As AI moves from experimentation to enterprise-wide deployment, CIOs and CTOs are feeling the pressure to deliver. 80% say they have technology-transformation mandates coming directly from their CEOs, but two-thirds admit they're being held accountable for AI systems they do not fully control.

Enterprises are expected to deploy an average of 1,661 AI agents by 2027, 38% more than today. But the majority (70%) of technology leaders say teams across their business are deploying technology faster than IT departments can track. Only 11% believe they are fully prepared for the scale of deployment expected over the next year.

These numbers come from a new IBM report 'Redefining the tech leader's mandate: Building the IT foundation for agentic AI at scale'. Based on a survey of 2,000 senior executives responsible for IT, technology or AI-related decision-making, the report suggests that one of the main challenges facing enterprises as they adopt AI at scale is that technology infrastructure, governance and investment processes were designed for "human-speed" change, while AI increasingly operates at a faster "machine-speed".

As a result, there's a widening gap between the pace of AI change and many organisations' ability to respond to it. Closing this gap will require businesses to redesign their architecture, their governance and controls and the way they manage technology investments.

Here are three key lessons from the report.

1. AI thrives when infrastructure is adaptable

The current infrastructure in many enterprises lacks the flexibility to adapt to change. That's because it was built for a slower, more predictable world, not an AI-driven landscape where change can happen fast and new opportunities and challenges appear with very little warning.

Organisations find it hard to move workloads, switch to better AI models, update workflows or bring in new capabilities when they need to respond to changes. As they scale from hundreds to thousands of AI agents, this inability to adapt starts to become a strategic problem.

The report suggests 88% of organisations are attempting or planning to move workloads to a different cloud provider, but technology leaders are finding that only 25% of those workloads are easily portable. The main barriers to data portability are data transfer costs or egress fees (69%), reliance on provider-specific services (61%) and the sheer technical complexity involved (52%).

The report does not suggest that every workload needs to be portable. Instead, it advises organisations to build what it calls "optionality" where it matters most: focusing effort on workloads, data and models where dependence on a particular provider or system could restrict important future business choices.

It's also important to always carefully review potential new services or deployments that support AI initiatives to ensure they don't rely heavily on proprietary platforms that might be hard to move away from, and clearly understand the costs of data egress and migration before making a decision.

For strategically important AI systems, infrastructure should be designed to make it as easy as possible to switch cloud providers, rotate AI models and embrace new capabilities without requiring re-platforming or major re-engineering.

The potential benefits are significant: adaptable organisations that are designed for greater optionality reported 10% higher returns from AI investments in 2025.

2. Manual governance has to be replaced with 'governance by design'

Governance and control in enterprise technology have traditionally relied on setting policies by committee and manual approval processes and review cycles. That worked well when systems moved at human pace, but isn't appropriate anymore with AI enabling systems to move at machine speed. Manual governance approaches make it harder to scale AI systems.

If AI agents are making thousands of autonomous decisions a day, control can no longer rely on human approvals. It has to be embedded directly into the technology architecture.

The advice is to use "executable" guardrails for AI, to enable the technology to operate and make observable decisions within set boundaries. The guardrails should help to determine and control what actions agents can take and when they must stop. All AI decisions need to be explainable across the AI lifecycle.

Practical steps mentioned in the report include establishing requirements for observability, auditability, rollback and ownership of AI systems, and building "mechanisms" to ensure these requirements have been met before systems are deployed.

Modernising governance and control in this way can make it easier to scale AI. Organisations that were classified as having "orchestrated control" were found to deploy 16x more agents than those relying on manual governance. They also spent 4x less of their AI budget and delivered 18% higher operating margins.

3. AI investments have to be managed as a portfolio

Because of the short lifecycles of AI models and the difficulties in predicting the returns from AI initiatives, investment in IT needs to be managed differently today.

Traditionally, technology investments were easier to predict. Their returns were likely to be stable, and they could be depreciated over time, often over several years. Their performance could be reviewed periodically according to fixed budget cycles. Good cost discipline was one of the main factors in getting the most out of those investments.

Not any longer. AI has introduced a class of technology investment that can have short lifespans and deliver returns that are uneven and very hard to predict. The average useful life of an AI model is around 14 months, suggests the report, which is short enough to require model refresh decisions to be made continuously.

While some AI investments might come with a proven track record, others will be more experimental. Some could fail quickly while others scale rapidly. Which is why companies must now view and manage their investments as a portfolio rather than trying to optimise each one individually.

Success is based on quickly identifying which investments in the portfolio are working and which are unlikely to succeed, and reallocating resources accordingly. Failure to do this risks funding projects that fail for too long while not scaling winning initiatives quickly enough.

Unfortunately, the financial visibility needed to make informed investment decisions like this isn't always available; 84% of technology leaders haven't "fully operationalized" AI financial management, according to the report. 85% lack full visibility into real-time AI spend.

Key information such as spend, model, business objectives and the expected return for each project needs to be available by use case to successfully manage AI technology portfolios. Organisations need to put in place recurring reviews of all investments. They should reassess funding, make changes where necessary and decide which projects to exit and which to support further as more evidence becomes available.

The potential rewards can be significant: organisations with strong financial discipline deploy 2.4x more AI agents with no higher AI/IT budget. And they're 3x more likely to say they are fully prepared for scaling AI.

Three interrelated requirements to help scale agentic AI

The three areas highlighted in the report are closely connected. An adaptable infrastructure allows organisations to embrace new opportunities quickly as they arise. But they also need effective governance to help them cope with the rapid pace of change that comes with AI; otherwise, that speed can increase risk. And along with adaptability and governance, organisations will need access to up-to-date financial information and performance data to manage AI initiatives as a portfolio. With this they can ensure resources move forward towards the systems and use cases that deliver most value.

This blog was originally published on the IBM Community.