says it beats frontier models by 24 points. The closest figure in the post is the 23.8-point gap to the coding agent on that band. Against the frontier LLM, the post shows 4.4 points overall and 13.9 on challenging.

None of this makes Spotter weak. Its trail is how you catch the wrong answers, and that only works if someone reads it.

Where the trail gets thinner

• Fast mode: everyday questions can take a shorter route, answering in under 15 seconds, and Spotter shows less of its working.

• Lost context: in Spotter 3, drilling down, editing an answer or applying a new filter means Spotter doesn’t recall that context for follow-ups.

• AgentQL output: the docs list no chart in the answer card, no pinning to Liveboards and no parameters in Phase 1.

• Audit history: events in the Spotter Conversations Liveboard are kept for 180 days on a rolling basis since July 2026. If you need a longer record, export it first.

How ThoughtSpot’s Pricing Model Works for Spotter

ThoughtSpot sells two products, Analytics and Embedded, and Analytics can be priced per user or by usage. These are list prices as of November 2026, and every figure shown is a lowest starting price.

ThoughtSpot doesn’t meter LLM tokens, and warehouse compute stays a separate bill. For the wider picture, see our ThoughtSpot pricing breakdown.

What counts as a Spotter query

The rule is in the consumption docs, not on the pricing page:

• Every natural-language question or follow-up counts. One question and two follow-ups is three queries.

• Clicked drill-downs don’t count as Spotter queries. On per-user plans, that makes clicking cheaper than typing. On credits, a drill-down is still a data query.

• Text-only replies count too. Asking Spotter to create a Liveboard counts one query per query on the board.

• On credits, data queries come on top. A question that generates several ThoughtSpot queries is billed for those as well as the Spotter query.

So 25 queries is about eight three-turn conversations a month. The pricing page doesn’t say whether the allowance is pooled or what happens at query 26. Ask before you sign.

A worked example: 200 users

Assumptions (ours, not ThoughtSpot’s): 200 users, lowest list prices, annual terms. Each user asks 25 Spotter questions a month and opens 20 Liveboards with eight charts in view, at one credit per query and one query per chart on screen.

• The crossover is 500 credits. $50 ÷ $0.10 means usage is cheaper at list until the average user generates 500 credits a month.

• Spotter is the small part of a usage bill, about 14% here. Dashboards are the real driver.

• A fully used seat costs $2 per Spotter query, though it also buys dashboards. Credit packages carry a one-year term, and the minimum isn’t on the pricing page.

Three smaller traps

• Rows count by availability. You license the rows made available for querying across all connections, so one large event table can use up Pro’s 250M.

• The MCP Server is an Enterprise add-on. The pricing table shows it only there. MCP connectors, which pull context in, are a different thing and are listed on Pro. Check before planning on Claude or ChatGPT access from Pro.

• Analyst Studio, the SQL, R and Python workspace, is an add-on on per-user plans.

Evaluating ThoughtSpot on SQL, Conversational and Agentic Analytics

SQL. Spotter doesn’t ask the model to write free-form SQL on the token and AgentQL paths. A compiler generates it and runs it live against your warehouse, inheriting joins, security and calendars. AgentQL covers per-group ranking and running totals, where tokens struggled. Spotter connects to 30+ sources including Snowflake, Databricks and BigQuery with no data copy.

Conversational analytics. This is Spotter’s home ground. You can start from a chart on a Liveboard, and token editing suits non-technical users. For Spotter 3, the documented limits are narrow: no chart configuration through natural language, and no parameter support.

Agentic analytics. Spotter 3 plans, checks and corrects its own work, and Deep analysis proposes a plan for you to approve, then runs a chain of queries into a report. Connectors to Slack, Jira and Asana can push actions out, but the capabilities page lists them as Early Access. Treat the analysis as the strong part and the action side as young.

Setup. ThoughtSpot’s FAQ says most of Spotter’s success depends on the data model, and recommends coaching on 20 to 50 questions before a pilot. Claire Gouze, Head of Data at nao, which makes its own tool, benchmarked it in February 2026 and found it powerful but judged the setup and lock-in excessive. Recheck that against SpotterModel, which ThoughtSpot now offers to guide model building.

Lock-in. Definitions can come from Snowflake or dbt, and ThoughtSpot says its semantics plug into other agents through MCP and Apache Ossie. Instructions, coaching and memory are still ThoughtSpot features, so check how they export before calling them portable.

When a Dedicated Data Agent Fits Better

Spotter is enough when your people already work in Liveboards, your data is modeled or you’ll fund that work, and the person checking an answer is a business user. Clickable tokens beat a SQL trace for that person.

A dedicated data agent such as Upsolve AI fits better in four cases:

• You don’t want BI seats to get an agent. Upsolve is priced by credits rather than per user, with no semantic layer needed to start. It works across Snowflake, BigQuery, Databricks, Redshift and Postgres at the same time.

• You’re embedding for customers on a published price. Team includes embedding and multi-tenant support at $2,000 a month, or $1,600 billed annually. In ThoughtSpot Analytics, multi-tenant Orgs are Enterprise only, the usage-limits page says to contact support at around 400 Orgs, and Embedded Enterprise is custom-quoted.

• Your verification model is tests, not coaching. Upsolve’s agent writes SQL and checks it against the database before running it. On top of that, each test pairs a question with reference SQL, and a CI switch can stop a failing version reaching production (the docs label Tests as Beta). Every answer has a step-by-step trace with LLM-graded evaluations, which suits a data team that wants regression testing for context.

• You want the agent wherever people ask. Any MCP client can connect, with the MCP App on every plan, alongside Slack and Teams.

You’d still be trusting one vendor with that context, but it isn’t tied to a BI platform. Upsolve’s row-level security and RBAC start at Team, and SAML SSO and HIPAA are Enterprise, where ThoughtSpot lists SAML on Pro. For other options, see our ThoughtSpot alternatives guide and the wider AI analytics tools roundup.

ThoughtSpot Spotter vs Upsolve AI: Comparison Table

As of November 2026, from ThoughtSpot’s pricing page and Upsolve’s pricing page.

Frequently Asked Questions

What is ThoughtSpot Spotter?

It’s ThoughtSpot’s AI analyst, a conversational agent that answers plain-English questions from your warehouse with charts, tables and written analysis. Spotter 3 adds forecasting, multi-model search and Deep analysis.

Is ThoughtSpot Sage still available?

No. ThoughtSpot’s release notes say Sage reached end of life, with removal in 26.6.0.cl and clusters moved to Spotter 2. The consumption docs still mention Sage indexing and the Sage search bar, so the name lingers.

How much does ThoughtSpot Spotter cost?

Spotter isn’t sold alone. It starts with Pro at $50 per user per month billed annually, or $0.10 per credit. Enterprise is custom. The per-user plan includes 25 Spotter queries per user per month.

Does ThoughtSpot Spotter generate SQL?

Yes, but mostly by compiling it. On the token and AgentQL paths, ThoughtSpot’s query layer generates the SQL deterministically from a structured query. Forecasting and multi-model stitching use Python.

Is ThoughtSpot Spotter good for customer-facing analytics?

It can be. You can embed it through ThoughtSpot’s web API and SDK, and row-level security applies. Check how Spotter is licensed on your Embedded plan and how your tenant count fits ThoughtSpot’s limits.

If you’re weighing Spotter against a dedicated data agent, book an Upsolve demo or compare the plans.

Sources

1. ThoughtSpot: Plans and pricing, including the 25-query limit, credit price and add-ons. https://www.thoughtspot.com/pricing

2. ThoughtSpot Docs: User actions and queries, including how Spotter queries are counted. https://docs.thoughtspot.com/cloud/26.10.0.cl/consumption-pricing-query-based-generate.html

3. ThoughtSpot Docs: Query-based usage: credits and base-credit packages. https://docs.thoughtspot.com/cloud/26.10.0.cl/consumption-pricing-query-based.html

4. ThoughtSpot Docs: Release 26.4.0.cl notes on Sage removal. https://docs.thoughtspot.com/cloud/26.10.0.cl/26-4-0-cl.html

5. ThoughtSpot Docs: Deprecated and removed features. https://docs.thoughtspot.com/cloud/26.10.0.cl/deprecation.html

6. ThoughtSpot: Press release launching Sage, March 9, 2023. https://www.thoughtspot.com/press-releases/thoughtspot-integrates-with-gpt-3-to-launch-sage-a-new-ai-powered-search-experience-for-analytics

7. ThoughtSpot Docs: Spotter versions and data sharing with LLMs. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter-versions.html

8. ThoughtSpot Docs: Spotter overview, experiences and licensing. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter.html

9. ThoughtSpot Docs: Spotter 3 data handling, controls and retention. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter-data-handling.html

10. ThoughtSpot Docs: AgentQL: execution paths, verification and limits. https://docs.thoughtspot.com/cloud/26.10.0.cl/agentql.html

11. ThoughtSpot Docs: Getting started with Spotter: tokens, CSV upload, fast mode. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter-getting-started.html

12. ThoughtSpot Docs: Spotter limitations. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter-limitations.html

13. ThoughtSpot Docs: Spotter capabilities, including Connectors Early Access. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter-capabilities.html

14. ThoughtSpot Docs: Spotter FAQ on data models and coaching. https://docs.thoughtspot.com/cloud/26.10.0.cl/spotter-faq.html

15. ThoughtSpot Docs: Usage limits by edition, and how rows are counted. https://docs.thoughtspot.com/cloud/26.10.0.cl/editions-usage-limits.html

16. ThoughtSpot: Spotter product page. https://www.thoughtspot.com/product/agents/spotter

17. ThoughtSpot: Spotter Semantics. https://www.thoughtspot.com/product/spotter-semantics

18. ThoughtSpot: Introducing Spotter, launch post. https://www.thoughtspot.com/blog/introducing-spotter-ai-analyst

19. ThoughtSpot: AI Analyst Accuracy: The BIRD Benchmark Results (vendor-run). https://www.thoughtspot.com/blog/ai-analyst-accuracy-bird-benchmark

20. Claire Gouze, The New AI Order: I tested 21 analytics agents, February 6, 2026. https://thenewaiorder.substack.com/p/i-tested-21-analytics-agents-and

21. Upsolve AI: Pricing. https://upsolve.ai/pricing

22. Upsolve AI: Homepage and product overview. https://upsolve.ai/

23. Upsolve Docs: Tests. https://docs.upsolve.ai/ai-agent-builder/tests

24. Upsolve Docs: Agent Observability. https://docs.upsolve.ai/ai-agent-builder/agent-observability

25. Upsolve Docs: Use Upsolve from Claude and other AI assistants. https://docs.upsolve.ai/ai-agent-builder/deploy-agents/mcp-server

26. Upsolve Docs: Upsolve for Slack. https://docs.upsolve.ai/slack/overview