Your clients have a key-person problem. Claude + Obsidian fixes it. $2,400 setup. $9,600/month.

Cover image

Every business has one person who knows everything. Where the real client contract is filed. Why the invoicing process has that one weird exception. What happened the last time this vendor caused problems. Which client actually can't stand the phrase "circling back."

None of it is written down. All of it lives in one person's head.

When that person quits, retires, or just takes a two-week vacation — the business runs on guesswork until they're back.

Fortune 500 companies pay consultants six figures to solve this. It's called key-person risk, and it's one of the most expensive problems in business nobody budgets for.

Small businesses can't afford that consultant. They can afford $2,400.

Here's the entire system — what it is, why most versions of this fail within 60 days, and the one architectural difference that doesn't.

What most people missed

Every company already has a place where knowledge is supposed to live. A wiki. A Drive folder. A Notion page someone organized once in 2023.

The problem was never storage. It's retrieval.

What every business already has:
 
 Google Drive:     files nobody can find
  Notion:            pages nobody updates after month one
  Slack history:     searchable, but nobody searches 40,000 messages
  The one person:    knows all of it, answers questions all day,
                      is a single point of failure

What a queryable vault changes:
  
New hire asks a question    → gets a specific, sourced answer in seconds
  The one person goes on PTO  → nothing stops
  The one person quits        → the knowledge stays
  Any employee, any time      → same answer quality as asking a 5-year veteran

The tool was never the issue. Every company has a tool. What they don't have is a tool that talks back

Part 1. Why most "AI chatbot on your docs" services fail in 60 days

This is the part nobody selling this service explains — because most of them haven't solved it.

The generic pitch is simple: connect an AI to a folder of company documents, let employees ask questions. It sounds like a five-minute build. It usually is.

That's exactly why it fails.

The generic version:
  Dump all files into a vector database
  AI answers based on whatever it retrieves
  No verification step
  No source citation
  No mechanism to flag outdated content
  Result: confident-sounding wrong answers within week 2

What breaks trust fastest:
  Employee asks a real question
  AI answers with something that sounds right
  Answer is based on a process that changed 8 months ago
  Employee acts on it
  Employee stops trusting the tool
  Client cancels within 60 days

A knowledge system that occasionally hallucinates is worse than no system at all — because people act on confident wrong answers faster than they'd act on "I don't know."

The fix isn't a better AI model. It's a verification architecture most people never build.

Part 2. The architecture that actually holds up — loop, not lookup

Every answer this system gives passes through a hard rule before it reaches an employee. Not a vibe check. A rule.

The shape of every query, always the same four steps:

TRIGGER   →  employee asks a question in Slack or Telegram

RETRIEVE  →  system pulls the specific vault files relevant to it

VERIFY    →  checked against a hard rule: does the retrieved content
              actually contain a direct answer, with a source and date?

RESPOND   →  answer + source file + last-updated date,
              OR "not documented yet" if verification fails

Skip VERIFY and the system guesses when it doesn't know — the exact failure mode that kills client trust. Skip the source citation and employees can't check the system's work, so they stop trusting it the first time it's wrong.

This one architectural choice is the entire moat. Anyone can connect an AI to a folder in an afternoon. Almost nobody builds the discipline to make it say "I don't have that documented" instead of guessing.

Part 3. The stack — three pieces, verified pricing

Nothing exotic. The cost structure is what makes the margin work.

┌────────────────────┬──────────────────────────────────────────────┐
│ Component          │ Role                                         │
├────────────────────┼──────────────────────────────────────────────┤
│ Obsidian           │ The vault. Free, local, plain markdown files │
│ Claude Sonnet 5    │ Reads context, writes verified answers       │
│ Claude Haiku 4.5   │ Routes questions to the right vault files    │
│ n8n +Telegram/Slack│ The interface employees actually use         │
└────────────────────┴──────────────────────────────────────────────┘

Model routing matters for margin. Haiku 4.5 ($1 input / $5 output per million tokens) handles the cheap job — figuring out which 2-3 vault files are relevant to a question. Sonnet 5 ($2 input / $10 output per million tokens, introductory pricing through August 2026) handles the job that needs judgment — reading those files and writing a grounded, cited answer.

# n8n HTTP Request node — the answer step
POST https://api.anthropic.com/v1/messages
{
  "model": "claude-sonnet-5",
  "max_tokens": 600,
  "system": "{{ $json.verified_system_prompt }}",
  "messages": [{
    "role": "user",
    "content": "{{ $json.employee_question }}"
  }]
}

Real monthly API cost for a typical small business (10-30 employees, roughly 150-250 questions/month, plus the weekly self-audit pass): $5–$15/month, with prompt caching cutting repeated vault-context costs further on top of that. That's the number to quote clients when they ask what running this actually costs — and it's the number that makes a $1,200/month retainer look like the bargain it is.

Part 4. Building the vault — the part that's actually the product

The vault structure is the product. A vague vault produces vague answers nobody trusts.

Prompt to generate the initial structure for any client:

You are setting up an Obsidian knowledge vault for a
[business type] with [number] employees, connected to
Claude for employee Q&A.

Generate a complete folder structure and list of core
files. For each file specify: file name, exact content
it should contain, who should have access, and how
often it needs updating.

Business context:
- Industry: [describe]
- Team size: [number]
- Departments: [list]
- Processes that generate the most repeat questions: [list]

Prioritize files that answer what new hires ask in
their first 30 days. Output as a structured list.

Every file needs a hard rule baked into how it's written — specific names, specific dates, no vague statements. "Ask [person] if unclear" is not documentation. "Client X requires PO numbers on all invoices, confirmed with finance March 2026" is.

Part 5. The feature competitors don't build — a vault that flags its own decay

This is the second half of the architecture, and it's what turns a one-time setup into a service worth paying for monthly.

Most versions of this product are static: build it once, answer questions, done. That's a project, not a business — clients cancel once the docs go stale and nobody notices.

Add a self-audit pass on a schedule:

TRIGGER:   every Monday, before your weekly review

STEPS:
  1. Flag any vault file untouched for 90+ days
     that gets referenced in employee questions weekly
  2. Compare unanswered questions from the past week —
     find the pattern in what the vault is missing
  3. Draft a one-line summary of what needs updating
     and send it to your Monday review queue

VERIFY:    every flagged item includes the specific
           question that exposed the gap
STOP:      report sent, or nothing to flag this week

This is the retainer, made visible. Instead of "trust me, I'm maintaining it," the client sees a weekly note: "3 questions this week had no answer in the vault. Here's what's missing." That's the difference between a service that gets cancelled at month three and one that gets renewed for years — the client watches it actively getting better, not quietly going stale.

Part 6. Pricing, packaging, and who actually pays fast

Setup fee funds the build. Retainer funds the maintenance loop. Both matter.

Setup (one-time, delivered in 3 days):
  Day 1 — Vault architecture + core file list
  Day 2 — 10-15 core files written, Claude integration live
  Day 3 — Team onboarded, tested, self-audit loop scheduled
  Price: $2,400

Retainer (monthly, ongoing):
  Weekly self-audit report reviewed (~30 min)
  Gaps filled based on what the audit flags
  Monthly usage + cost report sent to client
  Price: $1,200/month

Who signs fastest, without negotiating:

> Law firms. Every associate asks the same questions about billing codes and matter procedures. One hour of a partner's time saved per week justifies the retainer on its own.

> Agencies. Onboarding a new hire eats real money in senior time. A vault that answers process questions instantly cuts that cost immediately — any agency hiring more than one person a year feels this.

> Medical and healthcare practices. Protocols and insurance procedures documented once, queried instantly. Compliance-sensitive data means this runs on a local model instead of the cloud API — the pricing structure holds either way.

Part 7. Getting the first client — the demo does the selling

Nobody buys "AI on documents" from a pitch. They buy it from watching it answer a real question about their own business.

Build a demo vault using nothing but public information — a prospect's website, LinkedIn, a few minutes of searching. Three to five files is enough: a team overview, a service list, one or two common client scenarios. Two to three hours of work, done before you ever get on a call.

The meeting has one job: let them ask it something.

Open with a question, not a pitch:
  "What do new hires on your team spend their first
   month asking about, over and over?"

Let them talk. Then show the vault answering exactly
that — built from nothing but what's already public
about their business.

Close:
  "Setup is three days. $2,400. Retainer is $1,200/month.
   Want to start this week?"

The first demo in any industry takes the full two to three hours. The second one in that same industry takes forty minutes — most of the structure already exists, only the specifics change. Every client you close narrows the setup time for the next one just like it.

The numbers at scale

Month 1:    1 client, first setup delivered        →  $3,600
Month 3:    3 clients, 1 new + 3 retainers          →  $6,000
Month 6:    6 clients, 1 new + 6 retainers          →  $9,600
Month 9:    9 clients, 1 new + 9 retainers          →  $13,200
Month 12:   12 clients, referral-driven by now      →  $16,800

Active time at 12 clients:
  Weekly audit review: 12 × 30 min           = 6 hrs/week
  New client setup:    3 days when onboarding
  Total at steady state:                     ~8-10 hrs/week

At $16,800/month and roughly 9 hours of weekly work, that's over $400 an hour of effective time. The leverage isn't the setup fee — it's a retainer clients renew because they can see the system getting better every week instead of quietly rotting.

Mistakes that kill 90% of attempts

**1. Skipping the verify step.**An AI that guesses confidently is worse than no AI. One wrong answer ends the contract.

**2. Selling "a chatbot" instead of "insurance against key-person risk."**Nobody budgets for a chatbot. Sell the fear you're removing, not the tech removing it.

**3. Building a static vault with no self-audit.**Docs go stale, nobody notices, clients cancel by month three. The weekly decay report is what justifies the retainer.

**4. Routing every question through the expensive model.**Haiku 4.5 does file-routing for a fifth of Sonnet's cost. Match the model to the step, not the whole pipeline.

5. Writing vague vault content."Ask Sarah if unclear" produces a vague answer forever. Specific names and dates produce specific answers.

**6. Pricing setup without a retainer attached.**A one-time build is a project with an end date. A retainer is a business.

Every business already has the knowledge. It's just trapped in formats that don't talk back.

The product was never "AI on your documents" — any freelancer can build that, and most get cancelled in two months because they guess instead of admitting what they don't know.

The actual product is a system disciplined enough to say "I don't know yet" and go find the gap first. That discipline is the entire business.

The person who knows everything is still going to leave. The only question is whether the business notices.

**👉 Follow for weekly AI business systems

🩷Like if you want more breakdowns like this

🔒Save this — the system is worth stealing**