Every year, boards approve bigger automation budgets. Every year, CEOs sign off on the next wave—robotic process automation (RPA), then machine learning, then generative AI, now agents. And every year, the savings fall short. Not catastrophically, not enough to kill the programs, but consistently, quietly, and by a margin that should be making executives uncomfortable.

Bain & Company’s survey of 951 global companies finds that while 37% targeted cost reductions of 11% to 20%, nearly 40% of those who measured outcomes landed in the 0% to 10% bucket instead (see Figure 1). The technology worked. The value didn't arrive. And rather than pausing to understand why, 90% of those same companies are now increasing their budgets again—this time for AI agents that will operate with even greater autonomy, complexity, and consequence.

Figure 1

But here's what the same data also shows: A meaningful group of companies is breaking the pattern. They are realizing the savings they targeted, deploying agents with genuine confidence, and funding the next wave from returns that actually materialized. They didn't get there by finding better technology or bigger budgets. They got there by treating data access, governance, and process redesign as CEO-level problems rather than IT problems. The gap between these companies and everyone else is widening. Understanding what separates them starts with three uncomfortable truths.

The agents aren't actually autonomous

Ask most executives about AI agents, and they'll describe a near future of autonomous systems handling complex decisions end to end. The data tells a more grounded story. Only 7% of companies are running fully autonomous agents in production today. The dominant model—cited by 38% of respondents—is "human approval required." Another 32% operate with guardrails and exceptions, meaning a human steps in whenever the agent encounters something it can't handle confidently (see Figure 2).

The problem isn't caution. Human oversight of consequential automated decisions is exactly the right posture right now. The problem is the gap between what the investment case assumed and what is actually running. If your agentic AI business case was built on the economics of full automation and the reality is a system routing a significant share of decisions to a human queue, the CFO approved one set of numbers, and the organization is living with another. This gap is wider for companies that missed their targets: Only 38% of companies that fell short have agents at guardrails-level autonomy or above, compared with 50% of those that delivered. Rather than rush to full autonomy, the imperative here is to close the gap between the business case and the operating reality, and to be honest about the economics of what is actually in production vs. what was promised.

The next wave is being funded by returns that haven't arrived

There is a second financial risk hiding in plain sight. When asked how they plan to fund generative AI and agentic AI investments, 44% of companies—the largest group—cited savings from prior automation programs (see Figure 3). Self-funding the next wave from past returns sounds like discipline. In reality, it is a circular bet with a structural leak.

Figure 3

The prior wave underdelivered. The savings pool is smaller than assumed. And the investment case for the current wave was sized against projections rather than actuals. This is not an edge case; it is the pattern our Automation Scorecard research has documented across every recent wave of automation. Companies that don't validate their reinvestment math against what automation actually returned, rather than what it was supposed to return, are compounding risk rather than managing it.

Data is still the wall—and it’s not coming down

Our survey finds that data access and integration is the single biggest barrier to AI progress, cited by 41% of respondents—above compliance concerns, budget, skills gaps, and executive buy-in (see Figure 4). Despite a decade of investments in data modernization running well into hundreds of billions of dollars globally, the No. 1 reason AI programs underperform is that companies cannot reliably get access to their own data.

Figure 4

Notably, the companies that delivered on their targets cite data as a bigger barrier than those that missed—44% compared with 40%. They have not solved the problem, but they have run into it harder because they are deploying at scale. The underperformers, by contrast, cite more organizational obstacles: insufficient budget, lack of a Center of Excellence, and competing priorities. These are not technology problems. They are signals that AI has not been given the mandate or the executive attention it requires. The leaders that outperform didn't solve the data problem faster; they stopped treating it as an IT problem and made it a board-level business prerequisite. Critically, they also stopped using data imperfection as a reason to defer action entirely.

Implications for business leaders

The companies on the right side of this performance gap made a small number of specific organizational decisions. For leaders still on the wrong side, those decisions remain available.

1. Pay down your workflow debt before deploying AI. The single most costly mistake in AI deployment is automating a broken process. Every organization has accumulated workflow debt—the unnecessary handoffs, redundant approvals, and embedded workarounds that make processes slower and more expensive than they need to be. AI doesn't fix workflow debt; it locks it in, speeds it up, and makes it vastly more expensive to unwind. The question to ask before any AI program is approved is not "Where can we apply AI?" but "If we were designing this process from scratch today, what would it look like?" Only then should the technology conversation begin.

2. Validate the investment case and name a governance owner before programs launch. Before approving the next wave of AI spending, CFOs should audit actual returns from prior automation programs, not projected returns. If the previous program delivered 60% of its targeted savings, size the current investment accordingly. The self-funding model that 44% of companies are relying on only works if the prior returns are real. Approving the next program on the assumption that they are—without verifying it—is the most avoidable financial risk in most companies' AI portfolios right now.

Alongside the financial audit, CEOs must answer one question their IT function cannot answer for them: Who is personally accountable when an AI agent makes a consequential wrong decision in production? Governance is currently split almost evenly between IT, business functions, and central teams, with no clear owner in most organizations. When an agent makes a consequential error in a production system, accountability cannot be improvised in the moment. It must be established in advance. This costs nothing and takes an afternoon. The risk isn’t technological; it’s organizational.

3. Use AI to solve the data problem; don’t wait for the data problem to be solved first. Imperfect data infrastructure is the most cited reason to defer AI investment, and the least valid one. The more productive posture is to sequence AI investments to start where the data is already bounded and accessible, and to use AI itself to improve how data flows through the organization.

The fastest path to value is often automating one repeatable, high-value workflow where humans are currently pulling data manually, consolidating spreadsheets, and producing reports, and replacing that entire sequence with AI. Amazon's Finance Technology team did exactly this with World Wide Watch, a generative AI solution that tracks valued-added tax (VAT) regulatory updates across global markets. What previously took tax teams 26 minutes per regulatory update now takes 2 minutes—a 92% reduction—with 80% of the AI generated summaries accepted without modification by human experts. That is not a moonshot. It is a bounded, specific workflow where the data was already there, and AI replaced the manual assembly.

Reserve the large-scale data modernization conversation for the use cases that genuinely require it, and let early wins like this pay for it.

4. Redesign the operating model, not just the process. Deploying AI agents without changing how people work around them pretty much guarantees an organization will underdeliver on the business case. In an agent-led operating model, employees are no longer moving work along a process; they are orchestrating, supervising, and making the high judgment calls that agents cannot. That is a fundamentally different role, and it requires deliberate investment in role redesign, new ways of working, and change management that most AI programs currently treat as an afterthought.

The organizations capturing transformational savings have leaders who recognize that the human operating model is as important to redesign as the process itself. Typically, they have funded and sponsored that change explicitly at the CEO level, before deployment rather than after. Old ways of operating will not work alongside agents. Left unchanged, it becomes the bottleneck that prevents technology from delivering the economics that the business case promised.

5. Measure outcomes at the enterprise level, not the program level. Programs will always optimize for what they were designed to measure—typically cost and hours saved. What matters for the enterprise is whether AI investment is producing better decisions, faster responses, and stronger customer outcomes. If those metrics aren't on the CEO's dashboard, programs will keep delivering the wrong things efficiently, and the value gap will persist regardless of how much the budget grows.

The turning point for most companies is not finding the best AI technology. It’s the moment when leaders decide—before the next budget cycle, before the next vendor pitch, before the next program launch—that they have a personal responsibility to create the organizational conditions for AI success. The window to make that decision ahead of the competition is still open, but it’s narrowing faster than many executive teams realize.