Conversational Analytics Tools: AI Analytics Software Pricing Compared Per Question
Every vendor meters this differently: tokens, messages, DBUs, capacity hours, included queries. Ask your own database a question below and see the SQL it writes.
PostgreSQL · MySQL · Snowflake · BigQuery · read-only · Last updated August 2026
Ask your data a question:
Writing SQL… Running (read‑only)… SQL Agentsql wrote
▋
Click a question. Agentsql writes the SQL, runs it read-only, and answers.
Direct answer
Conversational analytics means asking your data a question in plain English and getting an answer back, instead of building or reading a report. Be careful with the term: in the contact-center market the same words mean mining recorded customer calls for sentiment, which is a different product entirely. This page covers the data one. The part buyers get wrong is cost. Every platform now ships the feature, and every one bills it on a different unit: Google charges per data token, Snowflake per message, Databricks per DBU, Microsoft per Fabric capacity hour, ThoughtSpot per included query, AWS per account plus upgraded seats. None of those units is a question, so no two quotes are comparable and none can be forecast from headcount. Snowflake is the only vendor that meters the question itself, at 67 Platform Credits per 1,000 messages, which works out at 13 to 27 cents a question depending on your edition.
›_ conversational analytics pricing, vendor by vendor
Seven platforms, seven different meters, and not one of them counts questions.
Every rate read from the vendor's own pricing document or documentation on 30 August 2026.
Conversational analytics arrived as a feature of platforms people already owned, which is why its pricing is such a mess. Nobody designed a market rate. Each vendor bolted the AI onto whatever meter its billing system already ran, so the same capability shows up as tokens in one contract, messages in another and capacity hours in a third. That is fine until you try to compare two proposals, at which point there is no common denominator at all.
| Platform | Feature | Metered in | Published rate | What you pay before question one |
|---|---|---|---|---|
| Google Cloud | Conversational Analytics agent (Looker, BigQuery) | Data tokens | $3.00 per 1M input, $20.00 per 1M output | None published, but free only through 30 Sep 2026 |
| Snowflake | Cortex Analyst, called through the Analyst API | Messages | 67 Platform Credits per 1,000 messages | Warehouse compute to run the SQL, billed separately |
| Snowflake | Cortex Analyst, called through Cortex Agents or CoWork | Tokens | AI Credits per 1M tokens, varies by model | Same feature, different meter |
| Databricks | AI/BI Genie | DBUs | $0.070 per DBU, 150 free DBUs per user per month | Genie needs SQL Pro or Serverless compute underneath |
| Microsoft | Copilot in Power BI and Fabric | Fabric capacity CU-hours | $0.18 per CU-hour in westus2 | An F2 or P1 capacity, $262.80 a month at F2 |
| ThoughtSpot | Spotter | Queries, then seats | $50 per user per month, 25 Spotter queries included | The seat itself, whether or not it is used |
| AWS | Amazon Q in QuickSight | Account fee plus upgraded seats | Reader Pro $20, Author Pro $40 per user per month | $250 a month per account, triggered by any Pro user |
| Grafana | Grafana Assistant | Active AI users | $20.00 per active AI user, 40M tokens, $2.00 per 1M over | The $19 Pro platform fee, and tokens are exempt from discounts |
Sources, all read on 30 August 2026: Snowflake's Credit Consumption Table PDF, marked effective 26 August 2026; Google Cloud's Data Cloud agents pricing page; Microsoft Learn's Fabric Copilot capacity article; the Databricks AI/BI Genie cost documentation; AWS QuickSight pricing. The Grafana Assistant and ThoughtSpot Spotter figures were verified on 20 and 29 August 2026 respectively and are unchanged. Full workings on our Snowflake pricing, Databricks pricing, Looker pricing and ThoughtSpot pricing pages.
›_ what one question actually costs
Convert every meter into the same unit and the spread is enormous.
Arithmetic on the rates in the table above. No estimates, no modeled assumptions.
Snowflake is the easiest to price because it is the only vendor that bills the question itself. Its Credit Consumption Table lists Cortex Analyst at 67 Platform Credits per 1,000 messages, which is 0.067 credits a question. Multiply by the credit rate for your edition and you get a real number: 13.4 cents on Standard, 20.1 cents on Enterprise, 26.8 cents on Business Critical. A regulated buyer on Business Critical pays exactly double a Standard customer for the identical answer, and 1,000 questions a month lands at $268 before the generated SQL has run against a warehouse.
Snowflake Standard
$0.134
0.067 credits a message at $2.00 a credit. Exact, because Snowflake is the only vendor that meters the question itself. Running the SQL is extra.
Snowflake Business Critical
$0.268
The same 0.067 credits at $4.00. A regulated buyer pays double for the identical answer, so 1,000 questions is $268 before any query executes.
ThoughtSpot Pro, included allowance
$2.00
$50 a seat divided by the 25 Spotter queries that seat includes each month. That is roughly one question per working day before overage.
Microsoft, before question one
$262.80
The smallest Fabric capacity that can host Copilot, at 2 CU x $0.18 x 730 hours. Charged whether anyone asks anything or not, on top of every $14.00 Pro seat.
The Microsoft figure is the one most likely to catch a team out, because it is a floor rather than a rate. Microsoft's documentation is explicit that a Fabric Copilot capacity "must reside on at least an F2 or P1 SKU", and Copilot usage from Power BI Desktop and from Pro and Premium Per User workspaces all bills to that capacity. So even a team on Pro seats cannot turn Copilot on without a capacity behind it. At the $0.18 per CU-hour rate published in the Azure retail price API, an F2 is 2 x $0.18 x 730 hours, or $262.80 a month, charged whether anyone opens Copilot or not. Two footnotes in the same document are worth knowing before signing: capacities in Embedded license mode are not supported, and a Fabric Copilot capacity does not cover Fabric AI functions.
Three of the eight rows in the table charge something before a single question is asked: Microsoft's capacity minimum, the $250 a month per-account infrastructure fee AWS applies once any Pro user or Q feature is switched on in QuickSight, and ThoughtSpot's per-seat price. The other rows are pure consumption, which sounds safer and behaves differently: consumption pricing has no ceiling. Neither shape is wrong, but they fail in opposite directions, and a pilot rarely runs long enough to show you which one you have bought.
›_ the details that do not fit in a table
What each vendor's fine print actually says.
Snowflake bills the same feature two ways
The Credit Consumption Table effective 26 August 2026 carries a row reading "Cortex Analyst, 67 Platform Credits per 1,000 messages", with a footnote that says this pricing "will only be applicable when using the Cortex Analyst API". Reach the same feature through Cortex Agents or Snowflake CoWork and you land in Table 6(d) instead, priced in AI Credits per million tokens at rates that vary by model. One capability, two meters, and which one you get depends on an integration decision an engineer usually makes without seeing the price list. Snowflake also documents that executing the generated SQL incurs standard virtual warehouse compute charges, so the message price buys the thinking, not the doing.
Google's clock runs out on 30 September 2026
Google's Data Cloud agents page prices the Conversational Analytics agent, alongside the Data Science and Data Engineering agents, at $3 per million input data tokens and $20 per million output data tokens, and states that the free trial of all Data Cloud agents runs "through September 30, 2026, after which the new pricing will become effective". Output costs 6.7 times what input costs, which inverts the usual intuition: verbose answers, not lots of questions, drive that bill. Anyone piloting Conversational Analytics in Looker or BigQuery today is measuring a free product, and the meter starts on 1 October.
Databricks Genie is free for people, billed for robots
Genie One and Genie Agents are free through 31 January 2027, and every user gets 150 free DBUs a month that reset on the first. At the $0.070 Genie DBU rate that allowance is worth about $10.50 a user a month. The line that matters for anyone planning to embed Genie in an application: service principals are excluded from the free allowance. A human clicking around is inside the promotion; an automated integration is billed from its first question. Genie also needs SQL Pro or Serverless compute underneath, so adopting it can quietly move you off a cheaper SQL rate.
ThoughtSpot is the only one that caps questions
Spotter is not in the $25 Essentials tier. It starts on Pro at $50 per user per month, and that seat includes 25 Spotter queries a month, with usage beyond the allowance billed on top. Twenty-five questions is about one per working day, which is a reasonable fit for an executive checking a number and a poor fit for an analyst exploring. Divide the seat by the allowance and the included questions work out at $2.00 each, which is roughly fifteen times the Snowflake Standard rate for a comparable natural-language query, though you are also buying a full BI platform around it.
Two vendors are conspicuous by their absence from a per-question comparison. Tableau and Qlik both ship AI assistants, and neither publishes a rate you can divide. Tableau stopped publishing Creator, Explorer and Viewer prices altogether, so the widely quoted role figures come from resellers rather than from Tableau, and this page does not repeat them as though they were official. Qlik prices data capacity rather than headcount, with the phrase "No cost for additional users" appearing on its pricing page verbatim, which makes a per-question figure meaningless there for a different reason. Our Tableau pricing page and the Qlik and Power BI comparison cover both in detail.
›_ before you buy conversational analytics software
Four questions that change the number on the quote.
01
Find the meter before the feature list
Ask the vendor which unit the AI is billed in, then ask what a typical month of that unit looks like for a team your size. Tokens, messages, DBUs, CU-hours and included queries are not interchangeable, and a demo will never surface the difference.
02
Separate the thinking from the running
Almost every platform bills the natural-language step and the query execution on different meters. Snowflake says so plainly: the message price covers generation, and running the SQL draws normal warehouse compute. Budget both or the first invoice will surprise you.
03
Ask what happens when the promotion ends
Google is free through 30 September 2026 and priced from 1 October. Databricks Genie is free through 31 January 2027, then falls back to 150 DBUs a user. A pilot that is free today can be a line item next quarter, and neither date is in a sales deck.
04
Check whether it works without a semantic model
Most of these features answer questions inside a dataset someone modeled first. If your question needs a table nobody has published, you are back in the analyst queue no matter how good the language model is. Test the awkward question, not the demo one.
›_ a different shape of answer
Agentsql prices the seat, not the question.
The reason all of the above is complicated is that conversational analytics is a feature inside a platform, and platforms bill by whatever they already billed by. If you are already deep in Snowflake, Databricks or Fabric, using the native feature is often the right call, and the tables above are there to help you budget it honestly rather than talk you out of it.
Agentsql is the other option: a conversational analytics layer that sits on top of the database you already have, rather than inside a BI platform you have to adopt first. It connects read-only to PostgreSQL, MySQL, Snowflake or BigQuery, reads your live schema so it uses your real table and column names, writes SQL in that engine's own dialect, runs it and returns the answer with the generated query beside it. There is no semantic model to publish before a new question can be asked, no per-message meter, and no capacity minimum. Our pricing is a flat monthly figure and nobody counts your questions.
Where it is the wrong tool: if you need governed, certified metrics that every department reports against, you want a semantic layer and a BI platform, and you should buy one. If your data is scattered across a dozen SaaS applications rather than sitting in a database, the language model is not your problem yet. And if the recurring numbers your team watches every Monday are already on a dashboard that works, leave them there. This is for the long tail of questions nobody built a report for, which is where ad hoc reporting lives, and it works by text to SQL, with the accuracy caveats that come with it.
›_ frequently asked
Conversational analytics questions, answered.
What is conversational analytics?
Conversational analytics is asking a data platform a question in plain English and getting a number, a table or a chart back, instead of building or reading a report. The system interprets the question, identifies the metrics, filters and time periods you meant, generates a query, runs it and returns the answer.
What is the difference between conversational analytics and conversation analytics?
They are two different markets that share a name. Conversational analytics in the data world means querying your own database or warehouse in natural language. Conversation analytics in the contact-center world means mining recorded customer calls and chats for sentiment and intent. Vendors like NiCE and Sprinklr sell the second; Looker, Databricks and ThoughtSpot sell the first.
How does conversational analytics work?
The tool reads your schema or semantic model so it knows what tables and columns exist, interprets your question against that structure, writes a query in the right SQL dialect, executes it read-only and formats the result. The useful ones show you the generated query, so you can check the logic instead of trusting the number.
How much does conversational analytics cost?
It depends entirely on the unit your vendor meters. Snowflake charges 67 Platform Credits per 1,000 messages through the Analyst API, which is about 13 to 27 cents a question depending on edition. Google charges $3 per million input and $20 per million output data tokens. Microsoft requires a Fabric capacity starting at roughly $263 a month. There is no single market rate.
How much does Looker conversational analytics cost?
Google prices the Conversational Analytics agent as part of Data Cloud agents at $3 per million input data tokens and $20 per million output data tokens. It is free through 30 September 2026, and that pricing takes effect on 1 October 2026. Output costs about 6.7 times input, so the length of the answers drives the bill more than the number of questions.
Is there a conversational analytics API?
Yes, several. Google ships a Conversational Analytics API for BigQuery and Looker data, and Snowflake exposes Cortex Analyst as an API. Snowflake bills the API route by the message and the agent route by the token, so the same feature costs different amounts depending on which endpoint you call.
What is conversational analytics in Power BI?
It is Copilot. Microsoft documents that Copilot works in Power BI Desktop and in Pro and Premium Per User workspaces, but the usage still bills to a Fabric capacity, and that capacity "must reside on at least an F2 or P1 SKU". In practice that means there is no Copilot in Power BI without a Fabric capacity behind it, even for Pro seats.
Does Snowflake have conversational analytics?
Yes, called Cortex Analyst. Snowflake's Credit Consumption Table dated 26 August 2026 prices it at 67 Platform Credits per 1,000 messages when you use the Analyst API, and separately in AI Credits per million tokens when you reach it through Cortex Agents or CoWork. Only successful responses bill, and running the generated SQL draws normal warehouse compute on top.
What is conversational analytics in Databricks?
It is AI/BI Genie. Genie One and Genie Agents are free through 31 January 2027, and each user gets 150 free DBUs a month that reset on the first. At the $0.070 Genie DBU rate that allowance is worth about $10.50 a user. Service principals are excluded, so anything automated or embedded bills from the first question.
Is there an open source conversational analytics tool?
There are open source text-to-SQL projects and open source BI tools like Metabase and Superset with natural-language features bolted on, but the language model itself is almost always a paid API call. Self-hosting moves the cost from a vendor line item to an inference bill and an engineer, rather than removing it.
Do conversational analytics tools need a semantic model?
Most of the platform-native ones do. Looker answers against LookML, Power BI against a semantic model, ThoughtSpot against a worksheet. That is why they answer confidently inside the model and stall outside it. Tools that read the live database schema instead can answer a question about a table nobody has modeled yet.
Are conversational analytics answers accurate?
Accurate enough to be useful and not accurate enough to trust blind. The failure mode is rarely broken SQL; it is correct SQL answering a slightly different question than you asked. That is why the generated query needs to be visible. Reading the SQL takes ten seconds and is the difference between a tool and a liability.
Which databases does Agentsql connect to?
Four: PostgreSQL, MySQL, Snowflake and BigQuery. Each gets SQL written for its own dialect rather than one generic string, which matters more than it sounds, because these engines disagree on date truncation, string concatenation and identifier quoting.
›_ how it works
›_ connect your database
›_ what the AI features cost
Ask your database a question.
Read-only on Postgres, MySQL, Snowflake or BigQuery. The answer comes back with the SQL beside it, and nobody counts your questions.