Here's the most underappreciated investment a company can make right now: writing things down.
Not tickets. Not Jira descriptions. Structured, permanent, findable reference documentation that captures the hard-won knowledge your team has built over years. The kind of knowledge that currently lives in people's heads, in Slack threads that will eventually disappear, and in the tribal memory of engineers who've been around long enough to know where the bodies are buried.
This is the context layer that turns a generic AI agent into one that actually understands your business.
The documentation failure
Every time someone explains something in a Slack thread that they've explained before, that's a documentation failure. More importantly, it's an AI capability gap.
I've had this conversation with several founders recently. The pattern is always the same: they've given their teams access to every AI tool available — unlimited budgets, ChatGPT, Claude, whatever they want to try. But the results are inconsistent. The AI generates plausible-looking output that misses critical context. An engineer prompts Claude to build a feature, and the code compiles, passes tests, but subtly misunderstands a domain requirement that only three people in the company know about.
The problem isn't the model. The problem is the context.
AI consumes written text, not knowledge in heads
This is the insight that changes everything once you internalize it. AI operates on what's written down. If your pricing model assumptions live in an actuary's head, the agent can't use them. If your carrier appetite guides are spread across 47 email threads, the agent can't reference them. If the reason you structured a coverage a certain way is only known to the person who made that decision two years ago, the agent will make a different decision — and it might be wrong.
The companies getting the most leverage from AI aren't the ones with the best prompts or the most expensive tools. They're the ones with the best documentation. Period.
What this looks like in practice
At Shepherd, we're treating institutional knowledge as a product. Not a nice-to-have. Not something we'll get to eventually. A deliberate investment with the same priority as shipping features.
Concretely, this means:
- Carrier appetite guides documented in a structured, searchable format
- Regulatory constraints by state captured as reference documentation, not tribal knowledge
- Pricing model assumptions written down explicitly — the inputs, the logic, the edge cases that burned us before
- Architectural decisions documented with the why, not just the what
- Domain knowledge that used to only surface during onboarding conversations, now captured permanently
The goal is a system where a new engineer — or a new agent — can get oriented without needing to ask someone who's been at the company for three years.
The adoption stair-step
I recently spoke with a founder who's been particularly effective at driving AI adoption across his org. His framework is a stair-step:
- Opportunity — provide the tools, processes, and support. Make it easy.
- Champions — find early adopters and celebrate them publicly. Give them platforms to demo and teach.
- Requirement — progress to "you must use this."
- Accountability — non-adoption has real consequences.
But here's what he emphasized most: none of these steps work if the documentation isn't there. You can mandate AI usage all day, but if the agent doesn't have the context it needs, people will bounce off the tools and conclude they don't work. The documentation is the adoption strategy.
AI office hours and hackathons
Some tactical things that have worked, both for us and for other teams I've talked to:
- Weekly AI office hours where anyone can bring questions, share workflows, or demo something they've built
- AI-focused hackathons — pause the company for 2-3 days and build with AI. Nothing creates champions faster than hands-on experience with a real problem.
- Unlimited AI tool budgets — remove the friction entirely. Let people experiment with whatever tool works for them.
But the through-line across all of these is the same: the tool is only as good as the context. Invest in the context.
The compounding advantage
Here's why this matters strategically, not just tactically. Documentation compounds. Every piece of institutional knowledge you capture makes every future AI interaction more effective. Over time, the gap between a team with rich documentation and one without becomes enormous — and it's not a gap that can be closed quickly.
Your competitors can adopt the same AI tools tomorrow. They can't replicate your institutional knowledge. That knowledge, captured in a format AI can consume, is a durable competitive advantage.
The code is becoming commoditized. The context is not.
