Adaptrix Research · July 2026
Mainstream analytics platforms increasingly bill AI by consumption — per token, per credit, per DBU, per capacity unit. The result is a pricing structure where the faster a mid-market team adopts AI analytics, the more it is charged — and where the advertised licence is only a fraction of the real Year-1 bill.
Published 2026-07-17 · Method: public-data synthesis of published vendor rate cards plus a transparent, rebuildable cost model. Every figure carries a dated source (appendix below). Modelled findings are labelled as modelled — this is not a survey.
of the platforms we examined publish no AI overage rate at all — so the AI line cannot be modelled before the contract is signed.
questions per month is Qlik’s Answers allowance at both Premium and Enterprise — flat, not per user, so the allowance per head shrinks as you roll out.
dated AI price changes land inside six months: Snowflake on 1 September 2026, Looker overage billing on 1 October 2026, and the end of Databricks’ free Genie period on 31 January 2027.
is the cost-calculation error Gartner warns CIOs can make when they don’t understand how their generative-AI costs scale.
For two decades, analytics software was priced per seat: you knew the bill before the year started. That era is ending. Of the six leading analytics platforms we examined, four publish no AI overage rate at all — Microsoft Fabric Copilot draws Capacity Units per token and Snowflake Cortex bills credits per message, both of which a buyer can at least reason about. ThoughtSpot caps queries per user, Qlik grants a flat monthly allowance, and Databricks does not publish DBUs per question, so none of those three can be converted into an expected bill. Tableau gates AI behind a quote-only edition instead.
This mirrors the broader software market: in ICONIQ’s 2026 State of AI survey of ~300 software executives, 35% of companies now use usage-based pricing and 18% outcome-based — and 37% plan to change their AI pricing model again within a year. The meter is spreading, and analytics buyers are on the receiving end.
| Platform | AI feature | How AI is billed | Published rate (verified 2026-07-17) |
|---|---|---|---|
| Microsoft Power BI / Fabric | Copilot in Fabric | Capacity Units, drawn per token | 100 CU-seconds per 1,000 input tokens and 400 per 1,000 output tokens, against a paid Fabric capacity (minimum F2). Past the capacity you are throttled rather than billed, unless you opt into overage. |
| Snowflake | Cortex AI / Cortex Analyst | Credits per message | 67 platform credits per 1,000 Cortex Analyst messages. EU Frankfurt credits are $2.60 Standard, $3.90 Enterprise, $5.20 Business Critical. The rate for the newest model rises 50% on 1 September 2026. Warehouse compute to run the generated SQL is billed separately. |
| Databricks | AI/BI Genie | Free for users until 31 January 2027, then DBUs | Interactive Genie use by identified users is free until 31 January 2027. Service principals are excluded and bill from the first request, so the same volume of questions costs nothing or a great deal depending only on whether the caller carries a user identity. DBUs per question are not published. |
| ThoughtSpot | Spotter AI agent | Per-user query cap, no published overage | 25 Spotter queries per user per month, on Enterprise as well as Pro. The cap is per user and does not pool across a team. ThoughtSpot states separately that LLM tokens are included and not metered. No overage rate is published. |
| Qlik | Qlik Answers | Flat monthly question allowance | Allowances of 25, 200, 1,000 and 1,000 questions per month by tier — flat from Premium through Enterprise, and not scaled by user count. Qlik Answers starts at the Premium tier. No overage rate is published. |
| Tableau (Salesforce) | Tableau Agent | Edition-gated, quote-only | Requires the Tableau+ / Tableau Cloud+ edition, which is quote-only. Since October 2025 AI in Tableau no longer consumes Einstein Requests, though some AI features of Tableau Next still consume Flex credits, and the optional Einstein audit trail consumes credits in proportion to AI calls. |
A licence price is negotiated once and appears on the quote. The AI meter does neither. It is billed in a different unit from the licence — capacity units, credits, DBUs or questions — and in most cases the conversion between that unit and a question your team actually asks is not published.
Copilot in Fabric bills 100 CU-seconds per 1,000 input tokens and 400 per 1,000 output tokens against a paid capacity. Snowflake Cortex Analyst bills 67 platform credits per 1,000 messages. Those two can at least be reasoned about. ThoughtSpot caps queries per user without publishing an overage rate, Qlik grants a flat monthly allowance, and Databricks does not publish DBUs per question at all — so for three of the six, a buyer cannot turn expected usage into a number at any price.
Where an allowance exists, it is granted in a way that shrinks as the deployment grows. Qlik's Answers allowance is 1,000 questions a month at both Premium and Enterprise — a flat figure that does not scale with users. Twenty people share it comfortably; two hundred do not. ThoughtSpot caps per user rather than pooling across the team, so a heavy analyst cannot draw on a colleague's unused quota. Databricks grants its free units per identified user, and excludes service principals altogether.
The effect is the opposite of what an analytics platform should reward. The allowance is most generous when nobody is using the product, and tightest at exactly the moment adoption succeeds.
This is not a static picture, and the direction of travel is published in advance. Three changes to AI billing land inside six months of this study:
Two of the six platforms here impose no marginal cost on an interactive AI question today. One of those two has a published end date. A procurement decision made on this year's rate card is not a decision about next year's bill.
An externally metered AI endpoint is not just a cost line — it is an ICT third-party dependency. Under DORA (applicable since 17 January 2025), financial entities must record every contractual arrangement with ICT third-party service providers in a register of information (Art. 28(3)) — supervisors collected the first registers by 30 April 2025. A consumption-billed AI feature that routes questions to an external model endpoint belongs in that register like any other cloud dependency. Microsoft’s own documentation notes that Copilot in Fabric is powered by Azure OpenAI models deployed in a limited set of data centers, with optional cross-geo processing outside the user’s region.
The dependency question is structural for Europe: Synergy Research puts European cloud providers at just 15% of the European cloud market, with Amazon, Microsoft and Google holding 70%. And the EU AI Act timeline keeps moving — GPAI obligations have applied since 2 August 2025 with Commission enforcement powers from 2 August 2026, while the Digital Omnibus (May 2026) pushed high-risk obligations to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). Every metered external AI dependency is a line in that compliance ledger too.
Full disclosure: Adaptrix, the publisher of this study, sells the alternative it describes — an AI decision engine that runs open-weight models on Adaptrix-operated sovereign EU infrastructure. The contractually defined inference capacity is included in the annual price, and within that envelope there is no per-question and no per-token billing. Seats are tiered like any platform (the public Startup tier is €2,500/month for up to 25 users, billed annually; larger tiers are quote-based). We are not claiming a single flat number beats every meter at every scale — for genuinely light AI usage, a per-question meter can cost less. The point of this study is narrower and, we think, more useful: when AI is billed by consumption, the cost of asking is never zero and never predictable, and that is a structural fact worth pricing into any analytics decision.
The hidden AI tax is the gap between an analytics platform’s advertised licence price and what you actually pay once its AI features are used, because those features are billed by consumption — per token, credit, DBU, capacity unit or query. The licence is predictable; the AI line grows with every question your team asks. Most leading analytics platforms now bill AI this way, and several publish no overage rate at all, so the size of that line cannot be established before signing.
As of August 2026, based on published rate cards: Microsoft Fabric/Power BI Copilot draws Capacity Units per token, Snowflake Cortex bills credits per message, Databricks AI/BI Genie bills DBUs once its free period ends on 31 January 2027, ThoughtSpot Spotter caps queries per user, and Qlik Answers grants a flat monthly question allowance. Tableau gates AI behind the quote-only Tableau+ edition instead. Four of these publish no overage rate at all.
For most of these platforms it cannot be established from public information. Snowflake publishes a credit rate per message and Microsoft publishes a Capacity Unit draw per token, so those two can be reasoned about. ThoughtSpot, Qlik and Databricks publish no overage rate, and Databricks does not publish DBUs per question at all — which means a buyer cannot convert their own expected question volume into a number before signing. That is the finding, rather than any particular price.
All sources accessed 2026-07-17. Vendor prices move — treat every rate as “as of” that date.