Most enterprise AI does not pay back because companies buy a model and leave the work exactly as it was. McKinsey tested 25 different organizational attributes against real EBIT impact and found that redesigning workflows had the single biggest effect of all of them. Deloitte surveyed roughly 3,700 professionals and found 48 percent had introduced AI without changing the workflows or roles it sits inside. The tool is almost never the reason the return does not show up. The work around the tool is.
The returns are flat and the spending is not
McKinsey's 2026 global survey found that only 37 percent of organizations report any positive EBIT contribution from AI at all. That number is essentially unchanged from the year before. Spending went up. The share of companies seeing money come back stayed still.
Deloitte found something that explains part of it. Only 4 percent of organizations currently report AI value at the board level. Most companies can tell you what AI costs them. Far fewer can tell you what it returned.
MIT's GenAI Divide study put the same finding in harsher terms: 95 percent of enterprise pilots produced no measurable profit and loss impact. The study covered 52 executive interviews, 153 leader surveys, and 300 public deployments.
Three separate research groups, three different methods, one conclusion. The money is going out. The return is not coming back for most people.
What the companies that do get a return actually did
Here is the part that matters. The research does not just say most companies fail. It says what the small group that succeeded did differently, and the answer is consistent across all three studies.
McKinsey identifies AI high performers as about 6 percent of respondents, the ones attributing 5 percent or more of EBIT impact to AI. That group is nearly three times as likely as everyone else to say they have fundamentally redesigned individual workflows. Not added AI to workflows. Redesigned them.
Metrigy studied 759 companies for its 2026-27 best practices research and found the same split. In the less successful group, 56.7 percent focused on basic automation. In the successful group, 57.6 percent prioritized workflow redesign instead. The companies that redesigned reported a 25.9 percent bigger drop in future costs and a 13.0 percent better increase in revenue than the ones that automated what already existed.
Deloitte's version of this number is that 12 percent report redesign at scale with a new operating model behind it. Roughly one company in eight.
The gap between the 6 to 12 percent who get real returns and everyone else is not a gap in model quality. Every company in these surveys has access to the same models. It is a gap in whether anyone was willing to change how the work moves.
Automating a bad process just makes it faster
There is a difference between automation and redesign that sounds academic until you watch it play out.
Automation takes the process you have and makes each step quicker. The approval still goes to the same three people. The handoff still happens at the same point. The report still gets written, sent, read, and misunderstood. You just do all of it faster.
Redesign asks whether those steps should exist. It looks at a workflow and decides, step by step, whether a person should do it, a machine should do it, or it should not happen at all.
Metrigy found that 53.7 percent of organizations use AI to automate existing workflows rather than change how the work gets done. And 69.5 percent believe AI should support existing work structures. That belief is the whole problem in one sentence. If you start from the position that the structure stays, you have capped your return before you have bought anything.
Why this happens even when people know better
Redesigning work is politically expensive. Automation is not.
Adding an AI tool to an existing process requires a budget line and a vendor. Nobody's job description changes. Nobody's reporting line moves. No director has to give up a step they currently own. You can do it in a quarter and announce it.
Redesigning a workflow means telling people their job is now different. It means a handoff someone has controlled for six years goes away. It means the org chart and the work stop matching. That is a much harder thing to get approved, and it takes longer than the average patience for an AI initiative.
So companies do the easy version, measure it, find nothing, and conclude AI does not work for them. The MIT framing of this was that companies avoid friction. The friction was the part that produced the return.
The second problem nobody budgets for
Even companies that redesign one workflow often stop there, and the reason is that the tools do not talk to each other.
A redesigned workflow usually crosses several people and several systems. If each person has their own AI tool and those tools cannot pass work between them, you have not redesigned a workflow. You have built a faster version of each individual step, with the same gaps between them. The lag lives in the handoffs, and the handoffs are exactly what a single-user tool cannot touch.
This is why per-seat AI often produces a real feeling of productivity and no visible number. Everyone is faster at their own piece. The thing the business actually measures, how long it takes to go from decision to done, has not moved, because the waiting was never inside the individual tasks.
Deloitte's autonomy data supports this. 69 percent of organizations sit at the most conservative end, either allowing no AI autonomy at all or limiting it to low-risk reversible actions. Only 12 percent report AI running end to end with human auditing. If AI cannot carry work across a boundary without a person picking it up and putting it down again, the boundary is still the bottleneck.
What to do instead
Pick one workflow. Not a department, not a function, one workflow that a business owner can name and that has a start and an end.
Deloitte found that 37 percent of organizations making changes fully own one workflow, test it, and then scale up. That is the pattern that works, and it is also the pattern that survives a board meeting, because you can point at a single before and after.
Then do four things in order.
Map the current workflow as it actually runs, not as the process document describes it. Include the waiting. The waiting is usually most of it.
Decide for every step whether a person, a machine, or both should own it, and be willing to delete steps entirely. A step that exists only to inform someone is a candidate for deletion, not automation.
Give whatever AI you deploy the same access and context the employee doing that job has. A tool that cannot see what the person can see will always need the person to fetch things, and fetching is a handoff.
Measure the whole workflow end to end, not the individual steps. Step-level speed is where false wins come from.
The honest summary
The research does not say AI does not work. It says AI does not work when you refuse to change anything else. The companies seeing 5 percent or more EBIT impact are not using better models than the companies seeing zero. They changed the shape of the work and accepted the cost of doing that.
Everyone else bought a fast tool, bolted it onto a slow process, and measured the tool.
Frequently asked questions
Why do most enterprise AI pilots fail to show ROI? Because the workflow around the AI stays the same. McKinsey tested 25 organizational attributes and found workflow redesign had the largest effect on EBIT impact. Companies that add AI to an unchanged process cap their return at how much faster individual steps can run, which is usually not enough to show up in financial reporting.
What is the difference between automating a workflow and redesigning one? Automation makes each existing step faster while keeping the sequence, handoffs, and owners in place. Redesign questions whether each step should exist and reassigns it to a person, a machine, or neither. Metrigy found companies that redesigned reported a 25.9 percent bigger drop in future costs than companies that automated.
What percentage of companies actually get financial return from AI? McKinsey's 2026 survey found 37 percent report any positive EBIT contribution, and only about 6 percent qualify as high performers attributing 5 percent or more of EBIT to AI. Deloitte found only 4 percent report AI value at the board level.
Why does per-seat AI produce productivity that does not show up in results? Because the delay in most business processes lives in the handoffs between people, not inside individual tasks. Giving every employee a separate tool speeds up each person's own work while leaving the gaps between them untouched, so end-to-end time does not change.
How should a company start redesigning work around AI? Start with a single workflow that has a named owner and a clear start and end. Deloitte found 37 percent of organizations making changes use this pattern, owning one workflow fully before scaling. Map how it actually runs including the waiting, decide step by step who or what should own each part, and measure end to end rather than step by step.
