Last October a 60-person SaaS company in Boston watched its Senior Ops Lead give two weeks notice. HR ran the standard math: replacement cost around $95,000 once you counted recruiting fees, hiring manager time, onboarding, and ramp. That number turned out to be wrong by a factor of three.
What the company did not know, and what nobody had written down, was that this one person knew Vendor X's invoicing system sent corrupted CSVs every November because of a timezone bug on their side. She knew the three accounts that needed a Friday afternoon check-in call because they would otherwise escalate over the weekend. She knew that the staging deploy always failed on the first push after a long weekend because of a cache that had to be cleared manually. None of it was in Confluence. All of it was in her head. Her replacement rebuilt that context over about seven months. In the interim, two of those three accounts churned, Vendor X's November bill went unreconciled for six weeks, and the ops team shipped a hotfix at 11pm on a Tuesday that would have been a 9am fix if the right person had still been on payroll.
That is the real cost of a senior departure, and it is the single largest unmanaged risk in most services and operations-heavy businesses.
The number everyone quotes and the number that actually matters
The headline number is well-documented. Gartner's 2023 research on employee turnover put the replacement cost of a mid-to-senior employee at 50%-200% of their annual salary, with specialized or senior roles landing near the top of that range. (Gartner Turnover Research, 2023) SHRM's long-running benchmark sits at roughly six to nine months of salary for a typical professional role, and climbs past a year of salary for senior technical or operational positions. (SHRM Human Capital Benchmarking Report) Work Institute's 2022 Retention Report pegged the average turnover cost at $15,000 per employee across all roles, which is the conservative floor, not the reality for someone who has been around long enough to know where the bodies are buried. (Work Institute Retention Report 2022)
Those numbers are real, and they are also the easy part. They cover recruiting, onboarding, and ramp. They do not cover what the departing person knew that nobody else did. That cost does not show up on an HR dashboard because nobody is tracking it, and nobody is tracking it because it is invisible until the day it is gone.
Replacing a senior operator costs 50%-200% of salary. The tacit knowledge they take with them is not on that ledger, and it is usually the larger number.
Two kinds of knowledge, and only one is in your wiki
The distinction dates back to Michael Polanyi in 1966, who argued that "we can know more than we can tell." Ikujiro Nonaka built a whole theory of corporate innovation on top of that insight in the 1990s: the knowledge that makes an organization work is split between explicit knowledge, which can be written down, and tacit knowledge, which is bound up in the person who holds it. (Nonaka, The Knowledge-Creating Company, HBR 1991)
Explicit knowledge is what ends up in a runbook. How to deploy. How to reset a customer's 2FA. The steps to close the books.
Tacit knowledge is everything else. Which customers need the soft escalation path. Why the sales team always books the Tuesday standup on Wednesday. Which engineer on the integrations team will go heads-down for a week if you open a Jira ticket the wrong way. How much slack to give Vendor X on their November invoice before it becomes a collections problem. When to ignore the alert and when to wake someone up.
A widely cited Deloitte survey found that only about 12% of the knowledge that drives day-to-day work is formally documented. (Deloitte Human Capital Trends, Knowledge Management) Harvard Business Review's analysis of the same problem is blunter: the gap between what an experienced employee knows and what their wiki captures is usually the entire reason they are valuable in the first place. (HBR, "Your Company Is Full of Good Decisions. Learn How to Export Them," 2020)
This is why the standard corporate response to a departure, "please write up your handoff doc," almost never works. You cannot ask someone to write down a thing they do not know they know.
Why Confluence, Notion, and SharePoint all fail at the same job
Most companies with over 50 employees already have a knowledge management tool. They also still lose context every time someone leaves. The reason is not the tool. The reason is that the tools were designed to store explicit knowledge a human typed in, and tacit knowledge never makes it into a text box.
Three failure modes show up again and again:
- The writing tax. Documenting a process costs the senior employee 30 to 60 minutes per process, and they have 40 of them. They have a day job. They never finish.
- Staleness. Even the docs that exist go stale within a quarter. McKinsey's 2012 research on the social economy, still the most-cited study on this, found knowledge workers spend 19% of their week searching for information and another chunk working from information that turns out to be wrong or outdated. (McKinsey, The Social Economy, 2012)
- The wrong unit of capture. A runbook captures steps. It does not capture the decision pattern behind the steps, which is the part the new person actually needs. A checklist that says "escalate to engineering if latency is above 500ms" is useless if the actual rule the senior operator uses is "escalate if latency is above 500ms AND the customer is on the enterprise tier AND it is within four hours of their peak window."
The wiki captures the what. The person holds the why.
What modern AI-assisted capture actually looks like
The reason this problem is suddenly tractable is that three capabilities that used to be expensive are now cheap:
- Transcription at near-human accuracy. A one-hour recorded interview produces a clean transcript in under a minute for under a dollar.
- Structured extraction. An LLM can read 40 pages of raw transcript and pull out decision rules, exception patterns, named accounts, named vendors, and escalation triggers into a structured format.
- Queryable memory. That structured output can live in a retrieval system the whole team queries in natural language, instead of a doc nobody opens.
Stitched together, the pattern is: sit with the person for an hour, record it, let the model do the writing, review and ratify, ship it. The senior employee's job is to talk, not to write. Their time cost drops from 20 hours of documentation over a month to one hour of conversation. The output is richer because people speak more honestly about their real decision rules than they write.
This is not theoretical. The early results are already on the board. A 2023 MIT/Stanford field study of a 5,000-agent customer support team found that giving agents AI assistance trained on the top performers' transcripts produced a 14% productivity gain on average and a 35% gain for the newest hires, largely by giving novices access to the tacit patterns of the veterans. (Brynjolfsson, Li, Raymond, "Generative AI at Work," NBER 2023) The mechanism in that study is the same mechanism at play here: capture the senior person's decision patterns, make them queryable, and the rest of the team gets more of what that senior person knew.
Your first experiment: 60 minutes with the most senior person most likely to leave
You do not need a platform, a vendor, or a rollout plan to start. You need one meeting. The experiment is small enough to run this week:
- Pick the target. The single person whose departure would hurt the most. Usually a senior ops lead, a tenured support manager, a principal engineer, or the one account manager who knows the top five accounts.
- Record a 60-minute structured interview. Not a generic "tell me about your job." Specific prompts: walk me through the last five times you escalated something; what are the three accounts I should never let slip; what is the one vendor relationship that breaks every quarter; what decisions do you make that a new hire would get wrong.
- Run the transcript through structured extraction. Pull out decision rules, named accounts, named vendors, escalation triggers, seasonal patterns, and known traps.
- Review it with them for 30 minutes. They will correct the AI's interpretation, add the two things they forgot in the interview, and flag the parts that are actually sensitive.
- Ship it into a place the team can query. Even a shared doc with good search is a step up from nothing. A retrieval system they can ask in natural language is the target state.
Total time cost: roughly 2 hours of the senior person's time and 3 hours of whoever is driving the capture. Output: a living artifact that covers most of what would otherwise walk out the door.
Run it once, on one person, and you will find out three things. How much context your most senior people actually hold. How much of it was genuinely not documented. And how different their real decision rules are from the ones written on the wall.
The real play: capture as a workflow, not a panic project
One-off capture is better than nothing, but the companies that get this right do not run it as a project at all. They build it into the cadence:
- Every senior hire does a capture session at the 90-day, 1-year, and 3-year mark, when the shape of what they know changes.
- Every major incident or escalation gets a 15-minute post-mortem that feeds the same knowledge base.
- Every quarterly business review includes an "what did you learn this quarter that is not written down anywhere" prompt.
The point is that the knowledge base stays fresh because it is fed by the work itself, not by a documentation sprint that happens the week someone gives notice. The week someone gives notice is already too late. The tacit knowledge degrades the moment they know they are leaving, and half of it is gone by their last day.
The companies that will out-compete on operational excellence in the next five years are not the ones with the most AI tools. They are the ones who have systematically captured what their best people know and made it queryable by everyone else. That is the actual unlock. The tools are finally cheap enough. The question is whether you start before your next senior departure or after.
If you are sitting at 20 to 200 employees and can name the two or three people whose exit would actually hurt, you already know who to start with. See how we run institutional knowledge capture for growing companies, or reach out directly and we will scope a first session.