How your documentation can do the work of 3 people

You’re writing help articles and SOPs for your team. What if that same documentation could answer customer questions, check work against your standards, and never forget anything?

I spent September building Ken’s help center and SOP library. Halfway through I realized something obvious. This documentation doesn’t just train humans. It trains AI.

Most founders don’t know their docs are sitting there ready to become working AI agents.

Here’s what happened when we tried it.

We were getting crushed by support requests. Growing fast meant more questions. Team was spending 25+ hours a month answering the same stuff over and over.

I had two options. Hire someone at $3k-5k per month or build a system. I chose the system. Spent all of September on it. Help center, SOPs, all the boring operational stuff everyone ignores until it breaks.

Then I found something I wasn’t looking for. Your documentation works twice. Once for humans, once for AI. We built everything for the team. Then we fed the same content to AI agents. Same documentation, double the value.

Now we have AI handling about 40% of support requests. We have a copy checker that knows our 100+ guidelines. We also have a knowledge base that never forgets. This wasn’t even the plan. I just wanted better docs for the team.

There are three phases. You probably already have phase one done.

You need detailed documentation first. Not because AI is picky. Because specific documentation creates specific AI.

What counts as documentation:

At Ken we have help center on Intercom at help.getken.ai, SOP library in ClickUp, and Guidelines for everything from copy to delivery.

The more detailed the better because Vague docs create vague AI.

This is easier than you think. We use ChatGPT Projects and Claude Projects. You upload documents and write a prompt. That’s it.

The process:

Takes about 30 minutes for your first agent.

Don’t try to do everything at once. Pick one repetitive task where you already have documentation. We started with three agents.

This is what we built. AI that answers customer questions based on our help articles.

Intercom has this feature called Fin. It’s an AI agent that reads your help center. Setup took 20 minutes. Turn on Fin in settings. Tell it which articles to reference. Set rules for when to hand off to humans and then test with customer questions. We cut support requests by about 40% in the first month doing this.

What actually happens is customer asks question, then Fin checks our help articles. If it knows the answer it responds immediately. If unsure it routes to team. When team answers something Fin missed, we add that to the help center so Fin has more context for next time.

Before the AI we got about 100 support requests per month. Each took around 15 minutes to answer. That’s 25 hours of team time every month.

After we set up Fin we dropped to 60 requests per month. Each one now takes 5 minutes because we give a quick answer and link to an article. That’s 5 hours of team time. We saved 20 hours per month for $65.

This one solved a problem I didn’t know how to fix. We have over 100 copywriting guidelines. Never use certain phrases, always bold key points, keep paragraphs under 4 sentences, use contractions, avoid corporate jargon, and so on.

Our copywriter is good. But nobody can remember 100 rules while writing. So we built an AI that knows all of them. Setup was simple. We created Claude Project, uploaded entire copywriting SOP and wrote prompt that says review this copy against our guidelines and flag anything that breaks a rule.

Now when our copywriter finishes an email they run it through the checker which takes around 30 seconds. It catches when a forbidden phrase is used, when a paragraph runs too long, when corporate jargon slips in, and so on.

This wasn’t possible six months ago. You couldn’t train AI on your specific rules and have it actually enforce them. Now you can.

Still building this one but it’s already useful. We trained AI on all our documentation. Team uses it to look up processes instead of searching through ClickUp or asking me.

Way faster than searching. And it stays up to date when we change the docs. It’s not perfect yet but it’s getting there.

If you’ve already documented anything, you’re 80% done. Help articles. SOPs. Process docs. Whatever you have. The other 20% is just feeding those docs to AI. Here’s what we use to do that.

This is help center plus AI agent in one. Setup takes 20 minutes. No code required. Worth it just for the help center alone. Fin makes it way more valuable.

Better than ChatGPT for long documents. Handles our 100+ rules better. You upload docs, write a prompt, and you’re done.

Record your voice and it transcribes and formats everything. Way faster than typing. I talk faster than I type so this saves me hours.

Time investment:

Week 1: Build help center and write first 10 articles

Week 2: Add onboarding and expectation docs

Week 3: Configure AI agent and test with team

Week 4: Train team on new workflow

After that it’s 2 hours per month to update articles when questions repeat.

Our setup costs $65 per month for Intercom with Finn. $20 per month for Claude Pro. One week to build. 2 hours per month ongoing.

Alternative is support person at $3,000 to $5,000 per month. Training takes 2 to 4 weeks. Management takes 5+ hours per month. Even if the AI only handled 20% of questions it pays for itself in the first week. And it works 24/7. No sick days. Same quality every time.

1. Don't Write Vague Documentation

AI needs specifics.

Bad: "Be professional in emails."

Good: "Use contractions. Keep paragraphs under 4 sentences. Bold 1 to 3 key insights per section. Never start with 'I hope this email finds you well.'"

The more specific your documentation, the better your AI works.

2. Don't Try To Automate Everything At Once

Pick one task. Get it working. Then add another. We started with support. Then added copy checking. Then knowledge base. Try to do everything at once and you'll quit before anything works.

3. Don't Let Your Documentation Get Outdated

Your AI is only as good as your docs. When the AI gets something wrong it's usually because the documentation is unclear or outdated. Fix the docs and the AI gets better automatically. It's that simple.

4. Don't Assume This Is Too Technical

If you can use ChatGPT you can do this. The process is upload documents, write prompt, test. The hard part is creating good documentation. But you should be doing that anyway for your team.

This isn't just for big companies. We're a team of 12 and we started this in September.

Do this if:

I think every agency will have AI agents handling 50%+ of routine work within 18 months. The ones that start now will have a 12 month head start. And this isn't even about replacing humans. It's about humans focusing on complex work while AI handles repetitive stuff. Document everything anyway. Then use it twice.

When we hire someone new they read the help center and SOPs. So do the AI agents. Same training for humans and machines. Is it perfect? No. AI still gets some stuff wrong. We update docs and it gets better. But it's already handling 40% of support. And we're just getting started.

Pick one repetitive task, then find or create documentation for it. Feed it to ChatGPT or Claude. Test it. Don't try to automate everything. Just one thing.

If you have a help center or SOP library sitting there doing nothing you're leaving money on the table. Make your documentation work twice.

P.S. The biggest mistake I see is waiting for the perfect time to start. There's no perfect time. But there's also never been a better opportunity than right now with AI. The early movers are going to dominate this space.

Where are you stuck with AI in your business? Documentation problems, not sure which processes to automate, tried it and it didn’t work?

Reply and let me know. I’m tracking what people are struggling with and I’ll dig into it in the next issue.

Cristian - Founder @Ken