The short answer: very little per question, and enough across a company that it needs a budget line. Until now most businesses have tried plain-language analytics free of charge, or inside a capacity they were already paying for. Google’s Looker pricing page says promotional, unmetered access to Conversational Analytics ends on 31 January 2027; from 1 February each Looker instance receives a monthly pool of data tokens, and usage above it is billed at USD 3 per million input tokens and USD 20 per million output tokens. Databricks’ documentation sets the same end date for free use of Genie One and Genie Agents by people. Copilot in Power BI has never been free: every question already consumes a paid Fabric or Premium capacity. The cost of one answer is small. The risk lies in volume, in agents that ask questions on a schedule, and in a shared capacity that slows every report when AI use climbs.
What changed, and when
| Date | Vendor | What happened |
|---|---|---|
| 8 October 2026 | Looker | Conversational Analytics agents on dashboards become generally available |
| 9 October 2026 | Looker | Token usage reporting becomes generally available on Google Cloud core instances; the default question mode moves from Thinking to Fast |
| 26 October 2026 | Looker | The Looker-managed MCP server, which lets outside AI agents query Looker, becomes generally available |
| 28 October 2026 | Microsoft | Reports more than 44,000 paid Fabric customers, up more than 50 percent in a year |
| 29 October 2026 | Looker | Gemini summaries in Explores and the Visualization Assistant become generally available |
| 31 January 2027 | Looker, Databricks | Promotional free AI usage ends |
| 1 February 2027 | Looker | Monthly token allocations and pay-as-you-go billing begin |
Looker: a shared pool, then a meter
Looker counts Conversational Analytics in data tokens. Input tokens carry the question, the conversation so far and the LookML context; output tokens carry the answer, the generated query, any chart and, in Thinking mode, the model’s reasoning. Each paid internal user, whether Viewer, Standard or Developer, adds 1 million input and 20,000 output tokens a month to a pool the whole instance shares, and each embedded user adds 250,000 and 5,000. A production instance never receives less than 60 million input and 1.2 million output tokens, a non-production instance receives a flat 40 million and 800,000, and unused tokens do not roll over. The meter covers questions in Explores and dashboards, standalone and embedded agents, the API and Looker agents published to Gemini Enterprise.
The pool is generous on input and tight on output. Taking the request size Microsoft uses in its own Copilot example, 2,000 input and 500 output tokens, the minimum pool covers 30,000 questions of input but only 2,400 of output each month. Beyond that, each such question would add about one US cent in output tokens. An instance with 200 internal users has an output pool of 4 million tokens, roughly 8,000 such questions. Long conversations, charts and Thinking mode answers use more.
Databricks: free for people, billed for machines
Genie One is Databricks’ plain-language interface for business users, and Genie Agents are the spaces where data teams define the trusted data, metrics and business rules its answers draw on. Until 31 January 2027 their use by people is free and appears on the bill under a separate GENIE_FREE_USAGE line with no list price. Use by service principals, the identities that automations and applications run as, is excluded and is billed today. Genie Code, the assistant for data teams, is already billed beyond 150 free DBUs per user each month, with a 25 percent discount until the same date. The documentation does not yet say what Genie One and Genie Agents will cost from February, so any 2027 budget carries a line whose price is still unknown.
Power BI: no separate meter, a shared capacity
Copilot in Power BI needs a workspace on a paid Fabric capacity of F2 or above, Power BI Premium of P1 or above, or a Fabric Copilot capacity. Pro and Premium Per User licences are not enough on their own, and trial capacities are excluded. Each request consumes capacity units: 100 CU seconds per 1,000 input tokens, 10 for cached input and 400 per 1,000 output tokens, so Microsoft’s example request costs 400 CU seconds. Copilot runs as a background job smoothed over 24 hours, which lets an F64 absorb about 13,824 such requests a day. An F2 has 48 CU hours a day, and 432 requests of that size would use all of it, leaving nothing for refreshes and reports.
When a capacity runs short, Fabric throttles, and Microsoft’s documentation warns that once a capacity is exhausted all operations stop. Capacity overage, switched on by default for new Fabric capacities, prevents that by billing the excess at three times the pay-as-you-go rate, up to a rolling 24-hour threshold that Microsoft says is not a hard spending cap. On Power BI the risk is less a surprise AI invoice than a finance dashboard that slows down because the sales team has discovered Copilot, or an overage charge nobody budgeted for.
Where the questions are processed
Microsoft’s Fabric region list, updated on 3 November 2026, shows every Fabric workload in UAE North, Power BI alone in UAE Central and Qatar Central, and no region in Saudi Arabia or Pakistan. Microsoft has said its Saudi Arabia East region opens to customers in November 2026, but it does not yet appear on that list. Copilot runs on Azure OpenAI models in a limited set of data centres; where they are not available in a capacity’s region, an administrator must allow data to be processed outside it. That setting deserves a decision from whoever owns data protection. Our article on data residency in Saudi Arabia and the UAE sets out the rules that apply.
Six steps before 1 February
- Measure the free period. Looker’s System Activity dashboards show token use by user and agent, Databricks records free use under its own SKU, and Fabric’s Capacity Metrics app lists Copilot as an operation. Twelve weeks of real usage is a better basis for a budget than any vendor estimate.
- Count the machine users. Scheduled agentic workflows in Looker, service principals in Databricks, outside agents connected through MCP and dashboards embedded for customers or partners all ask questions without anyone typing. Track them separately, because they grow fastest.
- Choose the mode by role. Thinking mode counts its reasoning as output, the more expensive kind of token. Keep it for analysts and leave everyone else on Fast.
- Fix the model before the AI. A plain-language answer is only as good as the semantic layer beneath it: clear measure names, one definition for each metric and verified queries for the common questions. A vague model produces longer conversations, more tokens and wrong numbers.
- Keep AI away from month-end reporting. In Fabric, a separate Fabric Copilot capacity carries the AI part of each request, so a rush of questions does not throttle the board pack. Set an overage threshold you have priced, or switch overage off on purpose.
- Give the cost an owner. Treat AI analytics as a cloud cost with alerts and a monthly review, not a feature that came with the licence. Our guide to protecting your 2027 cloud budget applies the same discipline to compute.
Where SoftwareBiz fits. Our analytics and business intelligence service sets up Copilot in Power BI and Gemini in Looker on governed data: the semantic model, the agreed measure definitions and the verified questions that keep answers correct and conversations short, around dashboards built for the decisions your leaders make, with training so teams answer their own questions. Where the numbers come from many systems, our data engineering and modernization team builds the pipelines and warehouse those models rely on. Where AI use now sits on a cloud invoice, our FinOps and cost optimization team sets the budgets, alerts and capacity sizes that keep it predictable.
What to do this quarter
Looker and Databricks customers have about twelve weeks of free usage left to measure, so switch on the usage reports now and review them each month until February. Power BI customers should check which capacity each Copilot workspace uses, whether overage is on and at what threshold, and whether the cross-region setting was a deliberate choice. Then decide who asks the questions, in which mode, against which model. The vendors have set their prices; the volume is still yours to decide.
