Why Managed Agents Are the Future

Back-office labor moved to outsourcing providers once running it well stopped being an advantage and became a cost. Servers moved to the cloud for the same reason. Both moves built markets worth hundreds of billions of dollars. Neither removed the function. Both relocated who runs it.

Agentic AI is repeating the pattern, in two years rather than two decades. The usual explanation is that agents will be sold by outcome, and that buyers will pay only for results. That is half right, and the wrong half is the one most vendors are building on. What wins is a managed layer. Someone is accountable for the finished work. The buyer gets a bill they can forecast. A human stands behind the agent. Pricing follows from that. It does not lead it.

The gap: companies want agents, but few can run them

The budget has moved. The capability has not.

Software used to organize work. People still did the job, only faster. The seat was the unit of sale because the human was the unit of work.

Agents break that link. When software can read the invoice, chase the payment and draft the quote, the buyer is no longer purchasing a tool. They are purchasing a result. Foundation Capital, a venture firm, calls this services-as-software and puts the prize at $4.6 trillion. Treat the figure as an investor's estimate. The direction is the point. The target is no longer the IT budget. It is the money spent on people and on outsourced services.

The capability is another matter. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The same firm calls most current projects early experiments, driven by hype and often misapplied. This is not a maturity curve that resolves itself with more internal effort. It is the gap that outsourcing and the cloud closed before: high value, high complexity, a poor success rate for the firm that tries it alone.

The signal: buyers want finished work, not outcome risk

The tempting conclusion is to charge only for results. Buyers do not want that.

It sounds fair. The vendor shares the risk. The buyer pays for value.

Buyers disagree. In a survey of 296 software buyers this spring, 89 percent said they had exceeded their AI budget. Seventy percent ranked cost unpredictability among their top three worries. Asked which pricing shape they preferred, they rated a base fee plus usage highest, at a net preference of +17. Pure outcome pricing came last, at -19. A little over half would accept a higher fixed cost if it made an essential AI product fully predictable.

The reason is plain. A pure subscription undersells what an agent creates. A pure outcome fee is too volatile for a CFO's fixed budget. It moves the variability onto the buyer's forecast. Volume swings. Definitions of success get argued. Finance cannot budget a number that depends on how well an agent performs next quarter. What the market is converging on, a fixed core with usage or outcome expansion, is not a new idea. It is a managed-services contract arriving in a new category.

The best-known example agrees. Sierra, which sells customer-service agents per resolution, publishes no rate. Third-party guides estimate between $1 and $2.50 a resolution, and Sierra has not confirmed it. The same guides report contracts that combine a platform fee, per-outcome charges and services. The flagship outcome-priced vendor sells, in practice, a hybrid.

There is a risk on the vendor's side too. A vendor that takes outcome risk on things it does not control is being paid for luck. A meeting booked is a result a vendor can influence. A deal closed is not.

The shape: what "managed" adds

An agent is not a finished product. It is a system that drifts.

Models change. Source documents change. A supplier alters its invoice format. Accuracy slips and nobody notices. Someone has to watch.

Managed means that someone is the vendor. Four things separate a managed agent from a license. First, the vendor sets an accuracy target before launch and tests the agent against the buyer's real documents. Second, a person signs off on anything that matters, such as a price gap or a large discount, before it posts. Third, every input, step and approval is logged, so an auditor can follow the trail. Fourth, one contract covers the agent, the integration and the operation. When the agent errs, the fix is the vendor's problem.

None of this is glamorous. It is the part mid-sized firms cannot build. They have no AI department and no wish to hire one. The pattern is familiar. Payroll moved to specialists. Servers moved to the cloud. Running reliable agents is the next discipline to leave the building.

It is also why managed beats self-serve for most buyers. Software that is cheap to buy and expensive to supervise is not cheap.

The verdict: three questions before you buy

The test is not the pricing page. It is the contract.

Who is accountable when the agent is wrong, and is that written into the agreement? What is the unit of finished work, and can the buyer verify it in their own systems? What happens to the bill if volume doubles?

A vendor who answers all three plainly has built a managed service. One who cannot has built a license with an agent attached.

The future is not outcome pricing. It is accountable work at a price the buyer can forecast. The open question is not whether agents will be managed. It is who earns the trust to manage them, and that gets settled client by client, not in a pricing model.

Sources: Gartner (June 2025); Foundation Capital; CloudTalk and eesel on Sierra pricing; a survey of 296 software buyers (February to April 2026).