Minor update to Responsible Use of Microsoft AI Services Web IQ added to Privacy & Security terms M365 G3 & G5 added as eligible pre-requisites for M365 Copilot Minor renaming of Microsoft Defender Expert services Updates to terms around Professional Services
Slightly bigger is the clarification/addition to terms that customers with SA who enable Azure Arc are eligible for ESU (Extended Security Updates).
There are some other changes that weren’t in the Product Terms:
A robot wearing a GitHub Copilot Workshop vest stands in a busy office with people collaborating in the background.
Microsoft have announced General Availability of a new feature within Copilot Studio – the GitHub Copilot harness. This sits alongside the already available Standard harness and Copilot Chat harness.
This new capability means your in-house created agents can now use the same coding and reasoning capabilities as Copilot Cowork and GitHub Copilot – putting new capabilities on the table for your AI agents.
While that is all well and good I would, not surprisingly, like to focus in on a licensing/pricing change. The Microsoft post casually states:
“Agents running on the GitHub Copilot harness use usage-based billing for all work, regardless of Microsoft 365 Copilot licensing.”
That, potentially, has quite the impact. Agents built using the 2 existing harnesses have their Copilot Credits covered by a user’s M365 Copilot license – this is not the case when agents are built with the GitHub Copilot harness. This means those agents will use Copilot Credits and burn pre-purchased amounts or incur PAYG costs.
To be clear, Microsoft aren’t removing anything or forcing this more expensive model on customers; the GitHub Copilot harness is a new, additional option that organisations can choose to use, or not. However it does add another potential way for people to inadvertently spend money with Microsoft and thus further highlights the need for storng internal governance and awareness/training programs.
A customer discusses a return with a store employee near a return policy sign.
From February 1st, 2027 – Azure Reservations will no longer be eligible* to be exchanged if the service also has a Savings Plan available. This applies to:
Azure Virtual Machines
Azure App Service
Azure SQL Database
and “similar services”.
*Reservations purchased before February 1, 2027, keep the right to one final exchange.
Equally, from that date any additional compute or database services that also become available for a Savings Plan will be subject to this change. Instance Size Flexibility isn’t impacted by this change.
The following are excluded from the change:
Reservations for products or services that are deprecated and approaching end of life.
Reservations for products and services that aren’t covered by savings plans, such as Azure VMware Solution. If you have a reservation for Azure VMware Solution, this policy change doesn’t affect it.
Cloud environments that don’t currently support savings plans.
Microsoft initially announced this change in October 2022 with a date of Jan 1, 2024. That was then pushed back to July 1, 2024 and then postponed indefinitely – with a promise of 6 months notice before it would be implemented again…and that’s what Microsoft have given. Could we see another postponement? Perhaps, but it seems unlikley tbh.
It’s clear that Microsoft want Savings Plans to be the primary option for organisations so here’s a quick overview if you’re not familiar:
Feature
Azure Reserved Instances
Azure Savings Plans
Commitment
Commit to a specific Azure resource (e.g. VM family, SQL Database, App Service)
Commit to an hourly spend amount on eligible compute & database services
Term
1 or 3 years
1 or 3 years
Discount
Typically offers the highest discount when usage is stable
Lower maximum discount, but greater flexibility
Flexibility
Low – tied to the purchased resource
High – discounts automatically apply across eligible compute resources
Changing VM Size
Allowed within instance size flexibility rules for some services
No action required
Changing VM Family
Previously possible through Reservation exchanges (being phased out)
No action required
Changing Region
May require a Reservation exchange (being phased out)
No action required
Workload Changes
Best suited to stable, predictable workloads
Designed for dynamic, changing cloud environments
Administration
Requires ongoing monitoring and optimisation to ensure Reservations still match usage
Much lower administrative overhead once purchased
Best For
Long-running production workloads with predictable resource requirements
Organisations with frequently changing compute usage, autoscaling or modern cloud architectures
Risk
Higher risk of underutilisation if workloads change
Lower risk, as discounts follow eligible usage automatically
FinOps impact
The ability to exchange Reserved Instances has been part of mature Azure commitment strategies for quite some time but this is bringing much of that to an end. Strategies will need to be reviewed and a decision made on whether Savings Plans are now a better option, even with the potential loss of discount.
Make sure you communicate this change internally and work to update your governance and strategy asap.
Microsoft has officially announced that Copilot Cowork is generally available. If you caught my LinkedIn post on this just after launch, you’ll know I flagged the headline numbers – and promised a deeper look. Here it is.
This isn’t just another Copilot feature release, Cowork represents a different, additional way of working…and brings with it an additional billing mechanism.
What Is Copilot Cowork?
Copilot Cowork is Microsoft’s agentic AI layer sitting on top of Microsoft 365 Copilot. Where standard Copilot handles in-the-moment assistance (drafting emails, summarising meetings, generating content), Cowork is designed for longer-running, multi-step tasks that span multiple apps and require sustained reasoning in the background.
How the Billing Works
This is where things get interesting and where CFOs – and everyone else – will want to pay close attention.
Copilot Cowork uses a seat + consumption model:
Seat requirement: Users must already hold a Microsoft 365 Copilot licence Consumption billing: Cowork usage is billed on top of that, via Copilot Credits
In other words, you’re not paying a flat per-user fee for Cowork. You’re paying based on what people actually do with it. That might sound fair – but as we’ll see, it creates significant cost unpredictability at scale.
The Three Prompt Types
To help organisations estimate their likely spend, Microsoft has defined three categories of prompt complexity:
Light Simple, quick tasks – lookups, short summaries, straightforward Q&A
Medium Multi-step tasks with moderate reasoning or tool use
Heavy Complex, long-running agentic tasks – deep research, multi-app orchestration, extended workflows
The credit cost per interaction scales accordingly. A user who primarily sends light prompts will consume far fewer credits than one regularly triggering heavy agentic workflows.
The Four Microsoft-Defined Personas
Microsoft has also defined four user personas to help organisations model expected usage – and therefore expected cost:
Knowledge Worker
Your standard office employee: using Copilot for day-to-day tasks like drafting documents, summarising emails, and pulling information. Predominantly light-to-medium prompt usage.
Customer-Facing Knowledge Worker
Staff in sales, customer success, support, or account management. Higher interaction volume and a greater likelihood of medium-to-heavy prompts – researching customer history, generating proposals, triaging complex queries.
Technical Worker
Developers, analysts, engineers, and data professionals. Usage tends towards heavier, more complex prompts – code generation, data analysis, technical documentation, multi-step problem solving.
Manager / Senior Leader
Executives and team leads. Usage is often more strategic – executive briefings, synthesising reports across sources, preparing for key meetings. Likely lower volume but higher complexity per interaction.
Microsoft has shared estimated annual Cowork costs based on data from early Frontier customers. Modelling the costs is where things start to get really surprising…the below uses list pricing and doesn’t factor in any type of discount.
As was pointed out in a comment, there is a small but important difference between the two. The Github hosted calculator estimates “Heavy” prompts at 2,500 credits while the spreadsheet version uses a value of 1,200. This means the latter version produces lower prices for those heavy prompt users.
Let’s look first at a small org of 60 users:
The Microsoft calculator uses the following estimates for the number and type of prompts each persona will use:
and these for the number of credits used per prompt:
That gives a final estimate of:
That’s right – over $164,000 per year for 60 people to use Copilot Cowork.
For a company of 1,680 staff:
You end up with an estimated annual bill of almost $5,000,000!
It seems impossible that companies are going to pay these amounts – surely? If these costs are real, it shows that customers are going to have to be much more realistic as to who gets access to Copilot Cowork.
The Guardrails Microsoft Provides
To be fair to them, Microsoft have built in some controls:
Spending limits – administrators can cap Cowork credit consumption at tenant, group, and user levels Usage alerts – notifications when consumption approaches defined thresholds
Usage reporting – Admins see usage broken down by user, group, and feature
User-level pricing – Users see what each task costs as they run it (coming soon)
These are sensible features, and their inclusion suggests Microsoft is aware of the sticker shock potential and are trying to get out in front of it.
This is what I really want to focus on, because the billing model isn’t just a procurement question — it’s an organisational design question.
Not all usage is equal value
A heavy prompt from a technical worker building an internal tool could save dozens of engineering hours. A heavy prompt from someone using Cowork to draft a quick internal update is a poor use of credits. The credit model treats both the same. Your organisation needs a way to distinguish between them.
You need a usage policy, not just a spending cap
A spending cap is a ceiling. What you actually need is a framework that answers questions like:
Which personas should have access to Cowork at all?
What types of tasks are appropriate for Cowork vs. standard Copilot?
Who approves high-complexity agentic workflows?
How do we measure whether Cowork usage is delivering value?
Without answers to these, you’re handing out a consumption-based service with no purchasing guidelines.
The ROI question is now urgent
With flat-fee AI tools, ROI questions are important but not time-sensitive – you’re paying regardless of use. With consumption billing, poor adoption and excessive adoption are both problems. You need a framework for measuring the value of AI usage – not just the cost.
Consumption models reward the vocal, not the strategic
In many organisations, power users will naturally gravitate toward the most capable features. That’s not always aligned with where the highest-value use cases are. Without intentional governance, Cowork credit consumption may cluster around enthusiastic individuals rather than high-value workflows.
What Organisations Should Do Now
If you’re already running Microsoft 365 Copilot – or planning to – here’s where to focus:
Map your personas
Use Microsoft’s four categories as a starting point, but refine them for your organisation. Who are your heaviest potential users? Where are the highest-value use cases?
Model your costs before you deploy
Use the per-persona estimates to build a realistic cost projection. Stress-test it against both optimistic and conservative adoption scenarios.
Define your governance framework
Decide who gets access to Cowork, for what purposes, and with what approval process for high-complexity tasks. Document this as policy, not just an IT configuration.
Set up monitoring from day one
Don’t wait for the first bill to understand usage patterns. Use Microsoft’s alerting tools, and complement them with your own reporting.
Establish a value measurement approach
Credits spent should be traceable to outcomes. What did that agentic workflow actually deliver? This doesn’t need to be complex – even a lightweight system for use case categories can help you build the picture.
As I say every few months, it’s another bumper quarter for Microsoft with all the numbers getting bigger.
Revenue = $82.9 billion (up 18%)
Net Income = $31.8 billion (up 23%)
Microsoft Cloud = $54.5 billion (up 29%)
Microsoft had Operating Expenses of $17.7 billion this quarter and say they were “primarily driven by continued investments in R&D compute capacity, AI talent, and data“.
Productivity & Business Processes
Revenue = $35 billion, up 17%
Microsoft 365 Commercial cloud revenue increased 19%
Dynamics 365 revenue increased 22%
M365 Copilot is now over 20 million paid seats. That’s a 33% increase over Q2 (where it was 15 million) but still a fraction of the overall customer base.
Paid M365 Commercial seats grew 6% YoY and ARPU increased driven by E5 & M365 Copilot.
Intelligent Cloud
Revenue = $34.7 billion, up 30%
Azure increased 40%
Earnings Call
Satya Nadella started by saying “We are at the beginning of one of the most consequential platform shifts that will change the entire tech stack as agents proliferate and become the dominant workload.”
He says there are “Tens of thousands of companies are already managing tens of millions of agents in Agent 365” – frankly, that surprises me.
Cosmos DB had 50% YoY revenue growth
Microsoft Fabric up to 35,000 paid customers, a 60% YoY increase
Copilot Credit spend almost doubled quarter over quarter
Nearly 60% of “service customers” are buying consumption credits
I’ve been banging on about consumption billion (aka Pay as You Go) billing for a few years now and it’s definitely where Microsoft (and a lot of the software industry) is going. Amy Hood, MS CFO, said;
“I start to think about it as a license business plus a consumption business, and really applying far more broadly than I think people have thought about that. And so, it starts to mean that over time, bookings will actually also look a little different. It’ll still have that per-seat license logic, but it’ll also have a meter, just like you see in Azure.”
and Satya said:
“seat-based pricing is just entitlement to some consumption”
While I think he means that seat based licenses are effectively a bundle of some consumption packaged up, that certainly sounds like it being a gateway to more consumption charges…which of course it is in many cases.
A group of software developers collaborate around computers with AI interfaces in a modern office
From June 1st, GitHub Copilot usage will start to consume GitHub AI Credits. Microsoft say that, as Copilot use has changed and become more complex, inference costs have increased and they can no longer sustain the current premium request unit (PRU) model.
Copilot features that consume AI credits include Copilot Chat, Copilot CLI, Copilot cloud agent, Copilot Spaces, Spark, and third-party coding agents.
Every Copilot plan will include an amount of GitHub AI Credits, and paid plan users will be able to purchase additional credits if needed. Microsoft say that “usage will be calculated based on token consumption, including input, output, and cached tokens“.
Copilot Business: $19/user/month, including $19 in monthly AI Credits
Copilot Enterprise: $39/user/month, including $39 in monthly AI Credits
There will be higher credits included for June, July, and August:
Copilot Business: $30 in monthly AI Credits
Copilot Enterprise: $70 in monthly AI Credits
1 AI credit = $0.01 USD so the standard inclusions are:
Plan
Total AI credits per user per month
Copilot Business
1,900
Copilot Enterprise
3,900
As is becoming common in FinOps for AI, the choice of models used for tasks will become much more important with this change. For example:
GPT 4.1 is $2 per input token and $8 per output token
GPT 5.5 is $5 per input token and $30 per output token
Claude Haiku 4.5 is $1 per input token and $5 per output token
Claude Opus 4.7 is $5 per input token and $25 per output token
Significant price differences for sure! See more here.
Usage will be pooled across an organisation which may help reduce the impact of this, depending how uniformly your teams use these features. A “Preview Bill Experience” has been introduced earlier in May to give users a view of what your consumption bill will look like. (That’s quite a good idea but more than a 1 month run up would have been better).
A few key things to note:
Base plan pricing isn’t changing.
Code completions and Next Edit suggestions will not consume AI credits.
Credits do not roll over from month to month.
Fallback experiences will no longer be available.
Copilot code review will consume GitHub Actions minutes AND GitHub AI Credits.
Digital cloud linking to a modern data center in a pixel art style.
A new addition, Azure Capacity Blocks (ACB):
“allow Customers to purchase a fixed duration block of capacity for a specific Microsoft Azure resource in a specified region, with a scheduled start date in the future.”
They can range from 1 day to 6 months, they are fully paid for upfront, and cannot be cancelled or refunded. Unused portions of ACBs will not be refunded either. Once the term ends, customers “will be evicted from the applicable capacity, and Microsoft will stop Customer’s use of the applicable Microsoft Azure Services.”
When I initially looked, there was a Learn site for this new release but I cannot find it now. Not sure if it’s been removed or I (and ChatGPT) just aren’t looking in the right place all of a sudden 🤔
This new offering must be related to the capacity issues that Microsoft have been having in their Azure datacentres – will it help prevent that from reoccuring?