Skip to main content

The handover: what you now know that no essay could have taught you, and where you take it

The short answer

The handover is the topic where your work stops being about you

A board inspection and a viva test whether you can defend the work now. The handover tests whether the work survives your absence. The systems outlast the people; the only question is whether the reasoning does too.

What you will be able to do

  • Name the specific expert judgment you now hold that could not have come from reading, because it came from building, breaking, defending, and shipping AI under real consequences across this program.
  • Distinguish the knowledge that transfers cleanly in a document (the dossier, the decisions, the evidence) from the tacit judgment that does not, and design a handover that carries both.
  • Assemble a handover pack from your finished dossier that a named successor can inherit and operate from on day one, without you in the room.
  • Identify the three things that reliably fail to transfer in a folder (tacit judgment, live ownership and relationships, the running currency watch) and build a deliberate mechanism to carry each one across.
  • Explain, using the Detroit facial-recognition audit, why an institution that keeps the harm but loses the record is condemned to repeat the harm, and how a living handover breaks that cycle.
  • Set your own operating cadence for the work after this program: the living credential you carry, the weekly frontier hour, and the review triggers that keep your governance from rotting into a snapshot.
  • Run a successor cold-read test that proves a stranger can navigate and act on your handover under time pressure, before a real successor ever has to.
  • Attach a one-line reason to every governance control you own, producing a guardrail register so no successor mistakes a hard-won limit for arbitrary friction and removes it.
  • Triage a handover when the overlap is short or absent, transferring the living and fragile knowledge before the settled record the dossier already holds.
  • State where you take this command next, in plain terms: the professional identity you now hold, why it out-operates a credential earned by study alone, and the ethical through-line that survives every tool change ahead.

The ideas, one by one

You now hold knowledge no essay could give you, and you must name it to hand it over

You know what it feels like when your own model fails, that every fix has a cost, what an attack on your file feels like, that the record is the decision, and, most valuably, where your own judgment is uncertain. That last one is the hardest to write and the most important to transfer.

The dossier transfers cleanly; the judgment that built it does not, unless you design for it

Tacit judgment crosses by worked example, live relationships cross by warm introduction, and the currency watch crosses by shared practice. All three transfer in motion, during an overlap, never as a document left on a desk.

The record outlives you, or the harm does

Detroit kept the harm of a wrongful facial-recognition arrest and lost the reasoning behind its searches, so a court had to force a retroactive audit of every case since 2017. An institution that designs its handover never ends up reconstructing its own record from the outside after the damage.

Every guardrail needs its reason attached, or a successor will cut it as friction

A rule with no recorded reason looks arbitrary to an efficiency-minded successor, who removes it, at which point the original harm returns. One line of reason per limit is the difference between a scar that is kept and one that is mistaken for a wart.

Filing is not receiving; the deliverable is a pack a stranger can act on

Prove it with a cold-read test: a real colleague, one hour, four questions, no clarifications needed. A record no one can navigate under pressure is functionally no record, as both NASA's lessons-learned experience and the Detroit audit show.

Your credential is living, not a transcript, and it is maintained by cadence

The credential opens your dossier so others see the evidence; it does not claim you know governance. Keep the dossier alive with a weekly frontier hour, event-based review triggers, and self-disclosure as a reflex, or the credential becomes a claim your own file no longer supports.

Your corrected mistakes are the cheapest lessons a successor can inherit

Somewhere in the year you changed your mind: approved then restricted, doubted a control until an incident vindicated it, trusted a vendor until they failed a test. Handing over "here is what I was wrong about and why I changed" saves the next person from paying the same tuition. A record of judgment that corrected itself beats a suspiciously clean one that never had to.

A short overlap is a triaged handover, not a failed one

The discipline degrades gracefully. When you cannot transfer everything, transfer the living and fragile first (open decisions, uncertainty map, guardrail reasons, warm introductions) and let the settled record, which the dossier already carries, come last. The failure is never a short overlap; it is spending it on the wrong things.

You take a jurisdiction-independent identity into a decade of churn

The specific laws and tools will change; the reflexes will not: build before you govern, assume attack, design evidence before you gather it, read the primary source. Underneath them is the ethical through-line, keeping faith with a person outside your institution who cannot see the system that could harm them.

Listen to it

The same lesson, talked through by two hosts as episode 98 of the podcast.

Real cases

These examples show institutional record and knowledge transfer at work, and what happens when it fails. The classification reasoning is stated so you can transfer it to your own seat.

Example 1: The Detroit facial-recognition audit as a forced handover. After Robert Williams's wrongful arrest, the 2024 settlement compelled the Detroit Police Department to audit every case since 2017 where FRT contributed to an arrest warrant, and to stop arresting on a face match alone. Source: ACLU, June 28, 2024. This is the anchor lesson made concrete: because no living record of FRT-search reasoning had been maintained and handed forward, the only remedy left was a costly retroactive reconstruction imposed by outsiders. An institution that designs its handover never ends up here. One that does not, inherits the harm and pays strangers to find the record.

Example 2: Amazon's Rekognition moratorium as a handover to legislators. In June 2020 Amazon paused police use of its Rekognition FRT for one year and said the pause was to give Congress time to legislate. Source: Amazon, June 10, 2020. Read as a handover, it is a vendor recognizing that a decision it could not responsibly keep making alone had to transfer to a body with the authority to make it. The instructive part is what happens to a handover with no receiving structure: no comprehensive federal FRT statute followed, and the question diffused back into a patchwork of state and city rules. A handover with no one designated to catch it is a drop, not a transfer.

Example 3: Aviation shift handover logs (established practice). Air traffic control and flight operations run on structured position-relief briefings: the person leaving a control position hands the incoming controller a defined account of traffic, equipment status, and anything unusual, and the incoming controller confirms understanding before authority passes. The discipline exists because the cost of a lost detail is measured in lives, and because the person arriving cannot see what the person leaving was worried about unless it is stated. The lesson for AI governance: authority should not transfer until the receiver has confirmed they can operate, not merely received a file.

Example 4: Clinical handover and the SBAR pattern (established practice). Hospitals moved patient handovers onto structured formats such as SBAR (Situation, Background, Assessment, Recommendation) precisely because unstructured verbal handovers dropped critical information between shifts. The structure forces the outgoing clinician to state not just the facts but their assessment and their recommendation, the judgment layer. This maps directly onto the handover pack: facts alone are the dossier; assessment and recommendation are the tacit judgment that a good handover forces into the open.

Example 5: ISO/IEC 42001 "documented information" and continuity. ISO/IEC 42001:2023, the international AI management system standard, requires an organization to create and control documented information so that the management system persists independently of any individual. Source: ISO/IEC 42001:2023. The standard is, in effect, a continuity requirement: the system must be reconstructable from the record, not from a person's memory. A governance function that meets this in spirit has already solved most of the handover problem, because it never let the reasoning live only in one head.

Example 6: NASA lessons-learned systems (established practice). After costly failures, NASA institutionalized lessons-learned repositories so that the reasoning behind a failure, and the guardrails put in place after it, would transfer to engineers who were not present when the lesson was paid for. The recurring difficulty, well documented in the organization's own reviews, is that a lessons-learned database is only as good as whether the next engineer actually reads and applies it. The lesson for you: a handover that is filed but not received is the same as no handover. Reception, not filing, is the deliverable.

Example 7: The successor who removes a guardrail nobody explained. This pattern is common enough across regulated industries to state as a general case rather than a single incident: a control is put in place for a reason, the reason is never written into the record, the person who understood it leaves, and a successor optimizing for efficiency removes the control as pointless friction, after which the original harm returns. It is the everyday, undramatic version of the Detroit failure. The mechanism is always the same: the guardrail transferred but the reasoning behind it did not, so the guardrail looked arbitrary and was discarded. Every limit in your dossier should carry, in one line, the reason it exists, so no successor mistakes a scar for a wart.

Example 8: The EU AI Act's own continuity logic for high-risk systems. The EU AI Act (Regulation (EU) 2024/1689) requires providers of high-risk AI systems to keep technical documentation and logs so the system's behavior and compliance can be reconstructed by someone who was not present when it was built. (see Topic 5.6) This is a legal handover requirement in disguise: the law does not trust that the people who built a high-risk system will still be around to explain it, so it mandates a record that outlives them. The timeline for those stand-alone high-risk obligations was deferred to 2 December 2027 under the 2026 Digital Omnibus simplification package, so the obligation is emerging rather than yet in force for most systems, but the design intent is already the lesson: a regulator, like a successor, must be able to inherit the reasoning from the file alone. Source: Regulation (EU) 2024/1689; Council of the EU, Digital Omnibus adoption, 29 June 2026, per CURRENCY-BRIEF.

Example 9: The whistleblower's records as a handover the institution did not want. When an individual inside an AI project raises a concern and preserves the documents that support it, they are performing a kind of forced handover of institutional knowledge to an outside body: a regulator, a court, or the public. (see Topic 10.6) The uncomfortable lesson for a governance professional is this: if your institution will not maintain an honest internal record, the record does not cease to exist. It simply gets assembled later, by someone else, on terms you do not control, exactly as the Detroit audit was. The defensible position is to keep the honest internal record yourself, so that the account of your systems is one you authored deliberately rather than one reconstructed adversarially around you.

Example 10: Open-source model cards as a vendor's handover to every downstream deployer. When a model provider publishes a model card or system card, it is handing over to strangers, the downstream deployers it will never meet, the knowledge of what the model does, how it was evaluated, and where it is known to fail. (see Topic 12.3) The good cards read like a handover pack: they state the intended use, the tested limits, and the honest uncertainty, so a deployer can inherit the model responsibly. The thin ones read like a document drop: a name and a benchmark score, with the reasoning and the failure modes left in the provider's head. The lesson transfers directly to your own handover: a card, like a pack, is judged not by how impressive it looks but by whether a stranger can safely operate from it, and the honest disclosure of limits is the part that makes it usable rather than the part that makes it look weak.

Example 11: Government transitions and the institutional-memory gap. When an administration changes, incoming officials inherit programs, systems, and commitments from predecessors who are gone, often with incomplete records of why decisions were made. Career civil-service structures and formal transition briefings exist precisely to carry institutional memory across the change, because the alternative is each new team relearning, at public expense, lessons the last team already paid for. The parallel to AI governance is exact: the systems and their risks persist across the staff change, and only a deliberate transfer of the reasoning, not just the assets, keeps the new team from removing a guardrail or repeating a mistake whose rationale left with the people who understood it.

Where people go wrong

  • "The handover is the dossier." The dossier is the "what." The handover is the "what" plus the "why" plus the living practices (the open decisions, the currency watch, the tacit judgment, the relationships). A successor who inherits only the dossier inherits a snapshot and none of the motion that keeps it true. The dossier transfers cleanly; the judgment that built it does not, unless you design for it.
  • "A handover is a moment, so a document is enough." The parts that matter most, tacit judgment and live relationships, do not transfer in a document. They transfer through an overlap: a period where you and your successor operate together and the invisible knowledge crosses the way it was built, by doing. A document left on a desk is the weakest possible handover.
  • "I should hand over only my confident conclusions." The opposite. The most valuable thing you can transfer is an honest map of where your own judgment is uncertain, because that is the knowledge a successor cannot get any other way and will otherwise learn only by being harmed by it. Handing over only confidence hands over a trap.
  • "Guardrails are self-explanatory; the rule speaks for itself." A rule with no recorded reason looks like arbitrary friction to a successor optimizing for efficiency, who will eventually remove it, at which point the original harm returns. Every limit needs its reason attached in one line, or it is a scar a successor will mistake for a wart and cut off.
  • "If the record is filed, the handover is done." Filing is not receiving. The NASA lessons-learned experience and the Detroit audit both show that a record no one reads or can navigate is functionally no record at all. The deliverable is a pack a stranger can act on under time pressure, confirmed by an actual cold-read test, not a folder you are confident is complete.
  • "The currency watch is a list of current laws I can write down." A list is stale the day you write it. What must transfer is the running process: the primary sources, the cadence, the triggers. Hand over the reading habit, not the reading's latest output. Otherwise your successor inherits a dossier that was accurate on a date that keeps receding.
  • "Once I leave, it is the successor's problem." The people harmed by a governance failure are almost never the successor or you; they are outside the institution entirely and cannot trace their harm back to a botched handover. The handover is how the institution keeps faith with those people across a change of staff. Treating it as an internal courtesy misses who it actually protects.
  • "The credential proves I know governance." The living credential does not claim knowledge; it opens your dossier so others can see the evidence of what you did. If you let the dossier rot after the program, the credential becomes a claim your own file no longer supports. The identity is maintained by cadence, not conferred by completion.
  • "The handover is just for the successor, so an audit response is separate work." The pack a successor can operate from is very nearly the pack an auditor wants to trust, because both readers ask the same underlying question from different angles. Build it once, to the standard that a stranger can operate from it and a skeptic can trust it, and you have pre-assembled most of your next audit response. Treating the two as separate builds the same evidence twice, badly, under time pressure.
  • "A short overlap means a bad handover, and there is nothing I can do." A short overlap means a triaged handover, not a failed one. The discipline degrades gracefully: save the living and fragile parts first (open decisions, the uncertainty map, guardrail reasons, warm introductions) and let the settled record, which the dossier already carries, come last. The failure is not the shortness of the overlap; it is spending a short overlap polishing the settled work while the fragile knowledge walks out with you.
  • "Handing over my mistakes makes me look bad." Your corrected mistakes are the cheapest lessons your successor will ever get, because someone already paid for them. A handover that says "here is what I was wrong about and what changed my mind" saves the next person from paying the same tuition twice. A record of judgment that corrected itself is more useful, and more credible, than a suspiciously clean record of judgment that never had to.

Questions people ask

What is handover?
The transfer of an AI governance function from one person to their successor, comprising not only the evidence file (the dossier) but the open decisions, the living currency watch, the tacit judgment, and the working relationships needed to keep the file alive and act on it. A handover is an overlapping period of shared operation, not a single document drop. More on Handover
What is handover pack?
The concrete artifact of a handover, built on top of the dossier: an open-decisions register, a tacit-judgment transfer note, a living currency-watch subscription, warm introductions, and a confidence-and-uncertainty map. Designed so a successor can operate from it during the outgoing lead's total absence.
What is dossier?
The structured, indexed, claim-to-evidence file assembled in Topic 13.1 from every artifact produced across the program. The "what" of governance: what an organization runs, what it decided, and the evidence for each. It transfers cleanly because it was built to; the judgment behind it does not. (see Topic 13.1) More on Dossier
What is tacit judgment?
Expert knowledge acquired by doing rather than reading, which the holder often no longer notices holding: instincts about which claims to distrust, which fixes carry hidden costs, and where one's own past reasoning is uncertain. It does not transfer by statement; it transfers by worked example.
What is open decision?
A governance decision that is pending or actively being avoided rather than settled. Open decisions are where a successor is first exposed to risk, so a handover surfaces them, with current thinking and the reason for the delay, before the settled decisions.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →