A new global study from SAP and Oxford Economics just dropped a number that should be on every CFO's radar: agentic AI is expected to generate an average ROI of $17.6 million per enterprise within two years. That's a 4x jump from the $4.3 million estimate of just one year ago. The market has never moved this fast on a technology-specific ROI projection.
But here's the catch — and it's a big one.
Only 3% of businesses say they are fully prepared for agentic AI deployment. The rest? Either partially prepared, or not prepared at all. The survey of 2,600 executives across 13 countries paints a vivid picture: a massive ROI prize is sitting on the table, and most enterprises lack the governance, data quality, and operational controls to actually claim it.
This isn't a technology readiness problem. It's a leadership, governance, and organizational readiness problem. And it's about to get much more expensive to ignore.
The ROI Numbers Are Real — And Accelerating
Let's start with what the data actually says, because the headline numbers are unusually clear for a study of this kind.
Today, enterprises globally expect a 21% ROI on AI investments — up from 16% last year. On a $28 million average spend, that's $5.9 million in returns. Two years from now, that expected ROI rises to 38%, or roughly $15.9 million on the same spend base. For agentic AI specifically, the ROI expectation jumps even higher: from $4.3 million last year to $17.6 million within two years.
These aren't speculative numbers. They're derived from 2,600 business leaders actively running AI in production. The pattern is consistent across markets, too: Brazil, UK, Australia, and Germany all saw increased investment this year, and the relationship between investment level and expected return tracked closely across all 13 countries surveyed.
What's driving the acceleration? Agentic AI workflows. Agents that can gather context, reason over data, and take autonomous action across enterprise systems are proving to be fundamentally different from the static LLM features that most enterprises deployed in 2024-2025. The productivity leverage is real — and the numbers are starting to show it.
For CFOs building AI business cases, this data matters. The ROI argument for agentic AI is no longer theoretical. It's benchmark data from peers across your industry.
The Governance Gap: Why Most Will Miss the Prize
Now for the uncomfortable part.
The same study that projects $17.6 million in agentic ROI also documents a governance landscape that would make any CRO or General Counsel lose sleep. The numbers tell a story of rapid deployment outpacing organizational readiness at nearly every level.
38% of enterprises have no human-in-the-loop process for overseeing agentic AI workflows. These are companies where AI agents are making or initiating decisions with no defined checkpoint for human review.
37% have no permission and access controls for their AI agents. Think about what that means in practice: agents with access to enterprise systems, data, and APIs, operating without defined boundaries on what they can read, write, or execute.
Only 44% have a registry of the agents currently running in their business. The majority of enterprises don't have a full picture of what AI agents are active, what they're doing, or who authorized them.
And here's the statistic that should drive board-level urgency: 69% of businesses either agree or can't confirm that they are deploying agents faster than they can govern them. Nearly seven in ten enterprises are running faster than their oversight capabilities can keep pace with.
Sean Kask, SAP's Chief AI Strategy Officer, put it bluntly in the report's foreword: "AI that lacks context — whether that's processes, data, or governance — at best creates activity without outcomes and at worst creates risk."
For CIOs and CTOs, this is a systems architecture problem masquerading as a policy problem. The controls that work for traditional software don't translate cleanly to autonomous agents. Agents don't have a single code path to audit. They reason over data, select tools, and chain actions in ways that are non-deterministic. Standard change management processes weren't built for this.
Shadow AI: The Invisible Fleet
Alongside the official agent deployments, there's a second, less visible problem growing in parallel.
69% of enterprises say shadow AI use happens at least occasionally. That's employees using personal AI accounts, unsanctioned tools, or unauthorized agent automations to get work done — outside the visibility of IT, security, or compliance.
In conversations with security and operations leaders, this is consistently one of the most underestimated risks in enterprise AI right now. It's not about employees being malicious. It's about employees finding faster ways to do their jobs, and the gap between what corporate-sanctioned tools can do versus what personal AI subscriptions can do is still wide enough that the temptation is real.
The governance implications are significant: data exfiltration risk, compliance violations, AI-generated outputs in business processes that have no audit trail, and agents operating on enterprise data that were never reviewed by security teams.
Shadow AI is what happens when you move fast on deployment and slow on enablement. If employees can't get to the right AI tools through official channels, they'll find unofficial ones.
Data Quality: The Foundation That Isn't There Yet
Agentic AI is only as good as the data it operates on. That reality is crashing into enterprises at scale.
73% of companies report challenges with incomplete data — and that number actually increased from last year. More enterprises are deploying AI, more AI is depending on clean data, and more organizations are discovering that their data foundations weren't built for this.
The practical consequence: 79% of businesses are experiencing rework, delays, or backlogs due to low-quality AI outputs. Nearly four out of five. If you've deployed AI and you're wondering why your productivity gains aren't showing up in P&L, this is likely a significant factor. Garbage-in, garbage-out is an old concept, but it hits differently when the garbage is coming out of an agent that touched 47 enterprise systems before producing a flawed recommendation.
In conversations with data engineering and architecture teams over the past several months, a pattern emerges: organizations that moved fast on AI feature deployment (the "just plug in the API" phase) are now circling back to do the data work they skipped. Master data quality, governance tagging, access control at the data layer, and lineage documentation — all of the boring, foundational work that AI pilots don't require but production-grade agents absolutely do.
The enterprises that got the data layer right before scaling agents are now seeing significantly better outcomes. The ones that didn't are spending their AI budgets on rework rather than value creation.
The Measurement Problem Nobody Talks About
There's a secondary finding in the SAP report that deserves more attention than it typically gets: only 33% of companies have KPIs at the executive level directly linked to AI implementation.
Think about the incentive structure this creates. Companies are spending $28 million a year on AI. They're projecting 21% ROI. But two-thirds of executive teams have no formal metrics tied to whether those investments are actually delivering. That's not a technology problem. That's a management accountability problem.
The companies that are capturing real AI value have done three things systematically: they defined what "success" looks like in measurable terms before deployment, they built infrastructure to connect AI task outputs to business process outcomes, and they have an executive owner accountable for the result — not just the deployment.
In a Fortune 500 company I spoke with recently, the AI program had been running for 18 months before anyone asked whether the efficiency gains from their AI-assisted process were showing up in headcount planning or cost forecasts. They weren't — not because the AI wasn't working, but because the AI savings had been absorbed into "we got more done" rather than "we spent less" or "we grew more." The value was real; the capture was not.
This is the difference between AI activity and AI ROI. The SAP data shows 69% of businesses are satisfied with their current AI ROI — even though more than two-thirds say they aren't close to realizing AI's full potential. You can be satisfied with a partial result if you don't know what the full result should look like.
What Leaders Need to Do Now
The data is clear on what separates enterprises that will capture the $17.6M agentic ROI from those that will miss it. It comes down to four fundamentals:
1. Build an agent registry before you build more agents. You cannot govern what you cannot see. If your organization doesn't have a centralized view of every AI agent running in your environment — who authorized it, what data it accesses, what actions it can take — you have a governance blind spot that will eventually become a compliance or security incident.
2. Define human-in-the-loop requirements by risk tier. Not every agent action needs a human checkpoint, but some absolutely do. Define the tiers: what can run autonomously, what requires notification, and what requires approval before action. This is the operational equivalent of a delegation of authority matrix — and enterprises that have it are moving faster and more safely than those that don't.
3. Fix the data layer, not just the AI layer. If 79% of your AI-assisted work is generating rework due to data quality issues, adding more AI won't fix it. The data foundation work is unsexy and unglamorous, but it's the multiplier that determines whether your $28 million AI spend generates $5.9 million in returns or $17.6 million.
4. Create executive AI accountability with real metrics. Link a portion of executive objectives — for CIOs, CTOs, and business unit leaders — to measurable AI outcomes. Not deployment metrics. Business outcomes: cost reduction, cycle time improvement, revenue impact, or error rate reduction. What gets measured gets managed.
The Bottom Line
The SAP Value of AI Report 2026 is one of the most data-rich enterprise AI benchmarking studies published this year. The headline is simple: the ROI from agentic AI is going to be enormous, and most enterprises are not positioned to capture it.
The gap isn't in AI capability or investment intent. Enterprise leaders across every major market are willing to spend, and the technology is delivering measurable results for those who deploy it well. The gap is in governance, data readiness, measurement discipline, and organizational accountability.
The enterprises that will reach $17.6M in agentic ROI within two years are already doing the foundational work right now. They're building agent registries. They're fixing their data quality before scaling their agent fleet. They're defining human-in-the-loop thresholds. They're tying AI outcomes to executive accountability.
The other 97%? They'll spend two years admiring the ROI projections while the gap between potential and captured value keeps widening.
The technology isn't the obstacle. Leadership and operational readiness are.
Source: SAP and Oxford Economics, Value of AI Report 2026, surveying 2,600 business leaders across 13 countries — Australia, Brazil, Canada, China, France, Germany, Italy, India, Japan, Singapore, Thailand, United Kingdom, and United States.
Continue Reading:
