This article has been written with the aid of Microsoft Copilot.

Over the last couple of years, we’ve been spending a lot of time building agents.

Some are simple. Some are quite sophisticated. Some are internal experiments and some are helping solve real business problems. Like many people working in this space, I’m trying to understand where AI is heading and, more importantly, what organisations need to be thinking about today.

One thing has started to become apparent.

Microsoft are moving from a world of fixed AI licensing to a world of AI consumption.

This is a big change; I think it could become one of the biggest changes organisations need to prepare for over the next few months.

And it has got me thinking more about the rise of AI FinOps and what we can do to help our customers.

We’ve Seen This Before

When cloud computing first became mainstream, many organisations struggled to adapt.

Traditionally, IT was relatively predictable. You bought servers. You bought software. You purchased licences. The costs were generally known upfront.

Cloud services changed this completely. Suddenly, organisations were paying for what they consumed: storage, compute, network traffic and databases.

The technology teams loved the flexibility. Finance teams sometimes had a different perspective when the invoices arrived.

This is where FinOps emerged.

FinOps helped organisations understand what they were consuming, why they were consuming it and whether they were getting value for the money that was being spent.

The goal was never simply to reduce costs. The goal was to maximise value.

I can’t help thinking that AI is about to follow a very similar path.

Why Copilot Changes the Conversation

Historically, most Microsoft 365 discussions have centred around user licensing.

How many users do we need? Who should get a licence? Have we assigned the licences correctly?

These are all sensible questions.

However, AI introduces another layer.

As Microsoft continues to expand the capabilities of Copilot, Copilot Studio, AI-powered development tools and agent experiences, we’re increasingly seeing consumption-based models emerge through Copilot Credits.

That changes the conversation.

The question is no longer simply: How many Copilot licences do we own?

It starts becoming: How much AI are we consuming and what value are we getting from it?

Those are very different questions.

What Is AI FinOps?

I suspect AI FinOps will eventually become a specialised extension of traditional FinOps.

Not because AI is completely different, but because the things we need to measure are different.

Instead of virtual machines and storage accounts, organisations may find themselves tracking things such as:

• Agent execution
• AI reasoning
• Copilot Credit consumption
• AI-powered processes
• Usage patterns
• Business outcomes
• Value generated by AI workloads

Ultimately, the challenge remains the same.

How do we ensure we’re getting value from the resources we’re consuming?

The Challenge With Agents

One thing I’ve noticed whilst building agents is that we often focus on capability first.

Can the agent do the job? Can it solve the problem? Can it complete the user’s request?

Those are absolutely the right questions when you’re learning and experimenting.

However, once agents move from experimentation into production, another question starts becoming important.

Can the agent solve the problem efficiently?

For example, imagine a user asks an agent to build a solution.

A very capable agent might analyse requirements, design a data model, build an application, create workflows, produce documentation, generate reports and configure automation.

On the surface, this sounds fantastic.

However, what has happened behind the scenes? How many reasoning cycles did it perform? How many searches did it execute? How many tools did it invoke? Did it arrive at the answer directly or explore five different approaches before choosing one?

The user may only see the final result. The organisation may eventually care about the journey taken to get there.

Why Agent Optimisation Matters

This is where things start to get interesting, **well interesting to me**.

Historically, optimisation has been difficult.

A lot of the intelligence in an agent was buried inside prompts and instructions. As agents became more sophisticated, prompts became longer, instructions became more detailed, and workflows became more complex.

One of the challenges was simply giving the AI enough context to make good decisions.

What I’m starting to see from recent updates is a move towards architectures that make optimisation easier.

The latest developments in agents, orchestration and harness capabilities are creating opportunities to shape how agents reason and how they execute tasks.

That might not sound particularly exciting. But from an AI FinOps perspective, it could be hugely important.

A good optimisation strategy isn’t necessarily about using less AI. It’s about using AI more effectively.

An Interesting Example: Work IQ

One capability that I think will become incredibly important is Work IQ. This is Microsoft’s intelligence layer for accessing your work context.

Traditionally, AI spends a lot of effort trying to understand context.

Who should I talk to? What project is this related to? Where is the information? Which document is relevant? Which team owns this?

That investigative work has a cost.

The more effort required to find context, the more effort the AI has to expend before it can begin solving the actual problem.

What I find interesting about Work IQ is that it helps the AI understand the organisational environment more effectively.

If the AI can find the right people, projects, conversations and information more quickly, in a more succinct way, then it will need less effort to reach the correct answer.

Better context means less work.

Less work means lower consumption.

Lower consumption means better results and value.

I suspect this is one of those areas we’ll be talking far more about over the next few months.

From Adoption to Optimisation

A lot of organisations today are still focused on adoption.

How do we build agents? How do we deploy Copilot? How do we train our users? How do we identify use cases?

Those are still the right questions.

However, with the changes to usage-based pricing, the conversation quickly moves to cost optimisation.

We are already seeing this in practice. A couple of organisations have started building agents with the assumption that the activity would effectively be covered, only to discover that costs can appear unexpectedly or be higher than they had imagined.

Cost estimation is hard at the best of times, and Microsoft’s recent move towards more usage-based pricing through Copilot Credits makes this more likely and, unfortunately, somewhat inevitable. The lesson is not to stop building agents, but to treat cost modelling, monitoring and optimisation as part of the adoption journey from the outset.

How do we govern AI consumption? How do we measure value? How do we optimise agent execution? How do we manage Copilot Credits? How do we scale AI responsibly?

I think that shift is coming sooner than many people realise.

Where Do We Start?

My advice now is simple.

Don’t wait for AI FinOps to become a formally recognised discipline before thinking about it.

Start building agents. Start experimenting. Start learning. Understand what works. Understand what doesn’t. Measure outcomes. Pay attention to consumption patterns.

And most importantly, focus on business value rather than technical activity.

However, put in place a budget, what is a reasonable monthly cost?

Make sure you are adding controls around that budget so that you don’t have unexpected costs appearing and have controls in place so that do not become a surprise.

We have a budget at iThink 365, who are a 9-person organisation of around about £250 per user per month with some variance based on the user role.

Microsoft have started putting cost management controls in place, but they need to be better with more granular controls for specific products and instances of those products. For example, being able to allocate a budget for a specific Agent or product within Microsoft 365.

You can start to read up on those controls here:

Microsoft Azure Cost Management

Overview of Cost Management – Microsoft Cost Management | Microsoft Learn

Microsoft 365 Copilot Cost Management

Usage-Based Billing and Cost Management for Copilot Credits | Microsoft Learn

Power Platform Copilot Studio / Copilot Credit Cost Management

Manage Copilot Credits and capacity for Copilot Studio – Power Platform | Microsoft Learn

Summary

Ultimately, organisations won’t be successful because they run the most agents. They’ll be successful because they generate the most value from those agents.

And that’s why I suspect AI FinOps is going to become a very important conversation over the next few months.

At iThink 365, we’re already having these discussions with customers through our improvement packages and soon via our Agent Adoption Service.

Today, the focus is helping organisations to get started with agents. Tomorrow, I think it will increasingly be about helping them optimise, govern and scale them successfully.

It is going to be fascinating to watch how this space evolves.