What Businesses Should Actually Watch in the State of AI Report

There are plenty of big predictions in there, but what interested me most were a few quieter trends that could have a much more immediate impact on companies building and selling technology around AI like where that growth is happening, where the bottlenecks are moving and which parts of the market become more valuable as AI gets cheaper and easier to access.

Here are a few of the things that stood out to me.

One of the numbers I kept coming back to was the enormous gap between average AI users and the heaviest adopters.

The report cites Ramp data showing that the median company in its sample spent about $12.50 per employee per month on AI. Among the top 1% of spenders, the median was more than $7,000 per employee per month. Another analysis found that roughly 1% of customers accounted for around 80% of observed OpenAI and Anthropic spending.

There are plenty of caveats around a dataset like that, but the gap itself is hard to ignore. There are companies using AI as another productivity tool, and there are companies starting to build significant portions of their operations around it. Those businesses are likely to have very different needs, budgets and expectations.

For companies selling into the market, that distinction matters. The customer experimenting with a handful of AI tools is not the same customer as one running AI across core workflows. They may care about completely different things and buy on completely different timelines.

That also means the usual “XX% of companies use AI” statistic is becoming less useful. The much more interesting measure may be intensity of use.

For a while, access to powerful AI models was itself a meaningful constraint, but that is becoming less true. The report points to improvements that come not from changing the model, but from improving the tools, context, feedback and systems around it. Across the report, you see the same pattern in other areas: security, permissions, infrastructure, power, data and integration all start to matter more as AI moves into real-world use.

That changes where the problems and opportunities are.

Once a company has access to capable AI, the next questions become more practical, like: Can it connect safely to the systems it needs? Can it access the right data? Can you verify what it does? Can you control what actions it takes? Can the infrastructure support it economically at scale?

The security examples in the report are a good illustration. Once AI agents can actually interact with external systems, the issue isn’t simply whether the model generates a bad answer, it’s whether the system has permissions to do something it shouldn’t.

That’s a very different problem, and for a lot of technology companies, it may be the more important market opportunity.

The report also makes clear just how quickly the economics of AI are changing. The cost of achieving a given level of performance has been falling dramatically. If increasingly capable general-purpose models become cheap and broadly available, some things that once looked like standalone products may start to look more like features. Capabilities that were difficult or expensive a year ago may become easy to replicate.

That puts pressure on companies whose differentiation depends primarily on access to the model itself.

At the same time, it could make other parts of the stack more valuable. Proprietary data, infrastructure, security, workflow integration, specialized expertise and access to physical systems become harder to replace just because the underlying model gets better. That is something I think a lot of companies should be watching closely. The question isn’t simply whether AI makes your product better, it’s whether cheaper and more capable AI makes your existing differentiation stronger or weaker.

All of this also makes me wonder whether some of the ways we currently measure AI are becoming outdated. Take adoption, if almost every company uses AI in some form, saying that AI adoption reached 80% or 90% doesn’t really tell us much.

Model benchmarks have a similar problem. They can tell us something about technical capability, but not necessarily how useful that capability is inside a business.

Even AI spending can be misleading if a relatively small group of companies accounts for a huge share of consumption.

The market probably needs to move toward more useful measures: AI spend per employee, percentage of workflows touched by AI, number of autonomous tasks completed, amount of human oversight required, or actual business outcomes attributable to AI.

Of course, those measures are harder to collect, but they would probably tell us much more about how far AI has actually penetrated an organization.

The biggest takeaway for me from the report isn’t whether AI keeps getting better (that seems like a safe assumption), it’s how the market changes as access to capable AI becomes less scarce.

• Which companies will move from experimenting with AI to actually restructuring around it? 
  

• Which parts of the technology stack become more important as the models become more interchangeable? 
  

• Where do new constraints show up? 

• And which companies discover that what looked differentiated a year ago has suddenly become a feature?

Those are the questions I think will matter most to a lot of the companies we work with.

At Voxus, we spend a lot of time helping technology companies figure out where they fit into fast-moving markets like this and how to explain why their piece of the stack matters. As AI gets broader and noisier, that positioning is only going to get harder, and more important.

If you’ve got the time, the full State of AI Report 2026 is worth a read, but hopefully you’ll find these brief takeaways worthwhile.