Enterprise AI applies machine learning and generative models to real workflows such as forecasting, risk scoring, customer operations and decision support so teams decide faster and more accurately. While 88 percent of organizations use AI, only about 5 to 6 percent capture significant financial value, because returns depend on workflow integration and governance, not the technology alone.
Almost every enterprise now runs on AI in some form, and almost none of them can tell you what it earned. That gap is the defining story of 2026. According to McKinsey State of AI 2025, 88 percent of organizations use AI in at least one function and 79 percent use generative AI, yet only about 5.5 percent report a material EBIT impact of 5 percent or more. Adoption is nearly universal. Value is scarce.
At PEKVOR we spend most of our engagements on the second number, not the first. The technology has largely been solved for common cases; the hard part is turning a working model into a workflow that changes a business result. This article explains the paradox, what actually stalls pilots, and the operating model the winners use.
The enterprise AI paradox: near-universal adoption, scarce returns
The headline figures look contradictory. McKinsey reports that 88 percent of organizations use AI and 79 percent use generative AI, but only around 5.5 percent see a material EBIT effect and only 6 to 7 percent have scaled enterprise-wide. The clearest external confirmation comes from MIT NANDA's 2025 study, The GenAI Divide, which found that 95 percent of enterprise generative-AI pilots delivered no measurable P and L return.
Read those together and the conclusion is uncomfortable but useful: the problem is almost never the model. Organizations are not failing because the AI cannot do the task. They are failing because a demo that works in isolation is not the same as a workflow that changes an outcome at scale, under governance, with people who trust it.
What decision support means beyond chatbots
Much of the disappointment traces back to a narrow mental model of AI as a chat window. A chatbot answers a question. Decision support changes a decision. Those are different products with different economics.
In practice, decision support means embedding models where choices already get made:
- Demand and revenue forecasting that feeds planning, not a slide.
- Risk and fraud scoring that routes a case before a human touches it.
- Customer operations that draft, triage and prioritize so agents handle exceptions.
- Next-best-action recommendations a person reviews and approves.
The common thread is that the AI output lands inside an existing decision, with an owner, a metric and a feedback loop. That is where McKinsey's finding bites: it reports that workflow redesign has the largest effect on value of any factor studied. The model is a component; the redesigned workflow is the product.
Why 95% of pilots stall — and it is not the technology
MIT NANDA's 95 percent figure is the most cited statistic in enterprise AI for a reason, but the more actionable finding underneath it is about how organizations build. The study found that buying capability from specialized partners succeeded far more often than internal-only builds. Teams that tried to construct everything in-house, from scratch, tended to underestimate the integration, evaluation and change-management work that surrounds the model.
The recurring failure patterns we see match that data:
- The pilot proves the model works but never touches the system of record.
- No one owns the workflow, so a promising demo has no home in production.
- There is no evaluation harness, so quality is a matter of opinion.
- Data access is solved by hand for the pilot and cannot survive at scale.
- Governance arrives late, and the project stalls in legal or risk review.
None of these are model problems. They are engineering and operating problems, which is precisely why capability alone does not translate into returns.
The economics flipped: collapsing inference cost
The reason the conversation has shifted from can it work to can we deploy it is cost. Stanford HAI's 2025 AI Index reports that the inference cost for GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. That is not an incremental improvement; it is a change in what is economically sensible to automate.
When inference was expensive, the model was the constraint and you rationed it. When inference is cheap, the constraint moves to everything around the model: data pipelines, integration, evaluation and governance. The market has voted with its wallet. Stanford HAI also reports that global corporate AI investment reached $252.3 billion in 2024, up 26 percent. The money is flowing; the discipline to convert it into EBIT is what remains scarce.
High-value use cases that move EBIT
The winners concentrate on a short list of workflows where AI changes a number an executive already tracks. McKinsey's 2023 analysis estimated that generative AI could add $2.6 to $4.4 trillion annually across use cases, but that potential is unevenly distributed. In our experience the EBIT-moving categories cluster in:
- Customer operations, where drafting and triage cut handle time and raise resolution rates.
- Software and engineering, where code generation and review compress cycle time.
- Marketing and sales, where content and next-best-action lift conversion.
- Risk, finance and supply chain, where forecasting and scoring reduce loss and working capital.
The pattern is that value shows up where volume is high, decisions are repetitive, and a small quality gain multiplies across thousands of transactions.
The data and governance foundation
Every durable deployment we build rests on two unglamorous foundations. The first is data: reliable access, lineage and freshness for the specific fields the workflow needs. The second is governance: clear policy on what the system may decide alone, what a human must approve, how outputs are logged, and how the whole thing is audited.
Governance is often framed as a brake. Treated correctly it is an accelerant, because it is what lets a promising pilot survive risk review and reach production. Projects that leave governance until the end are the ones that stall in the last mile.
From pilot to production: an operating model
The shift from pilot to production is organizational, not technical. The winners run AI like a product, not a science experiment:
- Pick one workflow with a named owner and a measurable outcome.
- Redesign the workflow around the model, keeping a human in the loop for high-stakes decisions.
- Build an evaluation harness before scaling, so quality is measured, not asserted.
- Instrument the workflow end to end so you can attribute results.
- Expand only after the first workflow proves value.
This is the discipline behind McKinsey's 6 to 7 percent who have scaled enterprise-wide. They did not scale a model; they scaled an operating model.
Measuring ROI honestly
The single healthiest habit an enterprise can adopt is honest measurement. That means a baseline before deployment, a clear metric tied to EBIT, and the willingness to kill workflows that do not move it. Given that only about 5.5 percent of organizations report a material EBIT impact, the goal is not to join the 88 percent who use AI but the small minority who can prove it paid.
It is worth keeping the labor picture in view too. The World Economic Forum's Future of Jobs 2025 projects AI will create 170 million new jobs while displacing 92 million by 2030. The net is growth, but the transition is real, and honest ROI accounting includes the change-management cost of getting people to trust and use the system.
How PEKVOR approaches enterprise AI
We start from the workflow and the number it is meant to move, not from the model. Our engagements begin by choosing one decision-heavy process with a named owner, then redesigning it around AI with a human in the loop where stakes are high. We build the data access and evaluation harness first, put governance and audit trails in from day one, and instrument outcomes so value is provable rather than assumed. Because the economics have flipped, our leverage is not in access to models but in the engineering and operating discipline that turns a pilot into EBIT. If your organization is in the 88 percent that uses AI but not yet in the small minority that profits from it, that gap is exactly the work we do.
Frequently asked questions
What is enterprise AI decision support?
It is the use of machine learning and generative models inside real business workflows to help people make faster, better-calibrated decisions, such as forecasting demand, scoring risk, prioritizing accounts or drafting a recommended action for a human to approve, rather than a standalone chatbot.
Why do AI projects fail?
Most stall because they are treated as technology experiments rather than workflow changes. MIT NANDA found 95 percent of enterprise generative-AI pilots delivered no measurable P and L return, largely because they were bolted onto processes without redesign, data plumbing, governance or clear ownership.
How much cheaper has AI become?
Dramatically. Stanford HAI reports the inference cost for GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024, which moves the economics from a capability problem to an integration-and-governance problem.
What ROI should we expect from enterprise AI?
Be honest and specific. McKinsey found only about 5.5 percent of organizations report a material EBIT impact of at least 5 percent, so expect targeted gains in the workflows you actually redesign, not blanket company-wide returns from a single deployment.
Where should we start with enterprise AI?
Start with one high-volume, decision-heavy workflow where errors are costly and data already exists. Redesign the workflow around the model, put a human in the loop for high-stakes calls, and instrument outcomes so you can prove or disprove value quickly.
Have a project where this matters?
This is the discipline we bring to every engagement. Tell us what you are building and we will show you how we would approach it.