How AI Agents Are Priced: Per-Seat, Per-Task, and the Microsoft Copilot Shift

On June 16, 2026, Microsoft made Microsoft Copilot Cowork generally available, and in doing so it quietly changed how a huge slice of the software market thinks about price. Cowork is an agentic assistant: instead of suggesting an edit while you type, it independently drafts documents, builds spreadsheets, and sends emails on your behalf. The headline was not the feature set. It was the bill. Microsoft moved Cowork to usage-based pricing denominated in Copilot Credits at one cent each, charged per task based on the compute each task consumes. Microsoft itself called this its first major pricing model change in nearly two decades.

That single move is a useful lens for understanding a broader shift. AI agent pricing is splitting into several distinct models, and the differences matter a lot once an agent starts doing real work at volume. This explainer walks through the main models, why agents specifically keep drifting toward usage-based billing, and what to check before you sign anything.

What Microsoft Actually Did

Copilot Cowork bills per task, and tasks fall into rough tiers. A light task (think a quick email draft or a short summary) runs roughly 100 to 300 credits, or about one to three dollars. A medium task lands around 400 to 700 credits. A heavy task, such as assembling a multi-tab spreadsheet from scattered source files, can run 700 credits and up, meaning seven dollars or more for a single job.

Three things are worth flagging. First, Cowork still sits on top of a paid Microsoft 365 Copilot subscription, so the credits are an additional, variable layer rather than a replacement. Second, the model you run changes the cost profile: at general availability, Cowork uses Anthropic's Opus 4.8 and Sonnet 4.6, Frontier customers can reach GPT 5.5, and Microsoft has signaled a cheaper Cowork 1 model aimed at everyday tasks. Third, the per-task framing means your monthly spend now tracks how hard your team leans on the agent, not how many people hold a license.

The Four Pricing Models You Will See

Across the agent market, four models dominate.

Per-seat or subscription. This is the legacy SaaS approach: a flat fee per user per month, often with tiers. It is predictable and easy to budget, which is why finance teams like it. The catch is that it decouples price from usage entirely. Tools built on assistants such as Claude or ChatGPT often start here because it is familiar, but it fits poorly once an agent does heavy autonomous work.

Usage-based, per-task, or per-credit. This is exactly what Microsoft just adopted. You pay for what the agent consumes, metered in tasks, credits, tokens, or runs. Automation platforms have priced this way for years: Zapier charges by task, Make charges by operation, and n8n meters workflow executions. The model rewards light users and scales the bill with real activity, but it makes forecasting harder.

Outcome-based or per-resolution. Here you pay for a result rather than the work behind it. Support agents are the clearest example: some vendors charge per resolved ticket, so you only pay when the agent actually closes a case. This aligns the vendor's incentive with your outcome, but it requires both sides to agree on what counts as a resolution, which is not always clean.

Hybrid. Most serious platforms end up here: a platform or seat fee for access plus usage charges layered on top. It blends a predictable baseline with variable consumption. Microsoft's setup (a paid Copilot subscription plus per-task credits) is a hybrid in practice, even though the credits are the newsworthy part.

Why Agents Push Toward Usage-Based Billing

The drift toward metered pricing is not a fad. It comes from the underlying economics. Every time an agent runs, it consumes a variable amount of large language model compute. A one-line reply and a fifty-step research-and-build task can differ by orders of magnitude in tokens, tool calls, and model time.

Flat per-seat pricing cannot absorb that variance gracefully. If a vendor charges one price per seat, a light user who runs three short tasks a month effectively subsidizes a power user who triggers hundreds of heavy jobs. The light user is overpriced and the heavy user is underpriced. At small scale nobody notices. Once agents run autonomously and at volume, the gap becomes a real margin problem for the vendor and a fairness problem for buyers. Metering the actual work is the natural correction, which is why even an incumbent as seat-anchored as Microsoft moved.

The trade-off is that cost becomes harder to predict. A per-seat bill is the same whether usage doubles or halves. A per-credit bill follows demand, which is great when usage is low and alarming when an agent loops or a team suddenly adopts it for everything.

Practical Advice for Buyers

Usage-based agent pricing rewards preparation. A few habits keep it from biting.

Estimate task volume before you commit. Map out how many light, medium, and heavy tasks your team realistically runs in a month, then price it against the tiers. With Cowork's numbers, a team running fifty heavy tasks a week is in very different territory from one running a handful of summaries. Do the arithmetic with your own usage, not the vendor's demo.

Watch for credit cliffs. Per-task pricing can hide step changes: a task that usually sits in the light band can jump to heavy when inputs grow, and a cheaper everyday model (like the planned Cowork 1) versus a frontier model can swing the per-task cost several times over. Know which model runs which job.

Set spend caps and treat AI cost as governance. The same discipline that applies to cloud compute now applies to agent compute. Databricks is among the platforms shipping AI spend controls so teams can cap and monitor agent costs rather than discovering them on the invoice. Budgets, alerts, and hard limits should be table stakes for any usage-based agent rollout. If you are formalizing this, our guide to the best AI cloud cost optimization tools covers the governance side in more depth.

Match the model to the workload. Per-seat still makes sense for predictable, light-touch assistants. Usage-based fits bursty or heavy autonomous work. Outcome-based shines when the result is clean and measurable, like resolved tickets. Hybrid is the safe default for most teams that want a predictable floor with room to scale.

For a wider view of the landscape, see our roundups of the best AI agents and the best enterprise AI agents, which dig into how individual vendors structure their plans.

The Bottom Line

Microsoft moving Copilot Cowork to per-task credits is a signal, not an outlier. As agents do more independent work, the cost of that work becomes too variable to bury inside a flat seat fee, so pricing follows consumption. The winners among buyers will not be the ones who avoid usage-based pricing. They will be the ones who measure their task volume, understand which model runs which job, and put spend controls in place before the agent starts working overtime. Price the work, not the seat, and budget like it can scale, because it can.