SAP and Oxford Economics published the Value of AI Report 2026 on July 15, 2026. They asked 2,600 business leaders across 13 countries what AI is actually returning and what they expect next. The headline is that AI has moved from experiment to execution, and the money is starting to show it. The quieter finding is the one worth reading twice. The companies reporting the best returns are, by their own account, mostly unprepared for the next phase.
If you have felt behind, here is some permission to stop. The large enterprises in this survey are spending an average of 28 million dollars on AI, and even they say only 3 percent are fully ready for what is coming. Nobody has this solved. This is the time to build the right foundation while the field is still level.
What the numbers actually say
The returns are real and rising, at least on paper. Companies in the SAP survey expect a 21 percent return on their AI investment this year, up from 16 percent the year before. They project 38 percent in two years. On an average spend of 28 million dollars, that 21 percent works out to about 6.3 million dollars returned this year, and the two-year projection sits near 15.9 million.
One caveat is important here. These are reported and expected returns, not audited ones. When a survey asks executives what AI is giving them back, you get confidence and forecasts, not a verified ledger. Read the direction of travel, not the decimal.
The bigger story is where the confidence is pointed: agents. The survey found companies expect an average return of 17.6 million dollars from agentic AI over the next two years, up from a 4.3 million dollar estimate a year earlier. That is a large bet on software that can take actions on its own, the kind that does the work rather than only answering questions about it. AI supports about 30 percent of business tasks today, and leaders expect that to reach 48 percent in two years.
So the spending is up, the expectations are way up, and the returns are trending in the right direction. Then you reach the readiness numbers.
The readiness picture is thinner than the spending
Only 3 percent of companies in the survey say they are fully prepared for agentic AI. At the same time, 83 percent see moderate-to-very-high transformation potential in it. That distance, between what leaders believe is possible and what their organization can actually support, is the whole story.
of companies say they are fully prepared for agentic AI, even as 83 percent see moderate-to-very-high transformation potential in it.
Source: SAP and Oxford Economics, Value of AI Report, 2026Look at how the work is organized. The most common approach is still piecemeal, a tool here, a pilot there, at 41 percent of companies. Only 17 percent report a strategic, company-wide investment approach. The encouraging note is that the strategic share nearly doubled in a year. More companies are moving from scattered experiments toward one plan.
The supporting structures are catching up slowly:
- 46 percent have a dedicated AI leader.
- 52 percent have clear frameworks for how they develop AI.
- 41 percent run AI training programs for their people.
Read those three numbers together. Roughly half of companies have named someone to own AI, half have written down how they build it, and fewer than half are teaching their teams to use it. That is the difference between owning a tool and building a capability.
The everyday version of the shortfall
You do not need a 28 million dollar budget to recognize what the survey describes. Two numbers translate straight to any team.
First, data. 73 percent of companies cite incomplete data as a challenge, and 79 percent report rework or delays caused by poor AI output. That second number is the one to sit with. Four out of five say the AI gave them something they had to fix or redo. The tool is fast. Feeding it the wrong context makes it fast at producing work you cannot use.
Second, shadow AI. 69 percent of companies report that their people use unsanctioned AI tools at least occasionally. Your team is already using AI you did not choose, on data you have not thought about, without a process around it. That is a signal. People reach for tools that help them. The question is whether you know which tools, on what, and with what checks.
Sean Kask, SAP's Chief AI Strategy Officer, put the mechanism plainly.
"AI has moved from experiment to execution, and that's beginning to show real returns. But there's still a long way to go. Because AI that lacks context, whether that's processes, data, or governance, at best creates activity without outcomes and at worst creates risk."
Activity without outcomes. That is the 79 percent rework number in five words.
Agents raise the stakes on the boring stuff
An AI that drafts an email is a productivity tool. An AI that can send the email, update the record, and move money is something else. The survey's agent-specific findings are where a P&L owner should slow down.
Among the companies deploying or planning agentic AI, 38 percent have no human in the loop for agentic processes. 37 percent lack permission and access controls for their agents. Only 44 percent have implemented an agent registry, a basic list of what agents exist and what they are allowed to do.
Put those next to governance readiness overall, which sits around 12 percent, and the shape is clear. The controls that keep an autonomous system from acting on the wrong thing are the least-built part of the stack, right as companies plan to hand agents nearly half their tasks.
This is an argument for sequence. Kask again:
"Businesses are quickly discovering that AI governance plays a foundational role in unlocking the value from AI."
Governance is not the tax you pay after the value. In this data, it is the thing that lets the value show up at all.
What this means for building an AI-capable team
Here is the part a leader can act on without a nine-figure budget.
The survey found 78 percent of companies are unsure their upskilling is keeping up with AI, or agree it is not. That gap is the real bottleneck, more than model choice or vendor selection. A tool you buy is a line item. A team that can direct, check, and improve AI is a durable advantage, and it is the part competitors cannot purchase in an afternoon.
This is where a shared way to measure capability helps. The 7 Levels of AI Proficiency describes what growing skill actually looks like inside a team, from a person taking their first useful steps with a tool, up through someone who can design a whole process around AI and keep a human check on it where it counts. It gives you a common language for a simple question: what can each person on my team actually do with these tools today, and what is the next level for them?
You can start this week, and none of it requires a big spend:
- Name what your team is already using. The 69 percent shadow AI number is your team. Ask what tools people reach for and on what work. You cannot govern or improve what you cannot see.
- Pick one process, not one tool. The 41 percent stuck in piecemeal mode are buying tools. The move is to choose a single workflow that counts, map how it actually runs, then decide where AI fits inside it.
- Put a human on the output. The 79 percent rework rate is what happens when nobody checks. A person reviewing AI work before it ships is the cheapest governance you will ever install.
- Teach before you scale. With 78 percent unsure their upskilling is keeping up with AI, a short, hands-on session that meets people where they are does more than another license.
SAP calls the destination the Autonomous Enterprise, AI connected to a company's real data, processes, context, and governance. You can strip the branding off and keep the instruction. Connect AI to the actual work, give it context, put a check on it, and teach the people around it. That order is the finding.
Where to start
Pick one workflow your team runs every week. Not the flashiest one. The one that eats the most time or produces the most rework. Write down how it actually works today, step by step, before any AI touches it. That single page tells you where a tool helps, where a human has to stay in the loop, and what your people need to learn next. The 3 percent who feel ready did not get there by buying more. They got there by connecting AI to work they already understood.
Related reading: Level 4: The Commander (Context Engineer).
Sources
Frequently Asked Questions
Is a 21 percent return on AI a solid benchmark for my business?
Treat it as directional, not a target. The figure comes from an SAP and Oxford Economics survey of large-enterprise leaders reporting and projecting their own returns, not from independently audited results. It tells you the trend is up and confidence is rising. It does not promise your number.
We are doing AI piecemeal. Is that a problem?
You are in the majority. 41 percent of companies in the survey are in the same spot, and only 17 percent report a company-wide strategy. Piecemeal is a normal starting point. The useful next step is picking one important process to organize around, rather than adding more disconnected tools.
Should I hold off on AI agents until governance is ready?
The survey suggests you do not have to choose, but you do have to sequence. With 38 percent of agent adopters running no human in the loop and only about 12 percent reporting governance readiness, the safer path is small: one agent, on a defined task, with a human check and a record of what it is allowed to do.
Is a 21 percent return on AI a solid benchmark for my business? Treat it as directional, not a target. The figure comes from an SAP and Oxford Economics survey of large-enterprise leaders reporting and projecting their own returns, not from independently audited results. It tells you the trend is up and confidence is rising. It does not promise your number.
We are doing AI piecemeal. Is that a problem? You are in the majority. 41 percent of companies in the survey are in the same spot, and only 17 percent report a company-wide strategy. Piecemeal is a normal starting point. The useful next step is picking one important process to organize around, rather than adding more disconnected tools.
Should I hold off on AI agents until governance is ready? The survey suggests you do not have to choose, but you do have to sequence. With 38 percent of agent adopters running no human in the loop and only about 12 percent reporting governance readiness, the safer path is small: one agent, on a defined task, with a human check and a record of what it is allowed to do.
What is the cheapest first move? Find out what AI your team already uses. Nearly 7 in 10 companies have unsanctioned tools in play. That conversation costs nothing and tells you where your real starting line is.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.