The numbers are staggering. Enterprises worldwide will spend $2.59 trillion on AI in 2026 — a 47% increase year-over-year. By 2027, that figure climbs to $3.5 trillion. And yet, a new Bain & Company survey of 951 companies found that nearly 40% of enterprises that actually measured their AI cost savings came in below 10% — despite targeting returns of 11% to 20%.
That gap should make every CFO uncomfortable. And it should make every CIO rethink how they're framing AI investments to the board.
I've had conversations with peers in enterprise leadership who are seeing exactly this. Budgets keep growing, pilots keep launching, and yet the quarterly reviews tell a quieter, more uncomfortable story. The ROI isn't landing. And nearly everyone is approving more spending for the next wave anyway.
This piece digs into why AI investments are underdelivering — and what the organizations actually getting results are doing differently.
The Ritual That Never Ends
Bain's co-authors — Michael Heric, Purna Doddapaneni, and Antoine Debarre — describe a pattern they've watched repeat across enterprise after enterprise:
"Every year, CEOs sign off on the next wave — robotic process automation, then machine learning, then generative AI, now agents. And every year, the savings fall short."
This isn't a bug in AI. It's a bug in how organizations deploy it. The Bain team puts it bluntly: "The technology worked. The value didn't arrive."
That's an important distinction. AI can absolutely automate workflows, generate content, analyze documents, and summarize calls. The capability is there. What's failing is the translation of that capability into measurable enterprise value.
And yet, 90% of companies in the Bain survey are ramping up their AI budgets again — this time to build and deploy agents "that will operate with even greater autonomy, complexity, and consequence." The pattern is accelerating, not slowing.
Three Root Causes Killing Your AI ROI
The Bain data points to three structural problems that are quietly draining AI budgets.
1. AI Isn't Actually Autonomous Yet
Only 7% of companies are running fully autonomous AI agents in production today. That means the other 93% still need humans in the loop — reviewing outputs, correcting errors, managing exceptions, and keeping workflows from going off-rails.
This matters enormously for ROI calculations. If your business case assumed a 60% reduction in headcount for a workflow, but the actual deployment still requires three full-time staff to supervise the AI, your savings model is broken before it starts. The human oversight cost was never in the spreadsheet.
From a technical standpoint, this is a gap between demo-grade and production-grade AI. Getting AI to 80% accuracy in a controlled environment is relatively straightforward. Getting it to 99.5% reliability at enterprise scale — with edge cases, regulatory constraints, legacy data, and exception handling — is a fundamentally different engineering problem.
2. The Circular Bet
When Bain asked how companies plan to fund generative AI and agentic AI investments, 44% said they're relying on savings from prior automation programs.
That sounds like fiscal discipline. It's actually a structural problem.
"Self-funding the next wave from past returns sounds like discipline. In reality, it is a circular bet with a structural leak. The prior wave underdelivered. The savings pool is smaller than assumed."
CFOs who approved RPA investments expecting 20% cost reduction — but actually got 12% — are now approving gen AI investments assuming those prior automation dollars are available to redeploy. They aren't. The math doesn't close, and by the time the next quarterly review surfaces the gap, another budget cycle has already begun.
3. Data Is Still the Wall
This one isn't new, but it's stubbornly persistent. Data access and integration remains the top barrier to AI progress for 41% of Bain's respondents — despite years of heavy investment in data modernization programs.
Talking to CIOs across industries, the story is consistent: there's no shortage of data. The problem is that the right data isn't accessible to the AI systems in real-time, in the right format, with the right permissions and governance controls. You can have a Snowflake data warehouse, a modern lakehouse architecture, and a dedicated data engineering team — and still find that your AI agent can't pull the specific records it needs because a business unit built its own shadow system three years ago and never integrated it.
For technical leaders, this is where RAG architectures and vector databases often overpromise. They solve document retrieval well. They don't solve the harder problem of enterprise data fragmentation.
What the Winners Are Doing Differently
The Bain team identified a set of practices that distinguish organizations actually capturing AI value. These aren't abstract frameworks — they're operational disciplines.
Don't Pave Cowpaths with AI
The worst use of AI is making an already-broken process 20% more efficient. You've just locked in the brokenness with automation.
The question to ask before any AI program gets approved: "If we were designing this process from scratch today, what would it look like?" That almost always produces a fundamentally different answer than "take the existing process and inject AI into it."
In conversations with business leaders, I see this mistake most often in finance and HR. AP automation that routes invoices through AI is a cowpath pave. Redesigning the entire vendor payment workflow — eliminating approval chains that exist only because manual processes required them — is process redesign with AI as the enabler.
CFOs: Audit Your Actual Returns, Not Your Projected Ones
The Bain recommendation here is pointed: "If the previous program delivered 60% of its targeted savings, size the current investment accordingly."
This is the kind of honest accounting that's politically difficult inside organizations. But it's the difference between a realistic AI investment thesis and a wishful one. If your last three automation initiatives hit 55%, 65%, and 58% of their projected savings, you don't have a 100% execution rate. You have a 60% execution rate. Your gen AI and agent investment should be sized accordingly.
CFOs should push for program-level audits of actual versus projected savings from prior automation before approving the next wave. Not because AI isn't worth investing in — it clearly is — but because accurate baselines produce better investment decisions.
Put a Human in Charge Before an Agent Makes a Consequential Error
Bain found that AI governance in most organizations is "split almost evenly between IT, business functions, and central teams, with no clear owner in most organizations."
When an AI agent makes a consequential error in production — and it will — accountability cannot be improvised in the moment. It has to be established in advance. Who owns the agent? Who owns the data it acted on? Who has authority to roll back the action?
For CIOs and CTOs, this is a governance architecture problem. For business leaders, it's a risk management problem. Either way, the absence of a clear owner is the fastest path to an enterprise AI incident becoming an enterprise AI crisis.
Use AI to Solve the Data Problem First
Here's a practical starting point for the data integration challenge: automate one repeatable, high-value workflow where humans are currently pulling data manually, consolidating spreadsheets, and producing reports.
Replace that entire sequence with AI. Not as a cost-saving exercise, but as a data confidence exercise. You'll learn more about your data quality, access permissions, and integration gaps from one focused deployment than from months of architecture reviews.
This approach also creates a credible internal case study — which is invaluable when you're trying to build organizational confidence in AI and secure budget for larger initiatives.
Redesign Employee Roles, Not Just Headcount
This is the piece that enterprise leaders most often get wrong — and it's also where AI investments most often stall.
The assumption embedded in most AI ROI models is that AI reduces headcount. In the short term, that's rarely true. In the medium term, it requires deliberate role redesign, not just reduction.
In an agent-led operating model, employees are no longer moving work along a process. They're orchestrating, supervising, and making the high-judgment calls that agents can't. A claims processor who used to handle 40 cases per day becomes a quality supervisor managing an agent handling 400 cases per day — but that transition requires deliberate investment in training, new workflows, and change management.
Organizations that skip this step find that their agents get deployed, but adoption stalls. The humans in the loop aren't equipped to supervise them effectively, so errors slip through, exceptions get mishandled, and leadership loses confidence in the technology.
Measure Outcomes at the Enterprise Level, Not the Program Level
This is the most important shift in how enterprise AI gets evaluated.
Individual AI programs can show positive ROI in their own metrics while the enterprise-level impact is flat or negative. A customer service AI that deflects 30% of tickets is a win by program metrics. If customer satisfaction scores drop, customer churn increases, and escalation costs rise, the enterprise outcome is negative — even though the program "succeeded."
What matters for the enterprise is whether AI investment is producing better decisions, faster responses, and stronger customer outcomes. Those are enterprise-level measurements that cross program boundaries and require executive-level ownership.
The Agent Wave Is Coming — Ready or Not
The 90% of companies ramping up agent budgets despite current ROI gaps aren't being irrational. Agentic AI represents a genuinely different capability step from traditional automation or even generative AI. Agents can plan, adapt, chain tasks, and operate continuously across enterprise systems in ways that point solutions cannot.
But the same structural problems that undermined ROI for RPA, ML, and gen AI will undermine agent ROI if they aren't addressed first. Agents operating with "greater autonomy, complexity, and consequence" — to use Bain's phrase — need better data foundations, clearer governance, and more realistic investment theses than the prior waves got.
The question isn't whether to invest in agents. It's whether you've fixed the plumbing from the last three waves before you add a fourth.
What This Means for Technical and Business Leaders
For CIOs and CTOs: The data problem isn't going away by itself. Prioritize one focused data integration project that unblocks a specific AI agent capability. Use it to build the internal proof of concept and the technical credibility needed to address the broader data architecture gaps.
For CFOs: Audit prior automation ROI before approving the next wave. Not to kill AI investment — the strategic case is real — but to right-size it based on your organization's actual execution track record.
For CEOs: The 7% of organizations with fully autonomous agents in production didn't get there by approving bigger budgets. They got there by solving governance, data, and role redesign — and then funding the agents on top of a working foundation.
The $2.59 trillion being spent on AI globally in 2026 will produce wildly unequal returns. The organizations capturing outsized value aren't the ones spending the most. They're the ones who've fixed the structural problems that prevent AI from delivering at scale.
That's the fix.
Sources: Bain & Company survey of 951 companies, Gartner 2026 AI Spending Forecast, Forbes analysis
Continue Reading:
