Agentic AI is changing more than how software gets built. It is changing what companies should buy, what they should build, how their application landscape needs to operate, which legacy systems are worth modernizing, and how IT work gets done.
One constraint runs through all these decisions: a lack of codified knowledge. The organizations that can capture what their systems do, how their business rules work, and where critical expertise sits will be better positioned to put agents to work without surrendering control.
Few organizations are ready to take on these tasks today. Our survey, The Oliver Wyman 2026 Parallel Surverys of CEOs and Tech Laders on Agentic AI and IT shows only 6% of CIOs say agents can reach governed, agent-ready data across most domains, and only 15% say someone in their organization could within a working day establish which requirements in the active portfolio conflict or are incomplete.
The economics of build versus buy are being repriced with agentic AI
Cheaper software development does not solve the build-versus-buy decision, although it changes where the decision matters. The same AI models that make building easier also allow software vendors to deepen their products and strengthen their positions.
Enterprise AI reached roughly $37 billion in spending in 2025, with 76% of use cases bought from vendors. That was up from 53% a year earlier.
CEOs are not betting on vendor access to carry them. Almost three-quarters expect durable advantage to sit in reengineered processes and operating models, and 64% in proprietary data and customer context, compared with 4% who point to privileged access to a leading model or vendor.
The first decision IT leaders must make is which parts of the agentic stack a company is prepared to hand to a vendor. Four layers matter: workflow, memory, interface, and orchestration. For each application, leaders need to ask whether it is differentiating, whether total cost still favors building once maintenance is included, and whether the resulting dependency can be reversed.
Building also no longer guarantees independence. It can simply move the dependency. one layer deeper. That, too, needs to be part of IT’s calculation.
Reversibility is largely unwritten. Only 10% of CIOs have a comprehensive operating plan to switch providers if a model supplier doubles prices or withdraws a version; 46% are developing one, and 41% have nothing. Yet few acknowledge the risk this creates, with less than half even listing supplier and model concentration on their enterprise’s risk register.
On an illustrative basis, disciplined sourcing can create a cost-and-optionality advantage worth roughly 3% to 6% of the addressable software estate each year compared with locking in. A credible ability to build can also create leverage in vendor negotiations.
The objective is not to build more. It is to know where control creates value — and preserve the option to exercise it.
Agents are becoming a new class of actor in enterprise IT
Agents need to become a third class of actor in the technological stack, alongside users and systems. That requires four shared foundations: agent identity, a gateway, observability, and isolated execution. Companies can build them around one real workflow, then extend them use case by use case.
Most technological estates cannot yet see the agents they already run. Only 10% of CIOs can produce an agent's owner, scope, cost, and last certification date on demand, while 58% have some of those attributes and 29% have no central registry at all. Just 25% can reconstruct what an agent changed, when, and who authorized it from a tamper-evident record.
At the same time, modernization priorities need to shift. Agentic AI rewards a well-documented application landscape, not simply a modern one. That means directing part of the modernization budget toward documentation: defining data meanings, extracting business rules, and creating tests that pin down current behavior.
Portfolio decisions change as a result. Leaders need to know whether an application is codified, whether agents can reach it, and how much of it agents can absorb.
Applications can then dissolve by layer. The workflow shell may disappear while the core system of truth remains, serving a new set of consumers.
How agentic AI is changing legacy system modernization
AI is changing the economics of legacy modernization, but not every legacy system should follow the same route.
For aging Java and .NET applications, obsolete frameworks, and old desktop code, AI-assisted migration can be industrialized: prove one migration pattern, measure how much human correction the converted code requires, and use that evidence to decide whether to scale it across the portfolio.
Mainframes demand a different standard. For mission-critical transactional systems, the challenge is not simply generating replacement code. It is proving equivalence.
While almost two-thirds of CIOs run a mainframe or legacy core, almost half of CEOs consider the legacy technology a significant enough constraint on progress to justify replacing it. But among those planning agentic modernization of the core, only 20% have both built a behavioral reference dataset and begun tracking verification costs. Forty-five percent have built only the dataset.
Agentic AI has made it more affordable to understand and translate decades-old software. It has not removed the burden of proving that the replacement behaves identically across every input — including edge cases that may exist only in the memories of experienced people.
That puts a premium on timing. People who understand legacy systems are retiring, while software license renewals create natural economic windows for change.
Capturing institutional knowledge before it leaves the organization can therefore be as important as the modernization technology itself. And modernization does not always mean migration. The cheapest application to modernize may be the one the organization can archive and switch off.
Redesigning the IT workforce for agentic AI
The goal is not an IT organization essentially run by AI with a few humans to make sure things keep moving. CIOs should be aiming for a supervised IT model in which people set intent, establish quality standards and escalation rules, and direct agents that perform more of the execution.
That will change the shape of the workforce, reducing entry-level positions and filling remaining roles with more supervisory technicians able to deploy and run agents. The constraint is capability, not cost, with 68% of CIOs naming the ability to find and reskill talent to work with agents as their primary talent constraint versus only 3% who said they would be looking to replace or reduce engineers.
Pure execution work shrinks, particularly at the entry level. Developers move toward orchestration and review, judging intent and outcomes rather than syntax alone. New roles emerge, including agent orchestrators, validation designers, and senior AI governance leaders.
Team structures change, too. Traditional squads are likely to give way to smaller, more senior teams directing fleets of agents. But capturing productivity too early creates a risk.
Reducing headcount before roles, development pathways, and talent pipelines have been redesigned may deliver one-time savings, but will likely weaken the organization's supervisory capability and talent pipeline over time.
The junior pipeline needs particular attention. Agents are now absorbing precisely the tasks that historically trained tomorrow's senior talent. If an organization chooses not to train a bench, it will have to compete for qualified workers in the broader labor market, a strategy that could prove increasingly difficult if too many enterprises take the same route.
Entry-level hiring is already declining, with almost one-quarter of CIOs claiming to have cut junior engineer hiring. Across agent-exposed functions, 17% of CEOs expect layoffs and another 11% expect to see reduction through attrition.
Creating a sustainable workforce therefore requires workforce redesign to become a joint project for both CIOs and chief human resources officers (CHROs).
But almost no one runs it that way today, with only 16% of CEOs saying they had charged their CIO and CHRO with a single joint mandate. On the CIO side, 25% have a single workforce redesign plan, covering both humans and agents, with one accountable owner.
Why Agentic AI strategy requires a new enterprise operating model
Sourcing, architecture, modernization, and workforce strategy can no longer be treated as separate decisions.
Architecture determines where agents can operate. Codified knowledge determines what they can understand. Sourcing determines where control sits. Workforce decisions determine whether the organization can supervise what agents produce.
Ownership is the gap to close first, with 45% of CIOs and 57% of CEOs confirming that no executive owns the task of redeploying capacity freed by agents, even as that capacity begins to emerge. Only 3% of CEOs are very confident that they could reconstruct, and stand behind, what an autonomous system did on their behalf in a customer or regulatory setting.
Agentic AI may lower the cost of producing technology, but haphazard workforce reductions could expose enterprises to considerable risk. The harder leadership question is how to redesign the enterprise around the new agentic operating model without giving away control, knowledge, or future capability. On that measure, many oranizations still have work to do.
