Who owns the answer
Every flag has exactly one role that can close it, and naming that role in the same sentence as the flag is what separates an escalation from a complaint.
Your flag log now has rows with a real observation in them and an empty owner column. That column is the whole difference between a flag and a complaint, and it is also where most well-meaning escalations die: the concern is correct, it goes to the person who happened to be in the room, and it is never seen again.
This lesson fills the column. It is short because the discipline is small: for every flag, name the single role that can close it, and put that role in the same sentence as the flag.
The one line a standard actually gives you
There is no standard that says “one owner per flag.” There is one that says accountability has to be legible, and it is worth having in your pocket because it is short and it is NIST:
“Roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear to individuals and teams throughout the organization.” (NIST AI RMF 1.0, GOVERN 2.1)
The same document gives the reason, a few pages earlier: “Different AI actors have different responsibilities and awareness depending on their roles in the lifecycle. For example, organizations developing an AI system often will not have information about how the system may be used.”
Read that as a description of your own position. You are the actor who can see the rendering surface and cannot see the retention decision, the contract, or the audit boundary. Somebody else is the reverse. That asymmetry is not a dysfunction to route around; it is the reason escalation exists.
The test: who can say the closing sentence
The owner is not the person who cares most, is not the most senior person in the thread, and is often not the person who caused the situation. The test is narrower and it is mechanical:
Which single role can say a sentence that would let me mark this row closed?
Run that on a flag and the answer usually collapses to one of a small set of roles. These are the ones this course keeps meeting:
| Role | Owns the answer when the question is |
|---|---|
| Component owner | What does this renderer permit, and what does it sanitise. Often you. |
| Application security | Whether a proposed defence is the right class of defence, and what the residual exposure is. |
| Platform or agent runtime owner | Which tools the agent may call, what the executor runs, what the trace store holds. |
| Telemetry stack owner | What an SDK records at the version installed, and what a version bump changed. |
| Privacy or data protection owner | Whether holding, recording or transferring this data is permitted, and for how long. |
| Compliance owner | What is inside an audit boundary, and what a report does and does not say. |
| Product owner | Whether a capability needs to exist at all. Usually the only person who can remove one. |
Two traps are worth naming. The first is routing everything to security, which is a catch-all that feels responsible and mostly produces a backlog. Almost nothing in the data module is a security question; a replay script recording an identity document is a privacy question with a telemetry-configuration prerequisite. The second is routing upward. The operating partner can make somebody answer and cannot answer.
A flag with two owners is two flags
This is the most common defect in a first flag log, and it is worth working through one example properly.
The observation from the data module: a session-replay script is running on the extraction review gate, which displays a source document crop and the fields pulled out of it. Write that as one flag and the owner column gets two roles in it, which means it gets neither.
Split it, and both halves become closable:
- What does this SDK record on this surface, at the version we have installed? Owner: the telemetry stack owner. This is a factual question with a documented answer, and it is answerable in an afternoon by reading the vendor’s own current documentation rather than remembering a default.
- Given that answer, is recording this surface permitted? Owner: the privacy or data protection owner. This is a judgement about purpose, basis and commitment, and the privacy owner usually cannot tell you what the SDK does, which is why it is second and not first.
The order matters as much as the split. The second question is unanswerable without the first, so raising them together produces a thread where the privacy owner waits on an engineer who was not told they were blocking anyone.
Where people get burned
A blank owner column is information and an invented one is not. If you genuinely cannot tell who owns a flag, the flag is “who owns this?”, addressed to whoever runs the engineering organisation, and it closes when somebody names a role. That is a legitimate flag and it is a much better look than guessing at a company you joined three weeks ago.
Check your recall
Answer from memory — no scrolling back.
Retrieval check
You raise the carve-out question from the SOC 2 lesson to the head of engineering, who is helpful and technical. Why is that the wrong destination even though something will probably happen?
Check your answer
Because they cannot say the sentence that closes it. The question is whether a vendor is treated as a subservice organization in a particular report, under the carve-out or the inclusive method. The person who knows is the one the service auditor talks to. Everyone else, however capable, is forwarding.
Something will happen, and what happens is a two-week relay with the detail worn off at each hop. Naming the compliance owner in the same sentence as the flag costs nothing and removes the relay.
Hands on
Fill the owner column, and split what will not fit
Done when: Every row in FLAG-LOG.md has exactly one role in the owner column, or is split into rows that do, or says “unknown” with a named person you will ask. No row has two roles. Each split records which of the two must be answered first, and why.
- Open
FLAG-LOG.mdand go row by row. For each, ask the closing-sentence test out loud rather than in your head — it catches the rows where you are about to write a team name instead of a role. - Write a role, never a person. “The privacy owner”, not “Priya”. Roles survive people leaving and they transfer between portfolio companies, which is the entire point.
- Where two roles appear, split the row in two. Give each half its own observation and its own question. Do not split by seniority; split by which question is factual and which is a judgement.
- Order each split pair. Write which one has to be answered first and one clause saying why. If neither depends on the other, they were probably one flag after all — recombine and look again.
- For anything left unowned, write “unknown” and the name of the one person you will ask who owns it. That is a real row with a real next action, and it is honest in a way a guess is not.
- Count the rows with two roles. The target is zero, and the count is the only pass condition this exercise has.
What this does not cover
Naming an owner is not raising the flag. A correctly routed flag with no severity attached, and no proposal, arrives on the owner’s desk as an open-ended obligation with a release date behind it — which is how an architect becomes the person whose flags get triaged last. The next lesson, on raising it without becoming the blocker, is about the other three things a flag carries: your read of the severity, whether you are actually asking to block, and a proposal small enough to say yes to.
It also does not cover what happens after the owner replies. A reply is not a closure, and the difference between them is the last lesson in this course.
Read this next — primary source
Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1National Institute of Standards and Technology, January 2023 — free PDF, fetched 2026-09-05. A standards body with nothing to sell, which is rarer in this subject than it should be.
This lesson takes GOVERN 2.1 and the short passage at 1.2.4 on AI actors having different responsibilities, which is the closest thing to a normative statement that somebody has to be named. Read the GOVERN function of the Core rather than the whole framework — it is a couple of pages of table. Where it stops: NIST says roles must be documented and clear. It does not say one role per issue, it does not tell you which role owns any particular flag, and it is written for the organisation rather than for one architect inside it. The narrowing is this course’s.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.