The Knowledge That Walks Out the Door: An AI Summer Project

How AI finally makes it worth documenting your school's work, and why that matters more than any flashy tool

Every school has a problem it has quietly stopped trying to solve. This summer, I became convinced that AI can finally fix it.

I had just come back from the AeXIS annual conference for advancement leaders at California independent schools, and two themes kept surfacing in session after session. Neither one was flashy, and both were quietly expensive.

The first was dirty data. Reports nobody fully trusts. Totals that do not reconcile. A ten-minute task that somehow takes three hours.

The second was turnover. The person who "just knew how everything worked" leaves, and the knowledge walks out the door with them. The systems start to degrade almost immediately.

The more I listened, the clearer it became that these are not two problems. They are one. Both come from the same place: the rules live in people's heads, not on paper. How we code a gift. Why we do it that way. The exception nobody ever wrote down. When none of it is documented, the data drifts, and the institutional knowledge is one resignation away from being gone.

Why this never got fixed

We all know we should write things down. So why don't we?

It is not laziness, and it is not for lack of good intentions. The honest reason is economic. Until now, doing the task again was always cheaper than documenting it. The veteran can finish the job in ninety seconds, but explaining every nuance in writing takes an hour she does not have. So she does the task one more time, and the knowledge stays in her head. Multiply that by every small decision she makes in a week, and you can see why the written guide, what many teams now call a standard operating procedure, or SOP, never gets created.

This is the part that changes everything. AI inverts that cost structure. Documenting used to be slow, manual, and dreaded. Now the slow part is handled for you.

What AI actually changes

It helps to be precise about what is genuinely new here, because it is not "AI writes your manual for you while you sleep." It is two specific shifts, and both matter.

The first shift: you no longer have to write. Someone can talk through their work out loud, or simply let AI observe the task as they do it, and the tool produces a written draft. The capturing becomes natural and fast. The translating into clean text, the part people dread, is the part AI does well.

The second shift is the one most people miss. AI can draw out the knowledge an expert does not even realize they have. A veteran will never think to write down "and if the gift is also matched, I handle it differently," because to them it is obvious. But when AI asks the question a newer colleague would ask, that hidden rule comes out. The tool surfaces the hard-won judgment that would otherwise be impossible to capture.

The real distinction: steps versus judgment

If you take one idea away from this article, make it this one. There are two very different kinds of knowledge locked in people's heads, and they are not equally important.

The first is procedural knowledge. The steps. How to pull a particular report, where to click, what order to do things in. This is useful, and it is also the easy part. It is also the part most likely to change when your software updates.

The second is judgment knowledge. The why behind the steps. When the normal rule does not apply. How you decide between two reasonable options. Why your team chose to do it this way and not the other way. This is the hard part, it is the valuable part, and it is the part that actually leaves when a person does. A newer colleague can usually figure out the steps. They cannot reconstruct years of judgment on their own.

Most documentation efforts capture the steps and miss the judgment. The opportunity with AI is to finally capture both, and especially the second.

Five ways to capture what is in someone's head

There is no single right method. Think of these as five approaches, and pick the one that fits the knowledge you are trying to preserve.

1. Record the doing. Have someone do a task they know well while they narrate what they are doing and why. AI turns that recording into a written SOP. This is the lowest-effort starting point, and it works well for routine, repeatable work. One advancement leader I heard from suggested a one-month sprint where each person records their core processes once. By the end, the whole office is documented.

2. Interview the expert. Instead of asking someone to write, put them in a conversation and let AI ask the questions. "Walk me through how you decide which fund a gift belongs to." "What do you do when it is unclear?" "What about a gift from a family foundation?" The follow-up questions draw out the exceptions and the reasoning. A thirty-minute conversation with a long-tenured colleague can surface a dozen decision rules that would never have made it into a manual. This is the best method for judgment knowledge, and it has the nice effect of treating that person as the expert they are.

3. Capture the decision while you are making it. When your team finally settles a long-running debate, such as how to handle a certain kind of gift going forward, record that conversation. AI can turn the discussion into a clean summary that captures not just the rule you landed on, but the reasoning behind it. The reasoning is what helps the next person apply the rule to a situation you did not anticipate.

4. Let your existing data show you where the gaps are. This is where the two problems meet. Ask AI to study your existing records and point out where they are inconsistent: the fund that is spelled three different ways, the totals that do not add up, the codes that seem to mean different things in different years. Those messy spots are a map. Each one usually marks a place where a rule was never written down, so the data tells you exactly what needs documenting. The same review can work in reverse: where a person has already left, AI can study the patterns they followed and propose the conventions they were using, so you can reconstruct what was in their head from the trail they left behind.

5. Draft from scratch and edit. The blank page is the enemy. Hand AI your existing reports and notes and ask it to produce a first draft of your definitions, naming conventions, or coding rules. It will not get everything right, but your team will be editing instead of authoring, and that is the difference between a project that gets finished and one that never starts.

What this looks like in a school

Because I heard this at an advancement conference, let me make it concrete there first, and then widen it out.

An advancement office runs on knowledge that almost never gets written down. How do we credit a gift when it comes through a donor-advised fund, and what changes if a spouse is also involved, or if there is an employer match? Why are our campaigns and funds named the way they are, and why did the naming change a few years ago? Which family always gets invited to a particular event, and at what level, because of a relationship that exists only in the director's memory? How exactly does money get from a dozen different online portals into the system of record?

None of that is in the database. It is in someone's head, and in a few spreadsheets only that person fully understands.

Now notice that this is not really about advancement. Although AeXIS is an advancement conference, every department in a school has its own version. Admissions has its own logic for how inquiries get classified and moved through the funnel, and that logic often lives in one person's judgment. The business office has a year-end close that only the bookkeeper truly understands. The front office has a dozen quiet workarounds nobody ever wrote down. Pick any team in your school and you will find at least one process that exists in exactly one person's head.

This shares knowledge; it does not replace people

It is worth naming a worry directly, because plenty of capable people feel it. If I write down everything I know, am I making myself replaceable?

It helps to flip the picture. Being the only person who knows how to do something is not job security. It is a burden. You cannot take a real vacation without your phone going off. You get interrupted constantly. And all the risk sits on you, because if you are out, the work stops. Writing your judgment down does not shrink your value. It makes your expertise visible, which is usually the opposite of how hidden knowledge gets treated.

There is a second benefit that is easy to overlook. Once your knowledge is written in a form AI can read, it becomes something the whole team can ask questions of. A colleague can get an answer without interrupting you. A new hire can come up to speed without shadowing you for a month. A leader can understand how the work actually happens without anyone feeling audited. That kind of transparency is good for the whole office, and it turns one person's private expertise into a shared asset that keeps its value over time.

This is the right role for AI here. It is capturing and sharing what people know. It is not doing their thinking for them.

The half AI cannot do

I want to be honest about the limits, because this is where a lot of AI advice quietly oversells.

AI can document what exists in someone's head. It cannot invent knowledge that was never recorded anywhere. If the only record of who graduated in 1978 was on paper that is long gone, AI will not recover it. Worse, if you ask, it may confidently make something up. Treat it as an assistant that organizes what you already know, not an oracle that knows your history.

AI also cannot decide your conventions for you. Whether a gift should be credited one way or another is often a genuine judgment call with no single correct answer. AI can capture the rule you choose and apply it consistently, but the choice has to be yours. Letting the tool decide is how you end up with a confident, well-formatted procedure that happens to be wrong.

And every draft needs a human to check it. A polished document that looks authoritative but contains an error is more dangerous than no document at all, precisely because people trust it.

So set a realistic goal. You are not going to get a hundred percent of the knowledge out of people's heads. But getting most of it written down, and knowing which few things remain undocumented, is a world better than where most schools are today.

The one thing you must never let it see

There is a real privacy line, and it is simple to state.

The moment someone narrates their work while real student or donor information is on the screen, the AI tool is seeing that sensitive data. That may be fine inside a properly secured, school-approved environment. It is not fine in a personal, consumer account, where your data may be retained or used to improve the model.

The safe path is to use an enterprise or school-managed version of these tools, the kind your technology leader has vetted, and to keep the documentation focused on the process rather than on real people. When in doubt, practice on a test record or remove the names before you record. If you remember nothing else: never put confidential family or financial data into a personal AI account.

Where the tools stand, in mid-2026

Everything above is durable. The advice in this section is not, because these tools change every few weeks. Read this as the current shape of the choices, not as a recommendation that will age well.

There are two broad families of tools, and they are good at different things.

The first is general AI assistants, such as Claude, ChatGPT, and Google's Gemini. These are where the real intelligence lives. Because the whole interaction is a conversation, they are built to capture the why, the exceptions, and the reasoning, not just the steps. Many can also save their output directly into your documents or knowledge base.

The second is purpose-built capture tools, such as Scribe and Tango. It is worth being clear about what these are. They are workflow tools, designed specifically to document a clickable procedure. As you work, they automatically record your clicks and assemble a clean, step-by-step guide with a screenshot at every step. Some now add AI features to tidy up the writing, but their core job is mechanical capture, not understanding your judgment. They are excellent at the steps and not built for the wisdom.

A simple way to choose:

• If you want to capture... The judgment, the why, the exceptions, the reasoning. Reach for... A general AI assistant (Claude, ChatGPT, Gemini

• If you want to capture... A screenshot-perfect, click-by-click visual guide to a procedure. Reach for... A purpose-built workflow tool (Scribe, Tango)

A note on video, because it is genuinely useful. Gemini is unusually good at understanding a recorded video, including both what is on the screen and what you say out loud. I tested this myself. I recorded a simple screen capture with narration, uploaded it, and asked for a written procedure. It produced a solid draft, and it could even point me to the exact moment in the video where I explained a particular decision. One thing to know: out of the box, it would not save the finished document straight into my files the way an assistant like Claude can in a connected workspace. There are ways to extend it that may close that gap, but it was not the default experience.

One small tip that will not go out of date: narrate as you work. Say what you are doing and why, out loud, even when it feels obvious. The spoken explanation is what lets any of these tools capture your reasoning, not just your mouse movements.

Two things you can steal today

Here are two simple, durable tools you can use no matter which AI product you prefer.

First, a prompt that flips the work from writing to talking. Paste this in before you start, and let the AI lead:

"I want to document how I handle [name the process]. Do not just list the steps. Interview me like an analyst trying to capture my judgment. Ask me three questions at a time about how I decide, what the exceptions are, and how I stay consistent. After I answer, ask the next three, until you have enough to write a clear procedure."

The reason this works is that it is far easier to answer a pointed question while the work is fresh than to sit in front of a blank page trying to remember every unusual case from the last five years.

Second, a simple format that forces the judgment onto the page instead of burying it. Ask for the final SOP to include a table like this:

• The situation: A standard annual fund gift arrives

• What you do: Record it to the current year's fund

• Why you do it: Keeps the annual campaign clean and comparable year to year

• The exception to watch for: If it arrives in the last days of the fiscal year, confirm which year it belongs to before recording

That "why" column is the whole point. It is the part that usually lives only in someone's head, and it is the part that lets the next person handle a case you never thought to write down.

Why this is worth your summer

This is unglamorous work. It will never feel urgent, which is exactly why it keeps getting skipped.

But it is one of the rare projects that pays off twice. It pays off now, because it steadies your data, it spreads knowledge across your team, and it protects you from the quiet damage of turnover. The work can keep moving even when one person is out. And it pays off later, because documented procedures are exactly what you need in hand before AI can begin taking real work off your team's plate.

So if you are looking for one concrete, high-leverage thing to do this summer, this is the one I would choose. Start with a single process that lives mostly in one person's head. Invite that person to walk through it in their own words while you record. In half an hour, knowledge that used to exist in only one place becomes something your whole team can rely on.