How to use AI without losing your institutional knowledge

Here’s what happens in a lot of small businesses right now.

Someone on the team asks Claude or ChatGPT how to handle a recurring task — closing out the register, prepping a client proposal, onboarding a new hire. The AI gives a solid, confident, generic answer. The person tweaks it slightly, uses it, moves on. The answer never gets written down. The next person asks the same question. The AI gives a slightly different answer. The output drifts. Nobody notices until the fifth or sixth time, when a customer complaint reveals that your team is doing the same thing five different ways.

The AI isn’t the problem. The lack of ownership over the AI’s answers is the problem.

What “AI + human hybrid” actually means for an SMB

The hype version of hybrid AI sounds like a Gartner slide: “AI handles the repeatable work, humans handle the exceptions.” That’s true at a 30,000-foot altitude and useless at a 10-person company.

The ground-level version is this: AI is extraordinary at producing a first draft of how to do something, and terrible at knowing how your specific business actually does it. The hybrid part is picking up where the AI stops — turning the generic draft into the version that reflects your decisions, your tools, your edge cases, your clients.

If that work doesn’t get captured somewhere persistent, you’ve just outsourced your institutional knowledge to a system that forgets by design. Every chat is a new chat. The AI doesn’t remember the five tweaks your team made last Tuesday. Your team does — until they leave.

The quiet failure mode

Six months into “we use AI for everything,” the pattern we see in SMBs looks like this:

• No one owns the “right” version of any given process. There’s a Claude conversation from February, a ChatGPT thread from July, a Notion page a contractor started, and three Slack messages where someone said “actually, we do it like this now.”
• New hires lose a week trying to figure out which version is authoritative. They give up and ask the AI, which gives them yet another slightly different answer.
• When someone leaves, the loss is worse than before AI, not better. Before, at least the knowledge was clearly in a human’s head and you knew who to ask. Now it’s splintered across three AI tools, two notes apps, and the person who’s walking out the door.

This is the bus-factor problem with an extra layer of plausible deniability. Everyone feels like the knowledge is “in the system.” It’s not. It’s in the ether.

The fix is dumber than it sounds

Pick one place where the authoritative version of each process lives. Make it the thing your team opens first, not the AI. Treat the AI as a drafting partner, not a storage layer.

That’s the entire mental model. Everything else — tool choice, workflow, team buy-in — follows from that one commitment.

The practical shape of this at a 5-50 person company:

1. The person doing the work drafts with AI. Claude generates the first version in 90 seconds.
2. They edit it to match reality. Our tools, our clients, our actual way of doing it. This is the 10-minute part.
3. It lands in the SOP library — the one that the whole team opens when they need to do that work. Not a draft doc, not a Slack thread, not a chat window.
4. Next person who touches it edits in place. The AI drafted version 1. The human owns versions 2 through N.

The AI contributes speed. The human contributes ownership. The library contributes memory.

Connect Claude directly to your SOPs

Here’s the shortest path to the hybrid workflow I just described, if you use Claude Desktop or Claude Code.

We built a Model Context Protocol (MCP) server that lets Claude read, create, update, and publish processes in your What’s the Process For library — using your login, with your permissions, nothing more. Setup is four lines in Claude’s config file:

{
  "mcpServers": {
    "whatstheprocessfor": {
      "command": "npx",
      "args": ["-y", "@whatstheprocessfor/mcp"],
      "env": { "WTPF_API_KEY": "wtpf_your_key_here" }
    }
  }
}

Mint a key at whatstheprocessfor.com/account/api-keys — there’s an “Install for Claude Desktop” button that generates this block pre-filled with your key. Paste it into ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows), restart Claude, and ask: “What’s my whoami on whatstheprocessfor?” If Claude answers with your account name, you’re connected.

After that, the workflow looks like:

You: “Hey Claude, draft a new-client onboarding SOP for our agency. We do website redesigns, 6-week projects, mostly SMB. Use our existing contract-review process as a reference.”

Claude: (reads your contract-review process from your library, drafts the onboarding SOP, creates it as a private draft in your workspace)

You: (open it in our editor, edit for 10 minutes to match reality, publish it to your team.)

The AI drafted it. You own it. Your team can see it. Three months from now when someone asks “how do we onboard a new client,” the answer is one place.

What this looks like for an AI agent (not just a human using AI)

The thing I’m watching closely is that the “hybrid” word is about to include agentic AI — not just a person chatting with Claude, but an AI agent handling an end-to-end task autonomously.

If you let an agent book meetings, triage support tickets, or close deals without a documented process, you’ve got the same drift problem a human team would have — except multiplied by the agent’s throughput. The agent will do the task 50 times before you notice it’s been doing it slightly wrong since day one.

The same fix applies, and it matters even more. Before you grant an AI agent access to a part of your business, write the SOP for it. The agent reads the same documentation your human team reads. If a human couldn’t do the job from the SOP, neither can the agent. If both can — you’ve just given your agent the institutional context that AI by itself doesn’t carry.

Our MCP server makes this concrete. The same whatstheprocessfor tool-set Claude gets in Claude Desktop is also what an agent would use to read your SOP library on behalf of a task. Same permissions, same audit trail, same place.

The actual commitment

I’m not pitching you a new tool to add to your stack. I’m pitching a discipline: the AI drafts, the human owns, the library remembers.

If you’ve got a tool that gives you the “library remembers” part — Notion, Confluence, SweetProcess, whatever — the discipline matters more than the brand. If you want the one we built, that works too. Pricing starts at $29/mo flat for your whole team. See the pricing or start a free trial — 30 days, no credit card.

The AI isn’t going to walk out the door. Your team eventually will. Decide now whose knowledge lasts.