When your AI employee gets a raise

For the last few years, many teams treated AI like a cheap experiment.

You bought a license, turned on a feature, and got used to the feeling that the marginal cost of another prompt was close to zero.

That era is ending.

GitHub and Microsoft are both moving toward models that make AI usage more explicit, more metered, and harder to ignore. GitHub says Copilot Business and Copilot Enterprise will move to usage-based billing on June 1, 2026, with GitHub AI Credits as the billing unit. Microsoft is moving AI Builder toward Copilot Credits and prepaid capacity packs, with overages handled through pay-as-you-go billing or blocked entirely if you do not allow them. GitHub docs, Microsoft Learn, Microsoft Learn

That is not just a billing change.

It is a signal.

The hidden message behind the pricing change

The market is telling us that AI is no longer something companies can safely subsidize forever.

For a while, the economics were simple:

• buy seats
• absorb the uncertainty
• let the team experiment
• celebrate adoption

But once AI becomes part of daily work, the cost moves with usage.

If your assistant answers more questions, runs more code, reviews more files, or touches more workflows, it should cost more.

That sounds obvious in hindsight. It is also a hard truth for anyone who started to think of an agent as “cheap labor.”

An agent is not free labor.

It is variable labor.

And variable labor is easy to underestimate when the first version is being subsidized by product strategy, investor expectations, or a vendor trying to win the market.

What changed at GitHub

GitHub has been steadily pushing Copilot from a simple seat license toward measured consumption.

Today, organization plans still include a per-user price, but the docs also show a shift to GitHub AI Credits for Copilot Business and Copilot Enterprise, with a fixed conversion rate of 1 AI credit = $0.01 USD. GitHub also documents monthly included credit pools and a path for additional usage when the pool is exhausted. GitHub usage-based billing

That matters because it changes the conversation from:

“How many seats do we need?”

to:

“How much work will our agents actually do?”

Those are not the same question.

A team can have a small number of users and still burn through a large amount of AI credit if the workflow is heavy, the models are more expensive, or the agent is doing deep work across many files and tools.

In other words, success now has a line item.

What changed at Microsoft

Microsoft is making a similar move in its own stack.

Microsoft says AI Builder credits are being phased out progressively, and the features will continue to work through Copilot Credits. For Microsoft 365 Copilot Chat and SharePoint agents, Microsoft documents prepaid capacity packs with 25,000 Copilot Credits per month per pack, plus pay-as-you-go billing for overages when needed. Microsoft Learn, Microsoft Learn

That is a very different posture from “AI is included.”

It means:

• AI usage can be budgeted
• overages can be enabled or blocked
• teams can be allocated credits by environment or group
• cost control becomes part of the operating model

Again, the signal is clear: AI is becoming a metered service, not a permanent giveaway.

The real cost of an agentic employee

If you want to think clearly about agent economics, do not ask what the tool costs.

Ask what the worker costs.

An “agentic employee” has at least six cost layers:

1. The model layer
   Premium models cost more than lightweight models.

2. The usage layer
   More prompts, more tool calls, more sessions, more tokens.

3. The workflow layer
   Retrieval, routing, approvals, memory, and integrations all add overhead.

4. The governance layer
   Logging, security, policy, auditability, and data controls are not free.

5. The human oversight layer
   Escalations, reviews, exception handling, and QA still require people.

6. The failure layer
   Bad outputs, retries, and manual corrections are hidden costs that often show up later than the invoice.

That means your agent does not just have a salary.

It has a salary, benefits, training, compliance overhead, and a cost of being productive.

The more useful the agent becomes, the more it may cost.

That is the part teams are not psychologically ready for.

The question every company should ask now

If your AI employee becomes twice as useful next quarter, will it also become twice as expensive?

And if the answer is yes, can your business still make money?

That question is especially important in workflows where AI is not just assisting, but acting:

• customer support
• code generation
• invoice handling
• legal review
• sales qualification
• internal operations

These are exactly the kinds of workflows where usage can grow fast once the team trusts the system.

The cost curve does not stay flat.

What to do before the bill surprises you

Teams should stop budgeting AI like a fixed subscription and start budgeting it like an operating system with variable usage.

That means:

• define what each workflow is allowed to do
• estimate usage by session, not by feature
• separate lightweight tasks from high-cost tasks
• set budgets and overage rules before the pilot expands
• track the workflows that create the most value, not just the most activity
• make sure governance and data controls scale with usage

The best AI teams are not the ones that spend the least.

They are the ones that can explain every euro or dollar they spend.

The uncomfortable but useful conclusion

If your organization is betting on agentic AI, the real risk is not that it gets worse.

The real risk is that it gets better, gets used more, and quietly gets more expensive than anyone planned for.

That is why the next phase of AI adoption is not just about capability.

It is about cost visibility, policy, and control.

If your AI employee is going to ask for a raise, you want to know:

• what it produces
• what it consumes
• where it is allowed to work
• when it should stop
• and who approves the bill

If you want to bring AI into real workflows without letting cost or data exposure run away from you, a control layer matters just as much as the model itself.

That is where the conversation should go next.

See the Privacy Gateway overview if you want to think through AI control, or contact us if you want help shaping a workflow that stays useful when usage starts to grow.