Every company can buy the same AI models at the same list prices. Yet AI use and AI value are concentrated in a small group of companies, and that group is pulling away. The data from the US Census, BCG, MIT, and OpenAI points to one cause. The leaders changed how the work is done. A newer model does not fix a workflow that was not redesigned.
AI Use Is Wide but Shallow
The US Census Business Trends and Outlook Survey found that 19.8% of US firms used AI in a business function in the two weeks ending May 3, 2026. Use rises with firm size.
Use is also shallow. A Census working paper by Bonney and coauthors found that 57% of firms that use AI use it in three or fewer business functions. Weighted by employment, 32% of firms used AI in some function.
The Leaders Pull Away
OpenAI's Enterprise Signals report ranks enterprise customers each month by output tokens per active user. In June 2026, the top 10% of firms used 8.3x as many tokens per user as typical firms, those near the median. In January, the gap was 2.6x. Token volume measures how deeply people use AI, not the value it creates. Still, in six months the most active users moved far ahead of the rest.
Value shows the same split. BCG's 2026 survey of 1,330 companies found 7.5% "future-built" leaders and 41% scaling. In 2024, 74% of companies reported no tangible value.
The Difference Is the Workflow
Leaders spend most of their effort on process. In BCG's 2024 study, leaders put about 10% of their resources into algorithms and 20% into technology and data. The other 70% went to people and processes. In 2025, BCG said companies must "redesign how work gets done."
Task-specific tools rarely reach production. MIT's NANDA study, "The GenAI Divide" (July 2025), looked at task-specific AI tools. 60% of organizations investigated them, 20% piloted them, and 5% reached production. The authors write that the divide "seems to be determined by approach," not model quality. The study is small and preliminary, with 52 interviews and 153 conference surveys, so read it as a direction.
Gains come where output is measurable. The Stanford AI Index 2026 finds the largest productivity gains in "structured, measurable work." It also finds that scaled use of AI agents was still in single digits for nearly all business functions.
The pattern is old. The essay The Commanding Heights by Will Manidis and Joe Roberts compares AI today to the first steam locomotives. They paid off first on captive routes at mines and ironworks, where the same loads already moved every day. The owners knew the work, the volume, and the cost, and the machine did the same job for less.
The Value Goes to Adopters
Economic history shows who gains when a technology spreads. Hornbeck and Rotemberg estimate a 48% annual social return on the $8 billion invested in US railroads in 1890. The railroads privately captured only 7% of that return.
The other 93% went to the farms, mills, and merchants that used the network. In AI, that is the companies that put models to work in their operations. Model makers will not keep most of the value of AI. The companies that change how they work will.
Redesign the Steps, Not Only the Tool
Two published cases show what a redesign looks like.
Uber: precision over volume. Uber's uReview reviews over 90% of about 65,000 code changes a week. The team scored each comment for confidence and removed whole categories of low-value comments. Developers act on 65% of its comments, against 51% for comments from human reviewers. The redesign was in what the tool chose not to say.
Ramp: people on exceptions. Ramp's policy agents review expenses against company policy. Ramp reports that they send only the 10% to 15% of expenses that need human judgment to a person. The finance team's job changed from reviewing every expense to reviewing exceptions.
The pattern is the same in both, and in large AI code migrations:
- Split the work into steps with a clear input and output.
- Add an automatic check after each step: tests, rules, or a confidence score.
- Send only exceptions to people, with the evidence attached.
- Measure every step, so you can see where the errors come from.
Our post on approval gates and exception routing goes deeper on step 3.
Pick the First Workflow
Start where the work already exists at volume. Good first workflows have high volume, written rules, an output you can score, a low cost of error, reachable data, and one owner. Our AI workflow readiness scorecard turns those criteria into a score. OpenAI's guide to building agents names refund approval, vendor security reviews, and home insurance claims as typical examples.
Measure Before You Build
Measure the workflow as it runs today, before any AI touches it. Track the same metrics after launch.
| Metric | What it shows |
|---|---|
| Cycle time per case | Speed |
| Cost per case | All-in cost, including model and human time |
| Error rate | Quality, from a sampled review |
| Share of cases sent to a person | How much the AI handles |
| Human minutes per case | Where your team's time goes |
Our guide to the AI workflow baseline shows how to collect these numbers in a week.
When a Better Model Does Help
A better model helps when the workflow is ready for it. You need a test set from your own cases, an automatic check after each step, and a gateway that can switch models. Then a new model is a measured upgrade: you run the tests and switch if it passes. Without those, a new model is a guess. This is why the redesign comes first.
Getting Help That Scales
Most companies lack the people to do this redesign. Hourly consulting does not solve that at scale, because growth is limited by hiring. Help scales when the delivery is a product.
| Hourly AI consulting | Productized deployment | |
|---|---|---|
| What the client buys | Engineer hours | A workflow in production, for a flat fee |
| What is reused across clients | Little | Tools, connectors, agent patterns, eval methods |
| Who runs the system later | The consultant | The client's team, or the provider as operator |
| What limits growth | Hiring | How fast the patterns improve |
Redesign Your First Workflow With OpenNash
OpenNash FDE Agents take whole processes off your team's plate, measured in hours saved and cost per case. Our forward-deployed engineers:
- Score your workflows and pick the first one.
- Measure the baseline before we build.
- Redesign the steps with automatic checks and human approval on exceptions.
- Ship to production in 4 to 8 weeks, one workflow at a time, for a flat monthly fee.
- Report the same metrics every week, then hand the workflow to your team.
Book a call with an OpenNash forward-deployed engineer or email us at hello@opennash.com. In the first session, we score your top workflows together. Then we show you which ones deliver the highest business value and the fastest ROI.