The capability overhang in AI

“The future is already here—it's just not evenly distributed” - William Gibson

If you’re in tech, you have a keen sense of what’s happening in the world of AI agents right now. We’re seeing example after example of coding agents that can handle longer running tasks and take on much larger development projects, with teams increasingly calling out that their products are now written entirely with agents. The models have gotten insanely good at coding, and the agentic systems used for coding have quickly accelerated ahead to transform this particular domain of work.

Yet, when you talk to knowledge workers outside of the tech industry (even even some within), it’s pretty clear how early we still are; AI is often an incredibly helpful assistant that can answer quick questions and help look up information, though the production of large work output and automation is in its infancy. Even with today’s technology, there’s still so much more that’s possible, and as Ethan Mollick (@emollick) pointed out, “we really could stop AI development right now and it would still transform a substantial portion of white collar work … even given current models' limitations.”

And even these limitations are quickly disappearing. Increasingly, frontier AI models are being trained on almost every area of knowledge work, and they will only get more and more capable over time. They know esoteric legal topics, know deep areas in healthcare and life sciences, are increasingly getting good at creative fields like marketing and writing, and for anything they don’t know off the bat they can instantly pull up tools and talk to data sources for new information they need. They will almost assuredly surpass any human expert within the next couple of years in their respective field, at least in discrete tasks.

As AI models get smarter and smarter on various domains of knowledge work, the big opportunity now, for companies building and implementing agents alike, is how do we take the breakthrough capabilities coming out of the AI models, and apply them to enterprise workflows in a way that delivers real agentic work in the environments where knowledge work happens.

The context gap for agents

One of the biggest challenges to deploying agents is still getting the right context to agents. @buccocapital identified the problem for most agentic workflow automation quite succinctly as: “context is now (and perhaps always been) the bottleneck on growth, because anything with full context can be automated. The race is on to make the collective brain of the organization legible to AI.”

For most enterprises that were started before a world of agents, the context that agents need for doing their work is often not in a clean format readily available to agents. Dwarkesh Patel, in conversation with Dario Amodei, hypothesized that coding remains different from other areas of work because so much of the critical context for the work to be done is living right inside the codebase. Unlike in coding, in other areas of knowledge work, critical context is coming from fragmented systems, video calls, in-person meetings, external events happening in the world around the company and its customers, and so on.

And even when the context is digitized, most enterprises are still entrenched with many legacy and fragmented systems. Despite years and years of cloud growth, there's still a significant portion of software, data, and infrastructure that is stuck inside of on-premises or clunky systems that can't easily talk to agents running in the cloud. We see this all the time at Box, where companies still have large unstructured data repositories sitting inside of systems that will never play well with agents, and they’re finally looking to move these to the cloud. Over the coming years, there will be an aggressive move to modernize these platforms in cloud-native and AI-native architectures, but this still is a dependency for agents to be able to roam freely across an organization to automate work.

Then you have the issue of permissions and access controls. In most organizations, no two users have access to the same information. This is by design. A manager has access to the compensation details of their direct reports, but not their peers; a consultant can see the projects they’re working on, but not that client’s competitor’s in the same firm. There are limitless permutations of access controls in an organization that prevent the people from having access to the wrong information, which introduces all new complexities in a world where agents are doing 100X or 1000X more work within systems than people. In some scenarios agents should have access to far more than the user traditionally has, and in other situations an agent should have far less access than what the user technically has. Both become hard problems to ensure proper management around.

Finally, enterprises are dealing with a rapidly changing technical landscape. Right as they implement one solution, the best practices change, requiring another overhaul. Should an enterprise index all of their data in a vector database and use RAG to retrieve information? Should an agent tap into MCP servers from a variety of different SaaS tools? Should code execution agents be leveraging CLIs of different systems and have full capabilities on those platforms? In the span of 18 months, you could have landed on any one of these architectures and fully standardized your AI approach downstream from this, just to find out the world has yet again shifted the goalpost of what best in class looks like.

The massive opportunity ahead

The companies that can deliver on the right software and user experiences to make it easy for end users to be able to adopt agents and get them the right context will get ahead the farthest. Much of the adoption challenges will come down to how easy it is for non-technical users to take on the role of being a manager of agents within an organization. Just as coding tools have had to go from AI-assisted code writing in an IDE, to experiences that instead focus on managing and overseeing the work of agents, the same will have to happen in all areas of knowledge work, from legal and finance to consulting and manufacturing. This is not going to be an intuitive process for most knowledge workers, so it will mean that the software that makes this easy will tend to be the ones that win over time.

Some of these solutions will come from existing platforms that are deeply tied to existing workflows and data, and that can connect the intelligence of models to their domain with the right context (or harness) engineering to solve the customer’s problem in an end-to-end fashion. We’re focused on this at Box, securely connecting critical enterprise content to agents on and off our platform, to accelerate all domains of knowledge work. Equally, there will be huge opportunities for new players that reimagine workflows entirely in various verticals or lines of business use-cases, as well as new approaches to connecting across systems as a horizontal layer of intelligence for doing work.

There will also be a huge opportunity for the AI companies that understand how to deliver on the change management across the organization for these workflows. This is especially why the FDE trend is not going away anytime soon. Most organizations need real help in getting their environment set up for a world of AI agents, and this will be a heavily professional services-driven operation for most companies. They will need the support of teams to come in and deeply understand their workflows, and figure out how to best apply AI agents to these processes, and get their data in the right setup to power this.

This also creates a huge opportunity for new forms of system integrators and consultancies (and traditional ones that jump on the opportunity), because they will be the ones often doing the heavy blocking and tackling within organizations to make sure those companies are ready for AI in the workplace. The firms that already have existing relationships with large enterprises will be often in the best position to help those companies transform their work with AI agents, assuming they can rapidly evolve to support these practice areas. Don't write off the SIs yet.

And finally, this is a huge opportunity for resourceful and entrepreneurial talent within organizations to go in and reimagine workflows for a world of agents. Much of the change management needed to get agents the context they need to work with is really about retooling the underlying workflow to no longer be dependent on the tribal knowledge or analog ways that people work together. You will see a huge growth of roles within enterprises, and people that specialize in this will be hugely valuable in the economy.

A brief note on technology diffusion

It’s worth noting that even with the best architectures, data in the right formats, and up-to-date standards in place, diffusion still takes time. Because there are major workflows that need to be reinvented, data needs to get into well-organized environments, there’s technical literacy that needs to be established, change management needs to be executed to get users onboarded, and any regulated or large business has huge governance processes for deploying new tech or agents. Just to get agents deployed on a specific workflow in an organization, for instance, think through the necessary processes that you’d have to go through in a complex organization doing mission critical and your brain will hurt.

Here’s a point of comparison for technology diffusion of one of the fastest growing technologies in history: the cloud. In 2010, a time by which every person in silicon valley knew that cloud was the future, AWS revenue was estimated to be around $500 million, Azure had only just launched that year, and GCP was called Google App Engine. Fast forward to 2025, and these 3 platforms generated around $225 billion in revenue. And that’s only about 60% of the cloud market. So from the moment the tech industry saw the future of cloud to today, the market is nearly 1,000 times bigger. And it’s still growing at an insane rate.

Even in the fastest changing markets, diffusion of a new technology takes time to ripple through the economy. Agents will be no exception. This is the stuff that takes time in the real world of work, but that's exactly why there’s still so much opportunity ahead.

Overall, we are in an incredible moment where AI capabilities are increasing at an insane rate. We're going to be able to solve larger and larger tasks within more and more critical areas of knowledge work. This diffusion will necessarily take time, but the winners will be those that are able to bridge these breakthrough innovations with how the real world actually operates.