Home

Credits and capacity: why AI automations can run out

An AI-enabled flow can work perfectly in testing and still fail later because the organisation has run out of the capacity that pays for the AI step. This page explains the current transition from AI Builder credits to Copilot Credits and how to plan enough capacity for production.

Last reviewed 2026-09-05

Two credit currencies are overlapping during the transition

Credit typeWhat it isWhat happens next
AI Builder creditsLegacy capacity used by AI Builder features in Power Apps and Power Automate. Power Automate Premium currently contributes 5,000 seeded credits per active licence.Seeded credits are removed 1 November 2026. New customers cannot buy new AI Builder capacity add-ons; existing add-on customers have transitional rights according to contract.
Copilot CreditsMicrosoft's newer common consumption currency for Copilot Studio/AI tools and increasingly AI Builder scenarios.Can be supplied through prepaid capacity or pay-as-you-go. AI Builder in Power Apps/Power Automate can fall back to Copilot Credits when AI Builder credits are unavailable or exhausted.

What happens when the AI Builder pool runs out?

Microsoft calculates consumption during the month. If an environment moves into overage, it first tries to use available Copilot Credits. If neither AI Builder credits nor Copilot Credits are available, AI Builder model runs in apps and flows can be blocked and fail with capacity/entitlement errors. Microsoft also notes that consumption calculation can be delayed, so a flow may appear to keep working before enforcement catches up.

Why 'how many credits do I get?' is not enough

QuestionWhy it matters
How much capacity is available?You need the tenant/environment allocation, not just the licence brochure.
What AI action am I using?OCR, document processing and generative prompts consume at different rates.
Which model is selected?Prompt consumption depends on input/output tokens and the model class.
How often does the flow run?A cheap action repeated thousands of times becomes a meaningful monthly load.
How large is each input?Longer documents and prompts increase consumption.
What else shares the environment?Other makers, apps and flows can draw on the same capacity.
What happens at zero?Without fallback capacity, the AI step can fail even while the rest of the flow is correctly designed.

A simple capacity-estimation method

StepManager/maker question
1. CountHow many items will the automation process in a normal and peak month?
2. MeasureHow many pages/tokens/images does one average item contain?
3. IdentifyWhich AI action and model class will each item invoke?
4. MultiplyEstimate monthly credit demand, then add a sensible peak/error/retry margin.
5. IsolateFor an important pilot, consider a dedicated environment or clear capacity allocation so other workloads do not silently consume the same pool.
6. MonitorReview actual consumption weekly during the pilot and monthly in business-as-usual.
7. Decide at zeroChoose deliberately: block, reallocate capacity, buy prepaid credits, or allow pay-as-you-go overage.

Prepaid capacity vs pay-as-you-go

A prepaid Copilot Studio capacity pack currently provides 25,000 Copilot Credits/month for £153.80/month paid yearly at UK list price. Pay-as-you-go is also available and requires an Azure subscription. Prepaid capacity makes the maximum committed capacity easier to budget; PAYG avoids an upfront pack but introduces variable Azure billing. Organisations can also combine prepaid capacity with PAYG overage for continuity.

Related guidance

Sources

Related