Resources
Before an AI system can be useful to your company, it needs to know how your company works. That sounds obvious, but most companies skip this step and then wonder why their AI tools give generic answers. This is the unglamorous part that actually determines whether your investment pays off.
Dry Ground AI | Practical Guide
The Foundation
There's a useful distinction between what your company knows and how your company thinks. Both matter, and they serve different purposes.
The first layer is facts: what your company does, how processes work, who owns what. This is the handbook you'd hand a sharp new hire. The second layer is reasoning: why you made the decisions you made, what you considered and rejected, what lessons you've learned. This is the institutional memory that usually lives in people's heads and disappears when they leave.
Most AI implementations only capture the first layer. They build a searchable FAQ. That's useful, but it's not what makes AI genuinely valuable. An AI system that knows your facts can look things up. An AI system that knows your reasoning can actually help you think.
The facts. What your company does, how it works, who does what, what tools you use. The kind of information that should be in a good employee handbook but rarely is.
The reasoning. Why decisions were made, what was considered, what worked and what didn't. This is what lets AI reason about new situations instead of just retrieving old answers.
Layer one
Think of this as the handbook you'd give a smart new hire on day one. The test for whether something belongs here: if a new employee would need to know it to do their job, document it.
Here are the categories that tend to matter most:
| Category | What to include | Example |
|---|---|---|
| Company profile | Industry, size, locations, org structure, mission | "We're a 45-person logistics company in Dallas..." |
| People & roles | Key team members, who owns what, reporting lines | "Sarah Chen handles all vendor negotiations..." |
| Processes & SOPs | How recurring work gets done, step by step | "When a new client signs, we..." |
| Tools & systems | Software stack, what's used for what, where to find things | "We use HubSpot for CRM, Notion for docs..." |
| Terminology | Industry jargon, internal shorthand, acronyms | "QBR = Quarterly Business Review..." |
| Client info | Key accounts, relationship history, preferences | "Acme Corp is our largest account, contact is..." |
| Policies | Approval thresholds, compliance rules, what not to do | "Anything over $5k needs VP approval..." |
| Preferences | Communication style, meeting cadence, team norms | "Keep emails under 3 sentences..." |
One topic per document. This is the rule that gets violated most often. When you put your entire company handbook in one file, an AI system can't reason about the pieces separately or understand how they relate to each other.
Each document should have a clear title (what this document is about), 2 to 5 tags for categorization, who owns and maintains this information, what other documents it connects to, and then the actual content written in plain language.
A concrete example:
Title: New Client Onboarding Process
Tags: sales, onboarding, process
Owner: Sarah Chen
Depends on: Sales Handoff Checklist, Contract Templates
Triggers: Account Setup in HubSpot, Welcome Email Sequence
When a new client signs their contract:
1. Sales rep completes the handoff form within 24 hours
2. Ops team creates the client workspace in Notion
3. Account manager sends the welcome packet
4. Kickoff call scheduled within the first week
Common mistakes to avoid:
- Don't skip the handoff form. It causes gaps every time.
- Don't schedule the kickoff before the workspace is ready.
- Always confirm the client's preferred communication channel.
Tips that actually matter
Layer two
This is the harder one, but it's where the real value is. Without context, AI follows your SOPs mechanically. With it, AI can reason.
Imagine an AI that knows your company decided to switch from Salesforce to HubSpot, and knows why: licensing costs were too high, you weren't using most of the features, and the marketing hub integration let you drop another tool. Now when a client asks whether they should build a custom CRM, the AI can draw on that reasoning, not just retrieve a fact.
Document significant decisions using this structure:
| Field | What to write |
|---|---|
| The decision | What was decided, and when |
| The rationale | Why this option was chosen over the others |
| Alternatives considered | What else was on the table and why it was rejected |
| Outcome | What happened as a result (if you know yet) |
| Confidence | How settled is this? (tentative / developing / established) |
| Connections | What other decisions does this relate to, support, or contradict? |
A concrete example:
Title: Why We Switched from Salesforce to HubSpot
Tags: tools, crm, decision
Date: 2024-09
Confidence: established
THE DECISION:
Migrated from Salesforce to HubSpot CRM in Q4 2024.
THE RATIONALE:
- Salesforce licensing was $2,400/user/year. HubSpot Pro is $1,200/user/year.
- Our sales team (8 people) used maybe 20% of Salesforce's features.
- HubSpot's marketing hub let us drop Mailchimp ($400/mo).
ALTERNATIVES CONSIDERED:
- Stay on Salesforce, negotiate pricing: Rejected. Even at a discount, paying for features we don't use.
- Pipedrive: Rejected. Cheaper but no marketing automation.
- Build custom on Notion: Rejected. Not a CRM. Would break within 6 months at our growth rate.
OUTCOME:
Saved ~$38k/year. Sales team adopted within 3 weeks. Lost some custom reporting that we rebuilt in 2 months.
CONNECTIONS:
- Led to: Marketing Automation Overhaul
- Supports: "Buy vs Build" philosophy
The graph
A folder of documents is a library. A set of connected documents is a knowledge graph. The connections are what let AI systems navigate your knowledge instead of just searching it.
When writing each document, note the obvious connections. You don't need to get these perfect.
| Connection type | Meaning | Example |
|---|---|---|
| depends_on | You need to understand X to understand this | Onboarding Process depends on Contract Templates |
| triggers | When this happens, that process starts | Signed Contract triggers Onboarding Process |
| supports | This reasoning reinforces that reasoning | HubSpot Decision supports Buy vs Build Philosophy |
| contradicts | These are in tension | "Move Fast" contradicts "Everything Needs Legal Review" |
| supersedes | This replaces an older decision | New Pricing Model supersedes Old Pricing Model |
| led_to | This decision caused that decision | Salesforce Cost Analysis led to HubSpot Migration |
Getting started
You don't need to boil the ocean. A focused set of 15 to 30 documents is enough to make an AI system genuinely useful. Here's how to approach it.
One document per topic. Not one mega-document. 20 focused files beat one sprawling one.
Use a consistent structure for each doc: title, tags, owner, connections, then content.
Do knowledge documents first, context documents second. Get the facts down, then add the reasoning.
Start with what your team asks about most. If people are always asking "how do we handle X," that's your first batch.
Don't chase formatting perfection. A rough document with the right information beats a polished one that says nothing.
What comes next
The documentation itself isn't the end goal. It's the raw material. From here, the process typically looks like:
Convert your documents to structured formats with proper metadata
Build relationship edges between documents to form the actual graph
Generate vector embeddings for semantic search
Wire the knowledge base into your AI systems so they can search and reason across it
Let the AI's conversations with your team continuously add new entries over time
The better your starting material, the faster an AI system gets useful. But don't let perfect get in the way of good. A rough document with the right information is more valuable than a polished one that says nothing.
We help companies build knowledge and context graphs as part of our AI implementation work. If you want a structured process instead of figuring it out on your own, we're happy to walk you through it.
Get in touchWe use cookies to improve your experience. Cookie Policy · Privacy Policy