Two years ago, the only term you heard at conferences was AI. Last year, it was agents. This year, the term I’m hearing a lot is context, or more specifically, the context graph. We’ve reached a point where customers expect to be delighted with the value you are providing, not just whether you have AI, but that AI understands their business. You can’t simply say, “Yeah, we have AI in our product, but you have to turn it on.” AI has to be always on and deeply integrated.
The industry has moved into the high-stakes reality of customers expecting AI to understand their specific business. This is the “last mile” of AI: a general purpose model might know almost everything about the world, but a specialized agent needs to know almost everything about your company.
Specialization wins when the cost of being wrong is high and success depends on customer-specific context. In the enterprise, “pretty good” is synonymous with “too risky.”
The last mile defines category winners
In all my conversations with CIOs, I see two types of companies: those building agents from the ground up to solve the last mile, letting them scale without tech debt and constraints, and legacy SaaS companies that are struggling to fit AI into their offerings. This divergence between companies that are doing agents well and those that are merely claiming to do AI is only going to get wider. Most CIOs aren’t sure what it means to build a good agent. Because proof points are now possible, building successful agents in production is going to be the number one priority this year and next.
This is already showing up in customer interactions and in revenue. Coding is a clear category winner. GitHub had the first mover advantage, but it didn’t really capture the market. Cursor came from behind: it didn’t have the best models, the best infrastructure, or the Microsoft muscle, yet it won the hearts and minds of developers with a small wedge and kept growing from that. Cursor doesn’t just know how to code; it knows your company’s codebase. As a result, Cursor reportedly surpassed $2 billion in annualized revenue before SpaceX acquired it for $60 billion.
The next category after coding is vibe coding, which is dominated by Lovable and Replit. Everybody else is far behind. Like in coding and vibe coding, vertical agents are creating category winners. In legal, there’s Harvey, which hit a $190 million ARR in December. In customer support, there’s Sierra, which hit a $100 million ARR in seven quarters after its launch, followed by $50 million in one quarter. These startups have clearly established product-market fit and are scaling. They address the last mile problem in their domains by using AI to understand the specific systems, workflows, dependencies, risks, and business context inside each company.
Every category is going to find a winner. The IT automation category, for example, is still in its infancy, which means there will be winners when the right startups figure out the last mile to capture the category.
Why vibe coding fails at scale
The reason category winners exist is because not all AI tools are created equal. ChatGPT and Claude are accessible, but they rely on vibe coding, which is ineffective at solving the last mile problem. It’s easy to write a few prompts and say, “Hey, look, I have an agent that gave me a pretty good answer.” Unfortunately, because that answer is not grounded in your company’s data, context, and infrastructure, it’s effectively made up.
In IT, this shows up as infrastructure hallucination. Foundation models don’t understand your network topology. They can tell you how a Cisco switch works in theory, but they don’t understand how your switches are connected to your routers and your applications, not to mention the names of your applications. All of this is completely unique to your company, not general knowledge that LLM makers can scrape from the web. Without a context graph, an agent is just a tourist in your data center.
The worst-case scenario isn’t that a general agent misses the whole picture, but that it breaks something critical. Last year, a tech entrepreneur reported that Replit’s AI coding agent deleted a live company database during an active code freeze, and last month, the founder of car rental platform PocketOS detailed how a Cursor agent using Claude Opus 4.6 deleted a production database while in a staging environment. This is the price of ignoring the last mile.
You can’t vibe code your way into production. Vibes aren’t a substitute for deterministic logic. Vibe coding is good for prototyping, but when you’re dealing with production, you’re dealing with mission-critical work. Lives and people’s livelihoods are at stake. You don’t want to casually depend on vibe coding tools in production environments.
The horizontal threat (and its weakness)
That said, vibe coding has fueled several SaaSpocalypse market events, which ultimately come down to whether horizontal AI companies can go vertical. People are nervous about whether the foundation model makers, the AI startups that can do the horizontal play well, will also be able to help any enterprise build agents for their unique business.
That’s the horizontal play, and it’s scary. It’s hard for companies trying to play in that market because the model providers are going up the stack into the app layer, as Forrester and many others have noted.
Where you have room to capture and differentiate is on the vertical side because at least there, you can have domain expertise. To be fair, even that is questionable. Could OpenAI or Anthropic develop a model that is good enough to solve domains like legal, HR, and IT? The school of thought was that these models are too generic, but they have been getting better over time. Some of the latest foundation models are great at many tasks to the point where I’m inclined to call Claude Opus 4.8, for example, a horizontal agent.
At the same time, startups like Harvey and Sierra are still creating plenty of value on the vertical side. These businesses have created enough verticalized integration to claim their respective moats. Enterprise value depends on how well models are grounded in each company’s systems, workflows, and operating context. That said, the jury is still out on whether the horizontal model makers can simply release plugins in various verticals to take on the companies that specialize in each domain.
Specialization always wins in the end
Can the horizontal players win across all the vertical markets? Nobody knows, but I would rather bet on winning in one vertical than try to take on the horizontal players. Instead of bridging the last mile with people, I’d pick a platform that bridges it with data.
For IT automation, you don’t need a subsidized engineer to manually tune a model, just a platform that inherently understands the domain expertise of your specific vertical. The context graph is such an industry hot term right now because to build effective agents, you need to feed context into the LLMs, not just prompt your favorite LLMs.
The more you can discover a company’s context, the more you can solve the last mile problem. This is why specialization always wins; at some point, the generic approach just doesn’t scale.
