Microsoft's new Agent 365 went generally available on May 1st. It costs $15 per user per month. That's on top of the $30/user you're already paying for Copilot. And the $57/user you're paying for M365 E5.
Stack it up: $102 per employee, per month. For a 25-person company, that's $29,700 a year before a single AI agent actually does anything.
Read that again. That $15/seat fee doesn't build agents. It doesn't run them. It gives your agents an identity in Entra and extends your compliance tools to cover what they do. The governance layer. The permission slip.
Actually running those agents? That's a separate bill. Copilot Studio credits at $200 per 25,000 credits, or per-token pricing through Azure AI Foundry. A 100-seat company running modest agent workloads is looking at $3,000 to $5,000 a month, not the $1,500 the headline math suggests.
Three billing layers for one capability. Welcome to the per-seat trap.
The math gets worse in July
Microsoft announced global price increases effective July 1, 2026. Business Basic goes up 17%. Business Standard, 12%. The E5 tier that Copilot sits on top of gets a 5% bump. Copilot itself stays at $30, but the foundation underneath it keeps getting more expensive.
The new E7 bundle consolidates M365, Copilot, and Agent 365 into a single $99/seat SKU. Sounds cleaner until you realize you're paying $99 per person per month for the right to have AI agents exist inside your organization. The agents themselves still bill separately on consumption.
For a 25-person team, that's $29,700/year in licensing alone. Before electricity. Before the API calls. Before someone misconfigures a bot and it burns through your credit budget in three days, which, by the way, is one of the top three drivers of surprise bills in this ecosystem.
Per-seat pricing punishes the thing automation is supposed to do
Here's the part that should bother you.
The entire point of automation is reducing the number of humans needed for repetitive work. AI agents handle the stuff you used to hire for. That's the pitch. That's why you're interested.
But per-seat pricing scales with headcount. The more employees you have, the more you pay. The exact metric that automation is supposed to shrink is the one Microsoft ties their billing to.
Gartner predicted that 70% of businesses will prefer usage-based pricing over per-seat models by 2026. The broader software industry is moving toward outcome-based billing -- charging for what gets done, not who's in the building. One AI sales tool charges per qualified meeting booked. No seats. No headcount. You pay for results.
Microsoft went the other direction. They added another per-seat layer on top of existing per-seat layers, taxing headcount at the exact moment when AI's value is reducing headcount.
The adoption numbers tell the story
Only 3.3% of Microsoft's 450 million commercial M365 seats have paid Copilot licenses. That's 15 million seats out of 450 million. 70% of Fortune 500 companies adopted it, but mostly as pilots, not full rollouts. 35% of companies are delaying because of compliance concerns.
The accuracy numbers are rougher. Recon Analytics tracked Copilot's accuracy Net Promoter Score at -3.5 in July 2025. By September 2025, it dropped to -24.1. Partial recovery to -19.8 by January 2026, but still negative. 44% of people who stopped using it said they didn't trust the answers.
Now Microsoft is selling a $15/seat governance product for AI agents in an ecosystem where the underlying AI assistant has a negative trust score. Gartner predicts 40%+ of agentic AI projects will be cancelled by 2027 due to escalating costs, governance gaps, and unclear ROI. Organizations may end up paying for governance infrastructure around agents they eventually shut down.
What small teams figured out
A small team running self-hosted workflow automation on a $20/month server, making API calls to a capable language model at commodity pricing, handles 200 tasks a day for somewhere between $24 and $720 a month depending on how hard they push the models. No per-seat fee. No governance layer fee. No credit packs.
That same 10-person team on M365 E7 pays $11,880 a year before agents run a single task.
The self-hosted approach scales with usage, not headcount. Add five more employees and your automation cost doesn't change. Process twice as many tasks and your cost goes up proportionally to what you actually used. That's how automation pricing should work.
Open-source workflow automation tools have matured significantly. Production-ready platforms exist that handle everything from AI agent orchestration to complex multi-step workflows. Self-hosted, no per-seat licensing, infrastructure cost only. LLM API costs dropped roughly 80% year-over-year, which means the per-task cost of running your own automation keeps falling while per-seat licensing keeps rising.
This isn't a Microsoft problem. It's a pricing model problem.
I'm not here to tell you Microsoft is bad. Their tools work for a lot of organizations. If you're a 5,000-person enterprise with an existing E5 deployment and a dedicated IT team, the incremental cost of Agent 365 might be a rounding error.
But if you're a 10-person agency, a 15-person manufacturer, a creator with three contractors -- the per-seat model is actively working against you. Every new hire costs more in software before they produce anything. Every AI efficiency gain is offset by the licensing cost of governing it.
The question isn't whether AI automation works. It does. We automated our entire daily content pipeline -- topic selection, research, writing, quality scoring, video generation, publishing, distribution across six social platforms. Runs on timers. No human in the loop. Our cost scales with how much content we produce, not how many people work here.
The question is whether your automation should have a headcount. Whether the system you built to replace repetitive labor should be taxed like an employee.
What to actually think about
Look at your current M365 bill. Add $15/seat for Agent 365. Add $30/seat if you don't already have Copilot. Then add the consumption costs for actually running agents. Compare that to what a self-hosted automation stack would cost at your scale.
For some companies, the integrated Microsoft ecosystem is worth the premium. For others, it's a $30,000/year governance layer for agents they could run independently for a tenth of that.
The per-seat model is a tax on growth. The companies that figured this out early built automation that costs the same whether they have 5 employees or 50.
We build automation stacks like that for clients. Purpose-built, no per-head toll, scales with usage instead of headcount. First conversation is free -- kief.studio/contact.