Microsoft Fabric Data Agents: What They Cost and How to Size Your Capacity

Microsoft Fabric data agents let business users ask questions of their enterprise data in plain English "which regions missed SLA last quarter, and why?" and get governed, trustworthy answers without writing a line of SQL. Since April 2025, they no longer require an expensive F64 capacity; any paid Fabric SKU from F2 and up can run them.

That change opened the door for small and mid-market teams. But it also created a new question that almost no one answers clearly: what does a Fabric data agent actually cost, and how big a capacity do you really need?

This guide breaks it down no marketing fluff, just how the billing works and how to size it.

What is a Microsoft Fabric data agent?

A Fabric data agent is an AI assistant that sits on top of your governed data in OneLake. You point it at trusted, modeled data a warehouse, a lakehouse, a semantic model and it answers natural- language questions by interpreting intent, generating and running queries, and summarizing the result.

A data agent is only as trustworthy as the data beneath it. Pointed at raw, duplicated, ungoverned data, any AI produces confident-but-wrong answers. Pointed at a well-modeled Fabric warehouse with clear, consistent measures, it produces answers your leadership can actually act on. It also inherits each user's existing Fabric permissions so the agent can only access what that person could already access. No new security surface.

That foundation a clean, conformed, permissioned data layer is exactly what a good data platform engagement delivers, and it's the prerequisite most "just add AI" projects skip.

Do Fabric data agents cost extra?

The short answer -

No there is no separate license or per-message fee for Fabric data agents. Usage is billed as Capacity Unit (CU) consumption drawn from the same Fabric capacity you already pay for. There is no add-on SKU and no per-question charge.

If you buy a reserved F2 capacity roughly $156/month on a one-year commitment (about $262/month pay-as-you-go) that reservation is your budget. Data agent questions consume CU-seconds from that pool, alongside your normal warehouse queries, pipelines, and Power BI refreshes.

So the honest headline: your capacity is the cost. The real question isn't "what's the add-on fee" it's "will my capacity hold up under the AI load?"

What actually costs more than the capacity

Three things sit outside the flat capacity price. Budget for them:

Item	Extra cost?	Notes
OneLake storage	Yes, always separate	~$0.023/GB/month (~$23/TB). Usually small.
Capacity overage / autoscale	Only if you enable it	Bursts above your SKU bill at pay-as-you-go rates. Optional.
Throttling	Not a charge a slowdown	On a small SKU, heavy AI load can exhaust the pool and queue requests rather than bill you more (unless overage is on).

That last row is the one that catches teams by surprise. Let's size it.

How much capacity does a data agent actually use?

Every data agent question fires multiple internal AI calls it interprets your question, plans an approach, generates SQL, runs it, then summarizes the answer. Fabric bills AI work at roughly 100 CU-seconds per 1,000 input tokens and 400 CU-seconds per 1,000 output tokens.

In practice, a single data agent question consumes somewhere between ~800 CU-seconds (a simple lookup) and ~3,500 CU-seconds (multi-step reasoning), with a typical question around ~1,800 CU-seconds.

An F2 capacity provides 2 CUs, which is 172,800 CU-seconds per day. So:

• 1 user asking ~10 questions/day → ~11% of an F2's daily budget. F2 is comfortable.

• 5 users asking ~20 questions/day (100 questions) → over 100% of an F2's daily budget at typical query weight. F2 runs hot and will throttle at peak.

• 10+ users → well beyond F2. You'd need F4/F8 or capacity overage.

Fabric's 24-hour smoothing spreads background load, which helps but interactive AI needs headroom right now, so peak concurrency matters as much as the daily total.

How to size Fabric capacity for data agents

A practical sequence that avoids both overspend and throttling:

1. Start on F2 for pilots and single power users. It's the cheapest way to build, demo, and prove value. One heavy user barely dents it.

2. Install the Fabric Capacity Metrics app on day one. It shows exactly how many CU-seconds each data agent query burns replacing estimates with your real numbers.

3. Watch peak concurrency, not just the monthly total. Throttling happens when everyone asks at once during business hours, even if your daily percentage looks fine.

4. Scale to F4 or F8 when you onboard concurrent users. The step from F2 (~$156/mo reserved) to F4 (~$312) or F8 (~$624) is linear, and it's the difference between a snappy agent and a queuing one.

5. Consider reserved vs. pay-as-you-go carefully. Reserved saves ~41% but can't be paused; pay-as-you-go costs more per hour but can be paused when idle. For always-on team use, reserved wins.

The bigger picture: AI is only as good as your data

The temptation with Fabric data agents is to buy a capacity, switch on the agent, and expect magic. The teams that get real value do the unglamorous work first: consolidate sources into OneLake, model a clean warehouse with consistent measures, set permissions, and then let the agent reason over that trusted layer. That's what turns "an AI chatbot" into "a governed answer engine leadership trusts."

If you're evaluating Fabric data agents and want to know whether your data foundation is ready or how to size capacity for your specific team that's exactly the kind of thing we help with.

Frequently asked questions

Do Microsoft Fabric data agents require an F64 capacity?

No. Since April 2025, Copilot and Fabric data agents run on any paid capacity F2 and above. The old F64 requirement was removed. (F64 is still needed only for a shared Fabric Copilot Capacity across many workspaces.)

Is there a separate license or per-message fee for a Fabric data agent?

No. Data agent usage is billed as Capacity Unit consumption from your existing Fabric capacity. There is no add-on SKU and no per-question charge.

How much does a Fabric data agent cost on F2?

An F2 capacity is roughly $156/month reserved (about $262/month pay-as-you-go), and that covers data agent usage within capacity. Add a small OneLake storage cost, plus optional overage only if you enable it. The practical limit is throttling: F2 suits a single power user or a pilot, not a team of concurrent users.

Will F2 be enough for my team?

For one or two light users, yes. For five or more people asking the agent throughout the day, F2 will likely throttle at peak you'd move to F4 or F8. Install the Fabric Capacity Metrics app to measure your real consumption before scaling.

What makes a Fabric data agent give trustworthy answers?

A governed data foundation: cleaned, conformed data in a well-modeled warehouse with consistent measures and proper permissions. The agent reasons over that layer, so its answers match the KPIs your organization already trusts.

microsoft fabric data agents

At Aptocoiner Analytics, we help organizations:

Power BI Consulting Services - Dashboard and report development - Data modeling and optimization - Migration from legacy BI tools - Training and enablement	Microsoft Fabric Implementation - Fabric architecture design - Data lakehouse setup - Real-time analytics pipelines - OneLake configuration
Data Warehouse Implementation - Modern data warehouse design - ETL/ELT pipeline development - Data governance frameworks - Performance optimization	Business Intelligence Strategy - BI roadmap development - Tool selection and evaluation - Data strategy consulting - Proof of concept development