Embed the Expert

Problem

Expert judgment scales one conversation at a time

Every organization has a few people whose judgment others seek out. The senior PM who can read a project plan and say exactly where it will slip. The architect who spots the integration that will fail. The finance partner who knows which business case assumption won't survive the steering committee. Their judgment is among the most valuable things an organization has, and it reaches only the people who can get time on their calendar.

So it gets shared the way it always has: informally, when they're available. A hallway question. A review squeezed between meetings. A junior colleague who happens to sit nearby. Everyone else waits, or gets a weaker version from whoever was free. The quality of a project's review depends on who was available that week.

When I was at Planview, our CTO, Louise Allen, wrote articles for CIO magazine that showed exactly how she thought - and they were amazing. They reached far more people than her calendar ever could. But an article explains a way of thinking; it can't apply it to your project on a Tuesday afternoon.

The philosopher Michael Polanyi observed that we know more than we can tell. Much of an expert's judgment is that kind of knowledge, and no one can write it all down. But a good part of it can be drawn out - the questions they always ask, the red flags they always check - and most of that never gets captured either.

AI changes what's possible. The part of an expert's approach that can be drawn out - how they review a charter, pressure-test a business case, or diagnose a troubled project - can now be written down as a skill: a reusable set of instructions an AI follows to apply that approach on demand. Most organizations haven't done it. They use AI to produce more output, and leave their best judgment where it was: reaching only the people who can get a meeting.

Value

The best judgment in the building, reaching every project

Every project gets a first pass built on the expert's approach. Reviews that only happened when the expert was free now happen every time, consistently. The skill's findings are a first pass, not the expert's approval.

Less work waiting in line. Projects stop stalling for one person's review. Measure it: days work spends waiting for the expert's review, before and after.

The expert's time goes further. Routine reviews run through the skill; the expert spends their time on the hard cases the skill flags.

Judgment outlasts the role. More than 4.1 million Americans turn 65 every year through 2027. Not all of them will retire, but many of the senior practitioners organizations depend on are nearing retirement. A skill keeps their approach in a form others can run and improve - as long as it has a named owner and a regular refresh against the expert's current thinking.

New practitioners learn the expert's reasoning. A skill that explains why it flags something teaches as it reviews (see Pattern 09).

Context

When this pattern applies

This pattern applies when one person's judgment is a bottleneck: work waits for their review, quality varies by who was available, or their departure would leave a visible gap. It fits best when the judgment follows an approach that can be drawn out - questions they always ask, red flags they always check, a sequence they always follow - even if they've never written it down.

Start with the review your work waits on longest. That's where the waiting measure will move first, and it gives you a clear reason to choose one expert and one judgment over another.

It does not replace the expert. A skill captures the part of an expert's approach that can be drawn out, not everything they know. Novel situations, political judgment, and decisions with real consequence still need the person. The skill handles the first pass; the expert handles what the first pass surfaces.

It is the deliberate mirror of Pattern 12, Ghost in the Workflow. Ghost in the Workflow is what happens when knowledge gets trapped in AI by accident. Embed the Expert captures it on purpose, documented from the start.

Solution

Ask the expert, interview them, write the skill, and keep it current with them

Step 1

Ask first, and agree on credit and ownership

The skill packages someone's judgment, so start with them. Settle three things before anything else: they agree to it, they're named as the source, and it's clear who owns and maintains the skill once it's in use. An expert who helped build the skill will keep improving it. One who found out later won't.

Step 2

Pick one judgment, not a whole job

"How Maria reviews a project charter" is a skill. "Everything Maria knows about project management" is not. Choose a review or decision that happens often and that people already wait for.

Step 3

Interview the expert

Use Pattern 05's interview approach: have AI ask the expert hard, specific questions about how they do this one thing - what they look at first, what makes them worried, what they've seen go wrong, what they ignore. Capture their answers in their own words. You won't get all of it. Get what you can.

Step 4

Write the skill

Turn the answers into instructions an AI can follow: the questions to ask, the checks to run, the red flags to name, and when to say "this needs a human." Model it on Pattern 08, Assemble the Team, which is itself a pattern packaged as a skill. Keep it tool-neutral: a skill is saved instructions, and works in any AI tool that lets you save and reuse them.

Step 5

Test it against the expert

Run the skill on past work the expert already reviewed. Compare its findings to theirs. Where it misses something, ask the expert why they caught it, and add that to the skill. Repeat until the expert agrees it catches what they would catch first - not until they'd sign their name to every run.

Step 6

Give it a handoff record

Apply Pattern 12, Ghost in the Workflow: record whose judgment it captures, what it's for, what it was tested against, its known gaps, who owns it now, and when it was last refreshed. A skill without a record becomes the next ghost.

Step 7

Keep it current with the expert

On a regular schedule - quarterly is a reasonable start - and whenever their approach changes:

Review the advice the skill gavePull a sample of the skill's recent reviews and walk through them with the expert. Where would they have said something different?
Interview them againAsk what they've learned, what they now watch for, and what they've stopped worrying about.
Update the skill and the recordChange the instructions, rerun the Step 5 test, and log the date and what changed in the handoff record.

The handoff record is what makes this step work: it keeps the skill's advice, the expert's corrections, and each refresh in one place, so the next review starts from what changed instead of from scratch. If the expert has moved on, the record is how the new owner knows what the skill was built to catch.

Notes

This library is built this way

The "Use this pattern" instructions on this site are each a small Embed the Expert: one way of working, packaged so anyone can run it in any AI tool. Assemble the Team is the working example.

Related. Pattern 12, Ghost in the Workflow, is the mirror image: knowledge trapped in AI by accident instead of captured on purpose. Pattern 09, The Compressed Apprenticeship, is the companion: 09 helps people learn faster with AI; this pattern turns one person's judgment into something others can run. Pattern 05 supplies the interview.

Inspired by Michael Polanyi His idea of tacit knowledge - that we know more than we can tell (The Tacit Dimension, 1966) - frames what a skill can and can't capture.