AI for knowledge transfer: what it can carry and what it cannot

The line is not between important and unimportant knowledge. It runs between what once existed as text and what never left a pair of hands.

The question underneath the search is usually specific: can a machine take over the part of a handover that depends on somebody remembering. The answer is yes for a larger fraction than most people expect and no for a part no amount of tooling reaches, and most unhappy purchases here come from a buyer who never separated the two.

What an AI tool for knowledge transfer can carry

Everything that ever passed through text, which in an operating role is most of the material that matters.

• Correspondence. Years of mail with vendors, partners and regulators, including the small clarifying replies where the actual agreement usually lives.
• Documents and what came attached to them. The contract, the report, the quote, and the thread that carried each one in.
• Meetings that were captured. A transcript or a set of notes is text, and a decision made verbally survives if it was recorded somewhere at the time.
• The record built out of those. Commitments with dates, obligations with trigger events, decisions with the reason that was given for them.

The genuinely useful move is not summarising any of that. It is that a question which previously required a person becomes a lookup: why the second vendor was chosen, what was agreed about the extension, what is still open on a thread that has been quiet for two weeks. A machine is also good at the one job a departing human cannot do: reading back months of correspondence and rebuilding the state from it, without deciding halfway through that some of it is not worth mentioning. Practically the record is kept current by a scheduled pass, hourly at best, on a machine that is awake, and it runs against a single person's account, so what it carries is that person's stream rather than the company's. How that record gets written day to day is institutional knowledge capture, and what it becomes over a year is AI institutional memory.

What it cannot carry

Four things, and none of them are fixable by a better model.

• Skill that lives in hands. How somebody actually runs the technique, sets up the instrument, or steers a difficult call while it is happening.
• Judgment that was never articulated. The calibration that told a person a counterparty's silence meant a competing conversation, which they never said out loud because to them it was not a thought, it was just obvious.
• Anything decided in a room with no capture. A corridor conversation with no follow up email leaves nothing behind to read, and no tool recovers what was never recorded.
• What the person would have decided next. The record holds what was decided and the reason that was stated at the time. Those are not the same thing as the person, and treating a record as a stand in for their judgment is the failure mode worth naming out loud.

There is a fifth limit worth naming. Where the stated reason for a decision was a rationalisation written after the fact, the record carries the rationalisation in the same confident tone as everything else. That does not disqualify the approach, but a handover built on it should treat recorded reasons as evidence rather than as truth.

Tacit knowledge and merely unwritten knowledge

Most of the argument in this category collapses once you separate those two. Genuinely tacit knowledge is knowledge the holder could not write down if you paid them, and it moves only by working alongside somebody. Merely unwritten knowledge could have been written at any point and simply never was, because nobody was asked and there was no obvious place for it. In an operating role, the second category is much the larger of the two, and it is the whole of what a machine can reach.

The method transfer case, and what it exposes

An analytical method moving from one laboratory to another is the cleanest illustration in drug development. The written method and the standard operating procedure transfer as documents, and that part is easy. Whether the receiving laboratory reproduces the result usually turns on two other things: a handling detail that lives in the hands of the person who developed it and was never written, and the reason a parameter sits where it does, which is generally sitting in a thread about the run where the previous setting failed. A machine carries the second one and does not touch the first, so the honest promise is that it removes the archaeology and leaves the apprenticeship.

The same split runs through a program handover. The integrated timeline and the stage gate documents move as files. The trade made with the contract research organization to recover three weeks sits in mail, so it moves. Knowing that a particular study director says yes on a call and means probably is written nowhere and does not move at all. Which of those categories each item falls into is the subject of institutional knowledge transfer.

Using AI for knowledge transfer without overselling it

The practical shape follows from the split. Hand the recall half to the machine, and spend the scarce hours of a notice period on the half that has to be taught rather than read. That reallocation is most of the available benefit, and it is what the sequencing in the offboarding knowledge transfer template is built around.

One question separates the tools worth looking at. Ask what the thing actually reads. If it only searches documents you were already organising, it inherits your gaps, because the context you are trying to transfer was never in those documents. If it reads the correspondence where the reasoning genuinely lives, it can answer the questions a successor will really ask. The discipline level version of that distinction is AI knowledge management for a drug development team, and the underlying definition is what an AI company brain is.