The 2 AM Slack Message

You're half asleep when the message comes in. "Hey, wanted to let you know I accepted an offer. Last day is in two weeks."

It's your best support agent. The one who knew why that one enterprise customer's API integration broke every third Thursday. The one who'd built a mental map of every edge case in your billing system. The one new hires would ping before checking the docs because the docs didn't have half of what she knew.

Two weeks. That's what you've got to extract years of accumulated knowledge from someone who's already mentally checked out.

What Actually Leaves With Them

Most founders think losing an agent means losing headcount. Backfill the role, train someone new, move on. But headcount is the cheap part.

What you actually lose:

Undocumented workarounds go first. Every support team has them. The Shopify webhook that fires twice on certain plan types. The refund flow that breaks if you process it before the payment settles. The customer who insists on being called "Dr." and will escalate to your CEO if you don't. None of this is in your knowledge base. It lives in one person's head.

Then there's relationship context. Your best agents don't just resolve tickets. They remember that Company X had a rough onboarding three months ago and needs extra patience. They know which customers are evaluating competitors and should get white-glove treatment. CRMs capture some of this. Most of it evaporates.

Pattern recognition disappears too. A senior agent sees a spike in billing questions and thinks "we probably pushed a pricing page change." A new agent sees the same spike and panics. That instinct takes months to develop and zero seconds to lose.

And training capacity. Who was training your new hires? Probably the person who just quit. Now your remaining team is handling their own tickets AND onboarding the replacement AND covering the gap. Burnout compounds fast.

The Math Nobody Does

WaymakerOS published research showing knowledge silos cost organizations roughly $100,000 per 100 employees annually in lost productivity. That number covers duplicated effort, repeated mistakes, and time spent hunting for information that used to live in someone's brain.

For a five-person support team, scale that down and you're still looking at $5,000-$10,000 in productivity loss per departure. And that's the conservative estimate. It doesn't account for:

• The customer who churns because the new agent didn't know about their special arrangement
• The 6-7 months it takes a replacement to reach full productivity (industry data from TSIA backs this up)
• The tickets that get escalated unnecessarily because nobody knows the shortcut anymore

A support team with 30-45% annual turnover (the industry average per AmplifAI, with some segments hitting 60%) is rebuilding its institutional knowledge almost entirely every two to three years.

Why "Just Document Everything" Doesn't Work

You've tried this. Everyone has. The wiki initiative. The runbook project. The "let's record Loom videos of every process" phase that produced 47 videos nobody watches.

Documentation fails in support for a specific reason: the most valuable knowledge is situational. It's not "how to process a refund" (that's in the docs). It's "when customer X asks for a refund, check if they're on the legacy plan first because the legacy refund flow is different and they'll get a double credit if you use the standard process."

That kind of knowledge is hard to document because the person who has it doesn't even realize it's special. They just... do it. Ask them to write it down and they'll give you the standard process, not the 30 exceptions they've memorized.

What Actually Works

Classify and route, don't memorize

If your system can identify that a message is about billing (and specifically about refunds on legacy plans), it can route that ticket to the right person or surface the right instructions automatically. The knowledge lives in the routing rules and the intent taxonomy, not in one agent's head.

Supp's classifier handles this with 315 intents. When a message comes in about a refund, the system knows the difference between "refund request, standard" and "refund request, subscription, legacy plan" and can trigger different routing rules and auto-responses for each. That's institutional knowledge encoded into software instead of trapped in a human brain.

Record decisions, not just processes

When an agent makes a judgment call (waived a fee, made an exception, escalated differently than usual), capture why. A simple "decision log" field on tickets is worth more than a 50-page runbook. Over time, these decisions become your actual policy, not the idealized version in the wiki.

Pair new hires with the system, not just a mentor

A classification system that labels every ticket with its intent category teaches faster than a busy coworker. A new hire who sees tickets pre-tagged as "shipping_delay > international > customs_hold" immediately understands the pattern without requiring another human's time.

Make the first week about exceptions

Every new agent can learn the standard flows from docs. Spend their onboarding time on the weird stuff. The top 20 customers with special arrangements. The five integrations that break in non-obvious ways. The seasonal patterns in ticket volume. This is the knowledge that takes the longest to acquire organically.

The Real Insurance Policy

You can't prevent turnover. People leave for better pay, different roles, burnout, relocation, a hundred reasons you can't control.

What you can control is how much of your operation depends on any single person's memory. Every workaround that lives only in someone's head is a liability. Every edge case that only one agent knows about is a risk.

The goal is a support operation where someone can quit on Friday and the team handles Monday without skipping a beat. Not because people don't matter, but because the system is designed to capture what people know before they're gone.