The new economics of AI: Pay for outcomes, not effort

Ask three straightforward questions and watch the boardroom go quiet:

1. How many AI agents are live across your business today?
2. Who's accountable for what they cost?
3. What's the true cost of one problem an agent resolves end to end?

Few executives can answer all three. But that’s not because of a gap in technology. It’s a sign that AI has rewritten the enterprise's economics while the operating model stood still.

Unsurprisingly then, it’s expected that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and weak risk controls.

Generally, the technology works fine. What fails is the way organizations buy, govern and account for it.

For generations, service logic was simple. More work meant more people, and more people meant more cost.

AI severs that link. Spending moves from salaries to platforms, consumption and tokens, and most finance functions are pulling levers that no longer work.

Unit prices for AI keep falling, yet enterprise bills keep climbing, due to surging consumption. Every long context window, retry, human escalation and rework compounds the problem. A cheaper model that fails twice as often costs more, once you count the cleanup.

So, the critical number is cost per resolved outcome, not cost per token.

Pilots make this easy to miss. A closely supervised trial looks less expensive because the person watching every step quietly absorbs the errors and retries before they show up as cost. At scale, that supervision thins; those same failures surface as tokens, rework and escalations you pay for directly, and the economics that convinced everyone in the pilot can break down in production.

Model the scaled case before you commit.

Not all AI carries the same value or the same risk:

• The agent assists while a human works.
• It recommends, and a human approves each action.
• The agent acts with human monitoring.
• It’s autonomous, running on its own with evidence to prove the work was done correctly.

Value rises with every step, and so does the burden of proof. Autonomy you can trust in work that matters is worth paying more for. Autonomy you can't evidence should be priced at the cost of being wrong. This is where promising deals stall, clearing the pilot then dying in procurement because the evidence never appears.

Business teams are already switching on agents inside their SaaS platforms without telling IT. These “shadow agents” can act in production and quietly run up costs. Gartner puts only 17% of organizations at real agent deployment today, while more than 60% intend to get there within 2 years. Clearly, the gap between intention and control is about to widen.

The most useful model is the employment lifecycle. A capable agent needs:

• A job description and a clear task boundary
• A named owner and a sensible supervision ratio
• A person who owns the consumption bill that runs up each month
• Continuous assessment, not a one-time go-live test
• A termination process: a clean way to be deactivated that revokes agent access while retaining what it learned

An agent you can't observe is really an unmonitored employee with live access to production. That's why control, not raw capability, is the true constraint on scale. You can't safely buy an outcome you can't measure or defend one you can't trace. A living registry of every agent (owner, scope, entitlements, model and blast radius) should be as routine as the controls you apply to privileged accounts, keeping obligations like the EU AI Act manageable.

DXC Technology reached these conclusions by operating agents inside mission-critical estates, not by theorizing about them.

DXC OASIS (our agentic platform for enterprise IT operations built with Anthropic) now runs across more than 50 customers. If Skadeförsäkring AB, the largest property and casualty insurer in the Nordics, chose OASIS to unify and modernize a fragmented estate across Finland, Sweden, Denmark and the Baltics after a major acquisition. That’s a complex situation where uncontrolled automation only complicates matters.

The point is, no enterprise starts with a clean slate. So, anything that expects you to consolidate tooling first won't survive contact with reality. The goal isn't to automate everything, because the hard part is deciding where to keep human decision making. Measurable return has to be designed in from day one, since it's almost impossible to add later.

The proof is in the outcomes. DXC's Agentic Security Operations Center (SOC), built with 7AI, puts autonomous agents on triage, investigation and response for 3,200 security professionals, 12 centers and 4.5 million threats a day. Roughly 16 times a day, it sees rare problems that most enterprises meet once a decade. As a result, we improve faster than any client can; DXC has already returned hundreds of thousands of analyst hours to higher-value work.

DXC's Xponential blueprint tells the same story, cutting service desk tickets by 20% for 32,000 Textron employees, deploying more than 30 governed agents at infrastructure group Ferrovial and helping Singapore General Hospital reach 90% accuracy in antibiotic decisions. The point is never the model but the resolved outcome, evidenced.

This isn’t about technology deployment. It's an operating model change, and the first moves are unglamorous:

1. Build the agent registry and find the shadow agents.
2. Give the consumption bill a named owner and a forecast.
3. Model cost per resolved outcome on one real process at scale.
4. Define your autonomy ladder and the evidence each rung needs.
5. Rewrite one procurement template to ask for outcomes, not headcount and rate cards.
6. Start the workforce conversation and education early, because redeploying the people who hold your context beats replacing them.

Speed matters more here than in most transformations. Agents improve as they accumulate context, so every month of delay is a month of context that you never collect. The gap between fast movers and slow movers doesn't stay constant; it widens.

The real question for your next leadership meeting isn't whether to begin.

It's how much ground you're willing to concede while you wait.