When your best employee leaves — who owns what they knew

The person who knew which supplier picks up the phone at five on a Friday hands in their notice. What stays behind is a login, a folder, and a silence on the other end of the line. Most companies file this under HR, somewhere between the exit interview and the handover checklist.

Wrong drawer. That knowledge isn't soft. It's the most concrete asset your company owns, and it has just acquired a market price — because the first companies have started paying cash for it. Not to you.

Somebody pays $30 an hour to record how your people work

Shift, a product of the German company MicroAGI, pays people to strap a phone to their forehead and film their own manual work from a first-person view: changing a bike tube, loading a dishwasher, packing boxes, prepping food. The footage goes through review, faces and documents are blurred, and what remains is the movement itself — mapped joints and hands in space. That is fuel for training humanoid robots.

The scale isn't hobbyist. Shift's site{target="_blank"} claims 25,000+ workers across 15+ countries and over $5M paid out in the first quarter of 2026 alone. The platform{target="_blank"} is operated by MicroAGI, based in Aachen.

The rates are where it gets interesting:

Market	Rate	For what
US (global site)	$30/hr	one approved hour of footage
Poland, filming at home	30 złoty/hr	one approved hour of footage
Poland, filming at work	up to 50 złoty/hr	professional work, via the employer

Same number, different currency, roughly a quarter of the money.

Only approved hours pay. Breaks don't count as work. One tester filmed himself building Lego — rejected, because that's leisure, not productive labour. Washing dishes in heavy foam can fail too, once the algorithm loses the fingers under the bubbles. Historically about 70% of submitted footage clears the bar, so ten hours of recording is roughly seven hours of pay.

Why Poland

The clearest answer came from a Polish cleaning-company owner who is currently signing his crews up. He said it feels a bit like colonisers: someone with AI turns up in Poland, says "record yourself, I'll pay you", and is only here because labour is relatively cheap.

The other half of the same conversation runs differently. There are fewer people willing to do physical work every year, a repair job means months on a waiting list, and companies cannot staff their shifts. From that angle robots aren't a gadget — they're the only answer to a demographic problem anyone is actually proposing.

Both things are true at once. That same owner also worked out that he wins three ways: his crews earn a premium for wearing the mount, his clients pay less, and he takes a cut of the hours. Hard to hold that against him.

It is still worth being precise about what is being sold in that transaction.

Your company has the same problem, minus the cheque

Robots can't learn from YouTube. Footage scraped from the web has no depth data and no guarantee that it shows real work in real time — which is why Shift requires specific phone models with depth sensors and live capture. The largest public first-person video dataset, Ego4D{target="_blank"}, holds 3,670 hours. Shift's own people say teaching a robot general manual tasks needs hundreds of thousands of hours.

Hence the cheque. No record exists of what competent hands actually do, so it has to be bought from the people doing it.

This is where it stops being a story about robots. Your company stands on exactly the same kind of knowledge — unwritten, unstructured, living in hands and heads. How your best technician identifies a fault by sound. How your dispatcher knows that load won't fit on that truck. Which supplier answers on a Friday afternoon.

One difference: Shift pays for it. You lose it for free, with every resignation, every sick note, every holiday in peak season.

What you see the first time you open a company's files

In August we received two spreadsheets from a client — a sales register and a purchase register. The company remanufactures components and sells into four countries. It was the first moment we could see how they actually work, rather than how they describe their work. The gap between those two is usually wider than the owner suspects.

Three findings from that analysis repeat at almost every company we walk into.

The fragmentation is built into the tool. Sales split into one sheet per month, purchases into one sheet per supplier. Answering "what did we buy from this supplier this year" means opening a dozen tabs and adding it up by hand. That isn't sloppiness — it's the shape of the file they've worked in for years.

The company's most valuable dictionary lives in no system at all. A helper tab held their own taxonomy of parts: 36 categories named their way, twenty supplier types, six channels an order can arrive through. Their ERP has never heard of any of it. That is pure operational knowledge, written down by accident, in a side column.

The process stops halfway and nobody sees it. The company pays suppliers a deposit on cores and recovers it when the core goes back. In the sample we received, six deposits were entered as non-zero, totalling 2,500 złoty, and the number of settled returns was zero. It might be a posting delay — we don't accuse anyone on the strength of two spreadsheets. But the question itself is a test:

How many open deposits do you have right now, and for how much?

If the answer takes longer than thirty seconds, that isn't an accounting problem. It's knowledge with nowhere to live.

Three places knowledge leaks from

Where it sits	How you spot it	What you lose when the person goes
In one person's head	"Ask Mark" gets said several times a day	Decisions nobody can reconstruct
In a spreadsheet on someone's drive	Reports get built by copying between files	History and context; only numbers survive
In an inbox	What you agreed with a client lives in a thread	Commitments you hear about from the client
In a messaging group	A photo of the part instead of a ticket	The whole history, the moment someone leaves the group

None of these is wrong in itself. The problem starts when it's the only place.

Why "let's write documentation" doesn't work

Because you've tried twice and you know how it ends. Somebody sits down and describes the process from memory, the description lands on a shared drive, three months later it's out of date, six months later nobody opens it. A description of work is not the work.

Shift solved this the only way that works, and the lesson is worth stealing. They never ask anyone to describe how a tube gets changed. They record the tube being changed, at the moment it was going to happen anyway. The record is created where the work happens, not in a workshop with a flipchart.

That principle transfers to a business one-for-one. Operational knowledge only gets captured when it captures itself — as a by-product of something the employee is doing regardless. An order taken over the phone lands in the system together with its transcript and what was agreed. A deposit logged when the core arrives opens an obligation that chases itself. An enquiry from an inbox reaches the register without being retyped. That is what process automation actually means in practice — not a robot replacing a person, but a system that stops losing what the person already did.

It helps to think of this as the operating layer of the company rather than another tool. We went through that logic in more detail when writing about an AI operating system for a business, and the punchline there is the same: capture the process first, automate it second.

Want to see where yours leaks? Book a 30-min call — we'll walk one process end to end.

Where to start this week

No budget, no implementation, about an hour.

1. Ask three questions with a stopwatch running. How many open obligations do we have with suppliers? Where did the last ten orders come from? What did we agree with that client three months ago? Any answer over thirty seconds marks a process with no memory.
2. Count how many times a day somebody says "ask Mark". It is the cheapest risk metric available. Mark is simultaneously your best employee and your most serious single point of failure.
3. Pick one process and find where its record gets created. If the record is written after the fact, from memory, in the evening — that's the one to fix first.

Don't start with the most important process. Start with the one that most often requires asking a human.

What this story does not say

Honestly, because this topic is easy to oversell.

Robot manufacturers say humanoids will enter American homes in 2027. That's a vendor claim, not a fact, and it deserves the same treatment as any other vendor claim. Shift's people expect a "ChatGPT moment" for home robotics within a year — their forecast, not data.

The earnings deserve proportion too. Testers describe about half an hour of continuous wear before the forehead mount needs to come off. Realistically it comes to twenty or thirty hours a month. A supplement, not a profession.

And go carefully with the phrase "knowledge loss" itself. The goal isn't to trap people in procedures, or to prepare the company for replacing them with machines. The goal is that the company remembers what it learned — including on the days its best person is on holiday. If you're wondering where that knowledge would physically sit, there are now models that run locally, without sending data to the cloud.

Who digitises your company

Your company's operational knowledge is going to get recorded. That is already happening; the only open question is whose system benefits.

One version: somebody pays your people thirty złoty an hour and takes the record of how they work. They get a top-up on their wages, you get nothing, and the knowledge leaves the building in a format you cannot access.

The other version: the same record is created on your side, in your system, as a by-product of work that was happening anyway. Then a resignation is a staffing problem rather than an operational one.

The difference isn't technology. It's who treats that knowledge as an asset first.

Book a 30-min call. We'll take one process and show you where the knowledge leaks — and what can be done without rebuilding the company. If you'd rather read first, start with what AI for business looks like in practice, or with our work on voice agents that answer the phone.

Sources

• joinshift.org{target="_blank"} — rates, worker count, Q1 2026 payouts
• go-shift.app{target="_blank"} — platform operator, MicroAGI
• Ego4D{target="_blank"} — 3,670 hours, the largest public egocentric video dataset
• Video coverage of Shift's launch in Poland — 30 zł/hr and up to 50 zł/hr rates, acceptance rate, statements from the team and their partner
• Analysis of a Prospere AI client's operational spreadsheets, August 2026