Photo via Unsplash
The last three people who joined us got a notebook before they got a laptop bag. On day one, before any introductions, each of them was asked to write down every question they had about the firm. Then they spent the afternoon asking those questions of a NotebookLM notebook rather than of a colleague, and building a personal reference document out of the answers, each one carrying the citation for the document it came from.
Time to first meaningful deliverable across those three hires averaged 9 days. Our prior average was 42. All three said they felt up to speed by the end of week two, where six weeks had been the norm.
What was in the notebook
The upload list matters more than the tool, so here it is in full. The last four quarters of internal memos. The two most recent board packs. All product documentation. The full org chart, with role descriptions rather than just names and titles. And six months of All Hands recordings, transcribed.
The org chart with role descriptions did more work than I expected. A large share of new-hire questions turn out to be routing questions rather than substantive ones. Not "what is our pricing strategy" but "who decides pricing, and who has to agree before it changes." Those are answerable from documents and they are the questions people are most reluctant to ask twice.
The two kinds of knowledge, and only one of them was ever written down
Onboarding a knowledge worker at a mid-sized firm has taken somewhere around six weeks for as long as I have been hiring, and the reason splits cleanly in two.
The first part is procedural. How meetings run, which channel is used for what, who signs off on what. This can be documented and in most firms it partly is, sitting in a handbook nobody reads because it is organized by topic rather than by the question the reader actually has.
The second part is contextual. Why the strategy changed in Q2 2024. What happened with the client everyone goes quiet about. What the founder means when she says "Aunt Millie's rule." That layer is almost never documented, and the traditional way to acquire it is to bump into it over six weeks of coffee conversations and half-overheard Slack threads.
What surprised me is that the contextual layer was largely recoverable. Not from a handbook, which nobody had written, but from the memos, board packs and All Hands recordings where those decisions were originally explained to people at the time. The context was there. It was just distributed across forty documents in a form no new hire would ever think to search.
The Experiment, and Its Limits
- 3 hiresTotal sample. No control group, and the participants knew they were in an experiment
- 42 to 9 daysAverage time to first meaningful deliverable, before and after
- 2 weeksWhen all three reported feeling up to speed, against roughly six weeks previously
- 5 source typesMemos, board packs, product documentation, annotated org chart, transcribed All Hands
- $0Cost. The free tier covers this use, and there is no deployment step
What three people does and does not prove
Three hires at one firm, with no control group, is an anecdote. The people involved knew they were part of an experiment, which is the condition under which people work harder. Our prior 42-day average is drawn from a small number of previous hires in different roles. Any one of those things would be enough to disqualify this as evidence in a paper. The reason I keep doing it is that the direction was consistent across all three and the cost of being wrong is close to zero, which is a different standard from the one a study has to meet, and the right standard for an operating decision at this size.
What I would not do is present this to a board as a productivity finding. What I would present is the method and an invitation to run it on their own next hire, which is a claim I can stand behind.
How to run it on your next hire
Give the notebook on day one with the five source types above. Ask for the question list before anybody has a chance to answer them informally, because once the questions have been asked out loud you have lost the measurement and most of the point.
Then sit down at the end of week one and go through what they learned. This meeting gets skipped, and it matters for two reasons. It catches the confident wrong answers, and there will be some: the notebook answers from whatever it was given, so a superseded memo produces a wrong answer with a citation attached, which is considerably more persuasive than a superseded memo sitting in a folder. And it tells you where your firm's undocumented knowledge actually lives, because the questions the notebook could not answer are a precise map of it. I learned more from that list than the hires did.
Beyond onboarding
Six weeks of onboarding was never six weeks of learning. It was several weeks of asking basic questions of colleagues who were trying to do their own jobs, and the cost of that was carried by the people being interrupted rather than by the person onboarding, which is why it never appeared in any measurement. NotebookLM did not make anyone smarter. It removed a coordination cost that had been invisible because it was distributed.
That pattern extends past onboarding. Anywhere in your firm where one person interrupts another to retrieve context rather than to get a judgement is a candidate for the same treatment. Client history before a renewal call. Prior decisions before a design review. Precedent before a pricing conversation. The test is whether the interruption is a lookup or an opinion, and a lot of what gets called knowledge work is lookup.
Where this breaks is when someone applies it to the opinion half. A document-grounded tool will answer a judgement question with the same confidence it answers a factual one, and there is no citation that can save you there. Keep it on lookups.
NotebookLM lives at notebooklm.google.com. The free tier covers everything described here. Run it on the next person you hire and pay attention to the questions it cannot answer.
"The interesting number was not 9 days. It was the list of questions the notebook could not answer, which is the first honest map of our tribal knowledge anyone here has ever produced." - Adrian Dunkley, AI Boss
Frequently Asked Questions
What is NotebookLM and what does it cost?
NotebookLM is a Google tool that lets you upload a set of documents and then ask questions about them in plain language, with every answer citing the document it came from. It runs in a browser with no IT deployment and no procurement cycle. The free tier is sufficient for the onboarding use described here, which is why it is a reasonable first AI deployment for a firm that has never run one.
What exactly was uploaded to the onboarding notebook?
The last four quarters of internal memos, the two most recent board packs, all product documentation, the full org chart with role descriptions rather than just names and titles, and six months of All Hands recordings run through transcription. The org chart with role descriptions did more work than expected, because most new-hire questions turn out to be routing questions about who decides what rather than questions about the business itself.
What was the new hire actually asked to do?
On day one, before speaking to colleagues, write down every question they had about the firm. Then spend the first afternoon putting those questions to the notebook rather than to people, and build a personal reference document of the answers with the source citation attached to each one. At the end of week one, sit down and review what they learned and what the notebook could not answer.
What was the measured result?
Across three hires, time to first meaningful deliverable moved from an average of 42 days to an average of 9. All three reported feeling up to speed by the end of week two rather than at around six weeks. This is a sample of three at one firm, with no control group and with hires who knew they were part of an experiment, so it is an anecdote with numbers attached rather than evidence. The direction was consistent enough to keep doing it and nowhere near strong enough to generalize from.
Why does onboarding traditionally take around six weeks?
Two different kinds of knowledge are being acquired. Procedural knowledge is how meetings run, which channel is used for what, and who signs off on which decision, and it can be documented. Contextual knowledge is the history behind current decisions, why a strategy changed in a particular quarter, and what a founder means by an internal shorthand phrase, and it usually is not documented anywhere. New hires historically acquired the second kind by bumping into it in conversations over several weeks.
Does this replace human interaction during onboarding?
No, and treating it that way is the main way this fails. What it replaced was the several weeks of low-value question asking that used to come before real work, not the relationships. Every hire in this experiment still had the same introductions, the same one-to-ones and the same team time. They arrived at those conversations able to ask a second-order question instead of a first-order one.
What is the risk of running onboarding this way?
The notebook answers confidently from whatever it was given, so an outdated memo produces a confident wrong answer with a citation attached, which is more persuasive than an outdated memo sitting in a folder. The mitigation is the week-one review conversation, where a person checks what the new hire concluded. Skipping that review turns a good onboarding tool into an efficient way to propagate stale internal information.
Where else does this pattern apply inside a firm?
Any process where one person interrupts another person to retrieve context rather than to get a judgement. Client history before a renewal call, prior decisions before a design review, precedent before a pricing conversation. The test is whether the interruption is asking for a lookup or asking for an opinion. Lookups are what document-grounded tools handle well. Opinions are not, and trying to automate those is where these deployments go wrong.