Context is king when it comes to AI
AI tools won’t unlock real value or recurring revenue until they can automatically understand user context — turning today’s “vibe-driven” hype into tomorrow’s invisible, truly useful infrastructure, says Lewis Liu
There’s a funny TikTok/Instagram meme going around, clearly targeted at people like me, where a male influencer interviews attractive women asking, “What’s most attractive about a guy?” The women all invariably answer, “When they’re building vertical AI agents, of course!”
Naturally I am building vertical AI agents, and here’s my take on what is a far more complex task than just “building vertical AI agents”. As a side note, the term “AI agent” is super overused, but for the sake of this conversation, let’s define it as an AI application with some kind of conversational interface that can execute specific tasks: making slide decks, inputting Salesforce entries, redlining contracts, etc.
I was recently having lunch with a senior partner at a top law firm when the conversation inevitably turned to AI agents in legal practice. He described how whilst lawyers are experimenting with and deriving some initial value from legal tech AI tools like Harvey and Legora, for the vast majority of non-document data extraction tasks (which is mature technology that my first start-up Eigen pioneered), it takes longer to prompt these new legal AI tools than to do the job yourself. The example he gave was redlining: it takes longer to prompt the Harvey redline agent than, as an experienced lawyer, to redline the contract himself. Digging into tools like this and talking to other lawyers, I noticed it’s because it often takes enormous effort to copy and paste the right context into the prompt window in the right format. Ensuring your redlining agent has the right context often takes longer than just redlining the contract yourself.
Expanding beyond the legal space, I’ve noticed this is a common issue across many of the “fastest growing AI” tools. I was talking to a management consultant who complained that none of the slide deck generation tools – Google Gemini Slides, PowerPoint Copilot, or Gamma – really work because they just generate “generic BS”. Again, my experience using these tools is that they do generate great slides, but you need 500-1000 words in your prompt, including all your context. At that point, you might as well make the slide deck yourself if you’re a PowerPoint master. Full disclosure: I am also an ex-McKinsey consultant; thus, as a PowerPoint nerd, I naturally figured out how to make this work with AI.
Vibe revenue
In the broader venture capital context, investors are worried about what’s called “vibe revenue” versus “annual recurring revenue”. Annual recurring revenue is revenue you can count on recurring annually due to either multi-year contracts or highly predictable renewals. The problem is “vibe revenue”: a lot of these AI tools see huge sign-ups and initial growth rates, but customers sign up due to FOMO and don’t renew after trying it out for a few months. Many of these tools are “AI wrappers”: just some thin interface on top of GPT, often dubbed “ChatGPT for X,” without real workflows or context. So, people are buying AI for the sake of AI, but didn’t buy a product that uses AI to solve a real problem. As a result, they turn off the service after a few months.
So why is this context problem so widespread? The penny dropped for me when I was talking to an AI marketing expert friend of mine who is a “fuzzy”. Context: my wife went to Stanford, and apparently they divide people into “fuzzies” and “techies”. He asked me, “Are you getting all these ads on Instagram or TikTok on how to use AI better, how to prompt AI better?” I said, “No, I didn’t even realize this was a thing!”
See, for fuzzies, history majors who become lawyers or econ majors who become consultants, GenAI prompting doesn’t come as naturally as it might seem. Using the full power of GenAI still only pertains to a small population. Ensuring the AI has full context takes both time and a certain kind of technical training. Prompting, despite being in “natural language,” is not 100 per cent intuitive for professionals. For true adoption, AI applications need to become more user-friendly.
This is why so much AI revenue is still “vibe revenue” and not “recurring revenue.” To lock in true recurring revenue in AI, these AI agents need to do two things: (a) be designed to solve specific business problems (not just thin GPT wrappers) and (b) handle specific context in a highly nuanced way.
Building more verticalized AI agents (AI applications that focus on specific domains) is already a hot area of focus for venture capital: hence the memes I came across on TikTok. Handling specific nuanced context, however, is still a nascent field, and I think will be the next big wave in AI investment and product development.
As an example, let’s say we want to build a loan underwriting AI agent. Today’s AI tools ingest standard application data: financial documents, business plans, annual reports, news articles, etc., then make underwriting decisions. But imagine adding a contextual layer that captures the human underwriter’s working knowledge: their client emails, Slack conversations, calendar patterns, and decision history. Now the AI doesn’t just process documents; it understands the underwriter’s judgment calls. It knows that this client type usually over-forecasts revenue, or that similar deals required additional collateral based on past patterns. The AI learns not just what data to analyze, but how experienced humans actually make these decisions in practice.
The new set of companies that can turn deep user context into a ‘contextual layer for AI’: to make the AI easier to use: will be the next wave of unicorns. To be clear, this isn’t easy: there are technical challenges around data integrations and dealing with ‘AI memory’ (something I’ve written about previously at length). However, those who can solve context for AI will truly unlock the next phase of AI adoption.
The difference between today’s AI tools and tomorrow’s will be simple: today you feed context to AI; tomorrow, AI will already know your context. When that happens, we’ll stop talking about “AI adoption” and start talking about AI as invisible infrastructure: like electricity or the internet today. Until then, feeding AI context remains king.