Seventy-one percent of organizations now say they use generative AI regularly in at least one business function, up from 65% in early 2024, according to McKinsey. Ask most of those same companies what their AI assistant actually knows about their own contracts, their pricing history, or last quarter's board deck, and the honest answer is nothing. The model is fluent. It has no memory of the business running it.
That gap is the opportunity for founders this week, and it is smaller to close than most people assume.
Why a fluent model still gives a useless answer
"The AI models themselves don't really understand anything about your business... you have to connect the reasoning and generative power of the models with the context inside your company," Glean CEO Arvind Jain told TechCrunch in February 2026.
A general purpose model was trained on the public internet. It was never trained on your SOPs, your last ten customer contracts, or the reason your team stopped selling the enterprise tier in 2024. Without that material attached, every answer it gives is a guess dressed up as confidence.
This is also why a specific market now exists to solve exactly this problem. The retrieval augmented generation market, the technical name for systems that ground AI answers in a company's own documents, was estimated at $1.2 billion in 2023 and is projected to reach $11.0 billion by 2030, a 49.1% compound annual growth rate, per Grand View Research. Enterprise search and AI assistant startup Glean raised a $150 million Series F at a $7.2 billion valuation in June 2025, led by Wellington Management, according to CNBC, betting that most companies would rather buy this connective layer than build it.
Two ways to close the gap, starting this week
You do not need a $7.2 billion vendor relationship to get started. Two paths are already sitting in tools your team likely has access to.
The first is do it yourself. OpenAI shipped "GPTs" in November 2023, letting anyone build a custom version of ChatGPT with no code, just instructions and uploaded documents as extra knowledge. Anthropic followed in June 2024 with Projects for Claude.ai, which gives each project a 200,000 token context window, roughly the length of a 500 page book, so you can load your handbook, your top sales objections, your last twenty support tickets, and get answers grounded in that material instead of the open internet.
The second is buying the retrieval layer outright, which is the bet Glean and its investors are making: that connecting company knowledge to whichever model you use is hard enough, and valuable enough, to be its own category rather than a side project your engineering team maintains.
For most founders, the do it yourself path is the correct starting point. Not because it is cheaper, though it is, but because it forces you to find out what your team actually asks before you commit budget to a platform.
- Pick one narrow, recurring question your team answers from memory or by digging through Slack: how you price a specific deal size, how you respond to a specific objection, how a new hire finds the current version of a process.
- Gather the real documents that answer it, ten to twenty files, not everything in the drive.
- Load them into a Claude Project or a custom GPT.
- Test it with the actual questions your team asks, not hypothetical ones.
If the answers hold up against someone who already knows the material, you have a working prototype of the thing McKinsey's 71% are supposedly already using. If they don't, you've learned which documents are missing, contradictory, or out of date, the same audit you would need before buying anything from Glean or a competitor. Either way, you know more about your own knowledge gaps by Friday than you did on Monday.