Four Seats for 90 Seconds of Work: Why AI Agents Break the Software Bill

On Dec. 2, 2025, Rajesh Jha walked up on stage at the UBS Global Technology Conference and delivered good news for software investors. Microsoft’s then-executive vice president said he saw no sign that AI was cutting paid seats in Microsoft 365. Instead, he said, companies will one day run more AI agents than people. Many of those agents will have identities, mailboxes and a place in the address book. “To me, all of those embodied agents are seat opportunities,” he said.

The remark went viral in April 2026. Its logic was simple: If agents replace workers, sell each agent a “seat,” the per-user license that most business software charges.

Antonio Bustamante says that logic breaks the moment you try to count. He is chief executive officer of Bem, a San Francisco startup that makes money by turning messy files, such as PDFs, scans and emails, into structured data that software can use. He describes a document-checking job. One agent reviews a file. It then launches three sub-agents, smaller helpers that each handle one task: pulling out data, checking compliance and cross-checking records. Each helper lives for 90 seconds.

“Do I now have four seats?” Bustamante wrote in an email interview with CDOTrends. “If those sub-agents only exist for 90 seconds, do I pay for a full month?”

It is a fair question from a man who sells the alternative. Bem charges by the function call, a metered fee for each task its software performs. But follow the money and the picture gets messier. The seat and the meter are already merging and neither one fixes the problem Bustamante says sits underneath: data that agents cannot yet use. That problem belongs to chief data officers, not procurement.

Agents don’t keep office hours

The seat model assumes an agent works like an employee. “One agent does one job with roughly linear productivity similar to how an employee shows up, completes a set of tasks, and goes home,” Bustamante wrote.

Agents don’t work that way. A claims agent might handle five claims one hour and 500 the next, he wrote. He says Bem works with insurers, health care operators and financial services firms. There, one claims process can chain five agents. One reads the PDF, another extracts the fields, a third checks policy rules, a fourth flags exceptions for a human to review, and a fifth sends approved claims to payment.

“None of those agents map to a person,” he wrote.

Workloads also rise and fall with the calendar. Take a health insurer during open enrollment, the period each year when customers choose or change their plans. Hundreds of agent workflows may run at once. Afterward, volume drops to a fraction. “Under a seat model, are they supposed to license and pay for 500 seats year-round for capacity they only use two months a year?” Bustamante wrote.

His alternative: “The unit of work isn’t the agent, it’s the compute and the outcome.” In plain terms, pay for the computing power used and the results delivered, the way companies already pay for cloud services.

That is the theory. The price lists tell a different story.

The seat Microsoft actually sold

Despite Jha’s remark, Microsoft did not put a price on every agent.

Its agent-management product, Agent 365, became generally available on May 1, 2026. It gives IT and security teams one place to see, govern and secure agents. It costs USD15 per user per month, and one license covers every agent that person owns, manages or sponsors. The agents’ actual work is billed on separate meters: Copilot Studio credits and Azure token charges.

So the bill already has two layers. A per-person seat pays for oversight and a meter pays for the work.

“You can’t orchestrate agents on top of chaos.” — Antonio Bustamante @ Bem

Salesforce reached the same place from the opposite direction. Its agent platform, Agentforce, launched in 2024 at USD2 per conversation. In May 2025, Salesforce added Flex Credits, prepaid units that cost about 10 cents for each action an agent takes. Then it added per-user licenses, starting at USD125 a month, a flat fee that finance teams can budget.

Two of the biggest software vendors now charge for agents in the same way: a fixed fee per person, plus a meter for the work.

Bustamante calls the seat pitch “brilliant defensively.” His reasoning: seat revenue grows with headcount. If AI lets companies shrink headcount, charging for agents protects that revenue. “Who loses are enterprises locked into pricing models that penalize them for scaling AI efficiently,” he wrote.

Analysts expect the shift away from pure seats to continue. Gartner predicts that at least 40% of enterprise software-as-a-service spending will move to usage-, agent- or outcome-based pricing by 2030.

A meter sounds fairer. But it is also harder to predict.

The meter grows a seat

Bustamante admits finance teams will resist at first. “Finance teams hate surprises,” he wrote. His fix is to treat AI agents like cloud hosting. Learn what each task costs. Set usage budgets and alerts. If you process 50,000 claims a month and know the cost per claim, you can forecast the bill.

That works only if you know the number. Many companies don’t. A KPMG survey of 2,145 senior leaders, published in July, found that 29% struggle to understand their AI operating costs as they scale. Nearly half had changed the timing of AI rollouts when costs outran the expected value.

Uber shows how fast costs can run. The company gave about 5,000 engineers access to AI coding agents in December 2025. By April 2026, its chief technology officer said the full-year budget for those tools was gone. Uber’s response was a cap of USD1,500 per employee, per tool, each month.

Look at the shape of that fix. The meter’s first emergency patch was a budget per head. The logic of the seat crept back in.

Bem’s own pricing has a milder version of the same trap. Paid usage starts at 9 cents per function call. Sort a document, then extract data from it, and you pay for two calls. The document didn’t double; the number of steps did.

That detail matters more than it looks. Pricing does not just set the bill. It shapes the system.

How you pay decides what you build

This is Bustamante’s most consequential insight for CDOs. When a pricing model does not match how a technology works, he says, it changes what companies build.

Under seat pricing, he expects buyers to cram work into as few agents as possible. “Companies will over-invest in monolithic agents that try to do everything because it’s cheaper to license one seat than ten,” he wrote. Vendors, he predicts, will ship “super-agents” that do several jobs badly instead of specialists that do one job well. “That’s bad architecture because it makes systems brittle and hard to maintain.”

What gets left behind is the design he calls “the right way to build scalable systems”: lightweight agents that start when needed and shut down when finished. Testing suffers too. “If spinning up an agent to test a workflow costs you a seat license, enterprises will hesitate to try new things and innovation slows down,” he wrote.

The argument holds, but it cuts both ways. A per-seat bill punishes many agents. A per-step bill punishes many steps. Either way, the price list starts to act like an architect. Data, AI and security chiefs should know which price list is drawing their system.

Yet for most enterprises, Bustamante argues, the pricing debate is running ahead of reality.

Chaos is the real line item

Every pricing model, seat or meter, rests on one number: what the work actually costs. Finding that number is a data job. Someone has to track the cost of every step, every retry and every human review, then tie it to a finished claim. That work falls to the data team, which makes pricing a CDO’s problem as much as procurement’s.

But first, agents need data they can read. “Most enterprises are nowhere near ready,” Bustamante wrote. He describes claim data trapped in a 47-page document with inconsistent formatting, and loan applications with missing fields and illegible handwriting. “You can’t orchestrate agents on top of chaos.”

The research points the same way. IDC, in research with Lenovo, found that for every 33 AI proofs of concept a company launched, only four reached production. A proof of concept is a small trial built to test an idea. Bustamante blames the gap on two things. The first is data: Trials run on clean data and simple workflows but real operations are messy.

The second is liability. In a trial, a wrong answer is a lesson. In production, “if the AI approves a fraudulent claim or denies a legitimate one, someone is liable,” Bustamante wrote.

“The unit of work isn’t the agent, it’s the compute and the outcome.” — Antonio Bustamante @ Bem

In his view, that someone is the buyer. “Responsibility sits with the enterprise because the agent is a tool,” he wrote. It is a convenient position for a vendor, but the practical point stands. If a regulator asks why a system denied a claim three years ago, the company needs an audit trail: the data the system saw, the logic it followed and the reason for its decision. “The seat model doesn’t give you that,” Bustamante wrote.

Neither does a meter. A bill counts work. but doesn’t explain a decision.

Europe has set a deadline for that test. Under the EU AI Act, AI systems used to judge creditworthiness or price insurance face high-risk rules from Dec. 2, 2027. Those rules include keeping system logs.

So what should a buyer ask before signing?

Four questions for the vendor across the table

Bustamante suggests three.

First: How do you define a seat? “If the vendor can’t give you a clear answer about what counts as a billable agent,” he wrote, the vendor hasn’t thought through the economics, “and you’re going to get hit with surprise costs when you scale.”

Second: What happens to my costs if I double my workload without adding staff? If the seat count doubles too, the price is working against the efficiency AI is supposed to deliver.

Third: Can you show me a complete audit trail for every decision? “If not, then you’re buying a black box,” he wrote.

Add a fourth, and ask it of metered vendors too, Bem included: What does one finished outcome cost, with every step counted?

Neither a seat nor a meter answers that question. Only your own data can. A buyer who knows the cost per claim can judge any pricing model, whatever the vendor calls the bill.

Image credit: iStockphoto/Pressmaster