The gap
The capabilities are already mind-boggling. Frontier models reason through novel problems, write production-grade code, and hold coherent conversations across a million tokens of context. They analyse documents, summarise meetings, draft emails, and answer questions that would have seemed like science fiction five years ago.
Why does it still feel like AI isn’t doing anything in most organisations?
The answer isn’t the models. The answer is everything around them.
Three walls
Integration friction. Enterprise data lives in fifteen different systems, each with its own API - or worse, no API at all. Connecting them is expensive, fragile, and deeply unsexy. Nobody gets promoted for building a connector between Salesforce and the legacy ticketing system. So it doesn’t get done, and the AI that could transform workflows can’t access the data it needs.
Scattered documentation. Policies live in SharePoint, maybe. Exceptions live in email threads. Precedents live in people’s heads. The actual process - the way things really work - lives in the gap between the documented process and reality. An AI can only work with what it can see, and most organisational knowledge is invisible.
Workflows nobody wants to touch. Every organisation has processes that “just work” - fragile, undocumented, politically owned. Nobody wants to be the person who breaks the thing that’s been running for years. So these workflows persist, immune to improvement, while AI gets bolted onto the edges instead of transforming the core.
These aren’t technical problems. They’re institutional problems. And they’re why the gap between AI capability and AI implementation keeps widening.
The deeper issue: context is nowhere
You can give an AI access to your policies. You can connect it to your systems of record. You can even give it a clear goal. And it will still make decisions that feel wrong - technically correct but missing something essential.
What’s missing is context.
Not data. Context. The difference is everything.
Data tells you the policy says customers get 30-day payment terms. Context tells you this particular customer has been with you for fifteen years and always pays early, so when they ask for 45 days during a rough quarter, you say yes. Data tells you the escalation process requires three levels of approval. Context tells you that when Sarah from operations flags something, you skip the process because she’s been right every time.
This context currently lives in Teams threads, hallway conversations, sales desk calls, and people’s heads. It’s the institutional memory that makes organisations actually function. And it’s almost entirely inaccessible to AI.
Rules tell an agent what should happen in general. Context is what lets it exercise something like judgment - knowing when the rule applies, when the exception is warranted, and why. Without context, you get AI that follows rules rigidly. With context, you get AI that can actually help.
Captured context has its own failure mode: it rots. SharePoint becomes a dumping ground. CRMs accumulate three versions of the same customer. Policy documents drift out of date the moment they’re published. An agent fed stale or contradictory information will be confidently wrong - a harder failure to catch than a blank answer. Context isn’t a one-time capture. It’s a maintenance discipline. Whatever you feed the agent has to be kept current, or the gap just moves from “we have no context” to “we have context that lies to us.”
That’s the work. Not better models. Better context, delivered to the agent at the point of decision, and kept honest over time.
Where the leverage actually is
What can you actually implement next quarter that makes things work better?
Stop chasing frontier models. The capabilities already exceed what most organisations can absorb. A mid-tier model from two years ago is more than enough for 90% of use cases, if you could just get it connected to the right data with the right context.
The work is in the messy middle. Integration, documentation, workflow mapping - this is where the leverage is. It’s unglamorous. It doesn’t make good demos. But it’s the difference between AI as a toy and AI as infrastructure.
Start capturing context deliberately. This is why I built Arc. Not because I needed a chatbot, but because I needed a system that remembered context across sessions. The same principle applies at the organisational level. Every decision, exception, and precedent that gets captured is context that makes AI more useful.
The opportunity is implementation, not invention. The models exist. The capabilities exist. The connective tissue that lets them actually work inside real organisations doesn’t build itself - it takes people willing to do the unglamorous work, one integration, one documented process, one captured decision at a time.
The people who figure out context capture and integration will quietly build the most valuable AI infrastructure. Not by making better models, but by making existing models actually useful.
The gap isn’t capability. The gap is context.