Shipping the feature: rollout, monitoring, and the rollback you hope not to use
The short answer
Shipping an AI feature is a governance act, not an engineering event
A green pipeline proves the code runs, not that the model behaves on inputs your test set never saw. Launch is the moment your model first meets the real world, and it is where an owner, not just an engineer, is deciding to expose real people to a system whose behavior they cannot fully predict.
What you will be able to do
- Frame the act of shipping an AI feature as a governance decision (the moment your model first meets real users), not merely an engineering deploy, and state what that reframing changes.
- Design a staged rollout for your own feature using the exposure ladder (shadow mode, internal use, canary, percentage ramp, cohort or segment, general availability), and choose the right first rung for its consequence level.
- Build a monitoring plan that watches AI-specific signals (output-quality sampling, complaint and correction rates, groundedness, drift) and not just uptime, with pre-set thresholds and a named human who watches them.
- Write pre-registered rollback triggers: the exact, measurable conditions, agreed before launch, that fire a rollback without a meeting or a debate.
- Test a rollback before it is needed and record the result, distinguishing "we believe we can turn it off" from "we turned it off at 14:32 and it took ninety seconds."
- Distinguish the graduated responses available short of a full rollback (relabel, restrict scope, disable a subset, full off) and choose among them under pressure, as Apple did when it disabled one category and relabeled the rest.
- Account for the outputs already in the world that a rollback cannot recall, and design for that irreversibility in advance.
The ideas, one by one
Expose in stages that match the risk
A bad version should meet a recoverable number of users before it meets everyone. Climb the ladder (shadow, internal, canary, ramp, segment, general) only as evidence earns the next rung, and hold the highest-consequence segment at the lowest rung longest. Launching the riskiest content at full exposure is the rollout error underneath most AI headlines.
Watch AI-specific signals with instruments you own
The golden signals (latency, errors, saturation) are all green while a model returns a fast, successful, confidently wrong output. Add output-quality sampling, complaint and correction rates, and groundedness against source, set thresholds before launch, and name the human who watches. If your only detector is a user or a journalist, you have exposure, not monitoring.
Cadence is part of monitoring, and the watching does not end at launch
An automated alert fires at any hour; a human sample happens on a schedule, so during a live rollout the riskiest segment needs frequent sampling or continuous checks, not a once-a-day glance that leaves a failure live overnight. And because a model drifts as the world moves off its training window, a feature clean at launch can fail months later with no code change, so the sampling and drift signals run for the life of the feature (see Topic 4.5).
A rollback is tested or it is a belief
Almost everyone thinks they can turn the feature off; far fewer have done it and timed it. Test the rollback cold, record how long it took, build a feature flag so it is seconds not deploys, and name who is authorized to pull it at 2 a.m. without a meeting.
Rollback is a dial, not a button
Relabel, restrict scope, disable a subset, full off. The position you can reach is set by the granularity you built during rollout. Per-segment flags let you disable the harmful segment and keep the rest, which is exactly the scalpel Apple used to disable news summaries while keeping the others.
A rollback stops new harm; it cannot recall delivered harm
A sent notification, a published summary, a decision acted on: these are in the world, and no switch reaches them. For irreversible, public, attributed outputs, the only real protection is not exposing the risky version, which is a rollout choice made before launch, not a rollback made after.
Pre-register the stop line while you are calm
Deciding whether to roll back during an incident is deciding while tired, invested, and hopeful. Write the exact measurable trigger and the exact response before launch, so the crisis executes a decision already made. This is also your defensible record when the feature is attacked (see Topic 11.4).
A rollout plan can tell you not to ship yet, and that is the process working
If the riskiest segment never clears shadow mode, if you cannot build an instrument that speaks before an outsider, or if an irreversible output's failure rate stays above what the consequence tolerates, the honest move is to ship the parts that cleared, hold the part that did not, and say plainly what must improve first. A governance function that can never say "not yet" is narrating a launch, not governing one.
The plan is one page, and every line hides a choice
Five blocks carry it: the feature and its worst output, the per-segment exposure ladder, the monitoring with thresholds and a named watcher, the pre-registered triggers, and the rollback test record with the irreversibility line. If a reviewer can find a place where a specific was replaced by a hope, that is the line that will fail you under fire.
This plan is the first link in the shipping chain
The next topic's incident hits this feature and uses this rollback (see Topic 3.5); the post-incident review inspects this plan (see Topic 3.6); the eval suite is built to catch this failure class (see Topic 4.2); the logging records what your monitoring saw (see Topic 10.2). Build it weak and every link inherits the weakness.
Listen to it
The same lesson, talked through by two hosts as episode 20 of the podcast.
Real cases
These examples show rollout, monitoring, and rollback decisions in real systems, with the reasoning stated. The deep anchor is the Apple case; the others are referenced to sharpen a specific point and are treated in depth by their owner topics.
Example 1 (the anchor): Apple Intelligence notification summaries. Apple launched AI notification summaries as part of Apple Intelligence, grouping and condensing stacked notifications into a short line. For most content this was a convenience. For news, it fabricated. In December 2024 a summary of BBC notifications produced a false claim that a named suspect had shot himself; further examples falsely reported an arrest of Israel's prime minister (from three New York Times articles) and a darts result that had not happened yet (BBC News, January 2025). Read against this topic's three questions, Apple got exposure and monitoring wrong and rollback right. Exposure: the highest-risk content type, breaking news attributed to trusted outlets, appears to have been live at broad availability rather than held at a low rung, so the worst output met a very large audience quickly. Monitoring: the effective detector was external (a broadcaster's public complaint), which means Apple's own instruments were not catching source-unfaithful summaries before users did. Rollback: this is where Apple was strong. With the iOS 18.3 beta it disabled summaries for the "News & Entertainment" category specifically (a per-segment partial rollback, position 3 on the graduated dial) while keeping summaries elsewhere, and as an interim measure it relabeled summaries so users could tell an AI produced them (position 1), and it stated it would restore the feature once more reliable (CNBC; The Register, January 2025). The partial rollback was only possible because summaries could be governed by content category. The lasting lesson: a disciplined off switch does not compensate for launching an irreversible, high-consequence output at full exposure, because the notifications already delivered could never be recalled.
Example 2 (staged-rollout discipline, non-AI, for contrast): the CrowdStrike July 2024 outage. On 19 July 2024 the security vendor CrowdStrike pushed a faulty content configuration update (Channel File 291) to its Falcon sensor, crashing roughly 8.5 million Windows machines worldwide and causing one of the largest information-technology outages on record (CrowdStrike, "Falcon Content Update Preliminary Post Incident Report," 2024; CISA alert, 19 July 2024). The governance root cause is a rollout one: unlike the vendor's sensor binary updates, which went through staged "ring" deployment, this class of content update was pushed to the entire production fleet at once, with no canary. CrowdStrike reverted the update about seventy-eight minutes later, but by then the damage was global. This is not an AI incident, so it is not a program anchor; it is included because it is the cleanest possible illustration of the rung-3 lesson. A feature that goes to everyone simultaneously has no canary to die first, and a fast revert cannot recall the crashes that already happened. The staged-rollout principle is not AI-specific; AI simply raises the stakes because AI failures are quieter and harder to detect than a crash.
Example 3 (why fast delivery hides the failure): a summarization feature returning 200 OK. Consider any retrieval or summarization feature: it fetches sources, generates a summary, and returns it in a few hundred milliseconds with a success status. Every traditional signal (latency, error rate, saturation) is green. The one failure that matters, a claim in the summary that the source never made, is invisible to all of them. This is the general shape of the Apple failure and the reason 3D insists on groundedness sampling: for generative features, the signal that catches the real harm is a human or a check comparing output to source, not any measure of whether the output was delivered. A team that monitors only the golden signals is monitoring a summarizer's health with a thermometer that cannot detect the disease it has.
Example 4 (a fast crisis that a tested off switch answers): a live-trading deployment error. A deployment that goes wrong can do enormous damage in minutes when the system acts autonomously and fast; the canonical case is the Knight Capital trading loss, where a deployment error let an algorithm fire erroneous orders and lose about USD 440 million in roughly forty-five minutes in 2012 (SEC, Release No. 70694, 2013). The deep treatment of that night, the live incident with all hands, belongs to the next topic (see Topic 3.5). Here it makes one point about rollback: when a system acts on the world at machine speed, the value of a rollback is measured in the seconds it takes to reach and the minutes of damage it prevents, which is exactly why you test it cold. A rollback you have to figure out during the crisis is a rollback that arrives after the damage.
Example 5 (segmenting exposure by content risk, applied to your own estate). Suppose your organization ships an AI feature that drafts customer-facing replies. Not all replies carry equal risk: a reply about store hours is low-consequence; a reply about a billing dispute or a medical or legal question can create real exposure. The Apple lesson, applied, is to segment the rollout by reply type: hold the high-consequence categories at a low rung (shadow mode, human-approved) while the low-consequence categories ramp. This is not a hypothetical from a tech giant; it is a design choice available to any team, and it is the difference between a partial rollback you can execute (turn off the billing-reply segment) and a blunt one you cannot (turn off all replies). Your feature's segments come from your own "how my model fails" map (see Topic 1.6).
Example 6 (the ramp pause that earns its keep, illustrative method). Consider a team that ramps a feature 5 percent, then 25, then 50, pausing at each step. At 25 percent the correction rate on one segment ticks up past the baseline they measured in shadow mode. Because the ramp paused there, they caught it having exposed a quarter of users, not all of them; they held the ramp, sampled the outputs, found a class of inputs the model handled badly, and either fixed it or held that segment back. This is the ramp doing its only job: each pause is a decision point where the smaller population's data either earns the next step or stops it. A team that ramped straight to 100 percent would have learned the same thing from four times the harm and a much louder audience. The lesson is method, not a specific incident: the pauses are not delay, they are where the rollout pays for itself.
Example 6b (the silent failure the golden signals miss, illustrative). Picture a support-ticket summarizer that occasionally drops the one detail that mattered, a safety complaint buried in a long ticket. Latency is fine, no errors, resources normal; every golden signal is green, and the summary even reads well. The harm is invisible to delivery monitoring because the failure is an omission, not a crash, and it surfaces only when an agent, trusting the summary, never sees the safety issue. The instrument that catches it is coverage sampling: checking whether critical categories present in the source appear in the summary. The lesson generalizes: for AI features the dangerous failures are usually quiet (a wrong fact, a dropped detail, a fabricated citation), and quiet failures need instruments that read the content, not the pipe.
Example 7 (relabeling as trust calibration, not a fix). When Apple, as an interim measure, changed how notification summaries were displayed so a reader could tell an AI had produced them, it did not make the summaries more accurate; it changed what a user did with them (CNBC; The Register, January 2025). This is the lightest position on the rollback dial, and its use and its limit are both instructive. Relabeling is genuinely useful: an output a user knows is AI-generated and unverified is treated with appropriate caution, which reduces harm from over-trust while you work on the underlying problem. But relabeling is not a cure: a confident, false, labeled statement can still mislead, and a label does not un-send a delivered fabrication. Reach for relabeling as a fast, honest stopgap that buys time, never as the thing that closes the incident.
Where people go wrong
- "Shipping is an engineering decision; the pipeline is green, so we are done." Wrong for AI. The green pipeline tells you the code runs, not that the model behaves on inputs your test set never saw. Shipping is the governance moment your model first meets the real world. "Done" is not launch; launch is when the watching starts.
- "We tested it thoroughly, so we can launch to everyone." Your test set is a sample; real users are the rest of the distribution. Thorough testing lowers the odds of a bad launch; it does not remove the reason to stage. Apple tested extensively and still shipped a fabricating summarizer, because the failing inputs were the ones no test set contained. Stage anyway.
- "We have a dashboard, so we have monitoring." A dashboard of latency and error rate is blind to the defining AI failure: a fast, successful, confidently wrong output. Monitoring means watching AI-specific signals (output-quality sampling, complaint rate, groundedness) at pre-set thresholds, with a named human accountable for watching. Everything else is decoration.
- "We will know if something goes wrong." From whom? If the answer is "users will complain" or "we will see it in the press," you do not have monitoring, you have exposure with a delay. The Apple timeline is what "we will know" looks like when the knowing comes from a broadcaster. Design your own instruments to speak first.
- "We can roll back if we need to." Belief, not capability, until you have done it. The moment you need the rollback is the worst time to discover the flag is miswired or the deploy takes forty minutes. Test the rollback cold, record how long it took, and name who is authorized to pull it at 2 a.m. An untested rollback is a rollback proposal.
- "Rollback is one button: on or off." Rollback is a dial: relabel, restrict scope, disable a subset, full off. The position you can reach is set by the granularity you built during rollout. A global flag gives you only relabel and full off; per-segment flags give you the scalpel Apple used to disable news while keeping the rest.
- "If we can turn it off, we are protected." Turning it off stops new harm; it does not recall harm already delivered. A sent notification, a published summary, a decision acted on: these are in the world, and no flag reaches them. For irreversible, public, attributed outputs, the only real protection is not exposing the risky version, which is a rollout decision, not a rollback one.
- "We will decide whether to roll back if it happens." Deciding during the incident is deciding while tired, invested, and hopeful, which is when stop lines quietly stop applying. Pre-register the trigger while calm and uninvested: the exact measurable condition and the exact response, written down before launch. Pre-registration protects your users from your own future motivated reasoning.
- "A rollback means the launch failed and the team looks bad." A rollback that fires on a pre-registered trigger is the plan working, not failing. Apple's rollback was the part it got right. The failure to fear is not a clean rollback; it is a feature you cannot watch, cannot undo, and launched at full exposure.
- "The riskiest content is fine at the same exposure as everything else." Consequence is not uniform across a feature's inputs. A wrong summary of a chat is embarrassing; a wrong summary of a flash-flood warning or a breaking news alert is dangerous or defamatory. Match exposure to consequence: hold the riskiest segment at the lowest rung longest.
- "A single global on/off flag is simpler, so it is better." Simpler for you, crueler for users. A global flag leaves you only two positions, live everywhere or off everywhere, so when one segment causes harm you must either leave the harm live or kill a feature that is working for everyone else. Per-segment flags cost a little engineering up front and buy you the scalpel to disable exactly the harmful segment. The "simplicity" of one flag is paid for by users in a crisis.
- "We will move the threshold if it fires too often." If a threshold fires on normal noise, the fix is to recalibrate it above the noise band and add a duration requirement, not to quietly raise it during an incident to silence an inconvenient alert. Moving a threshold after launch to avoid an alert is the motivated reasoning pre-registration exists to prevent, and it will read as a cover-up in any later review.
- "Shadow mode is wasteful because we already tested offline." Offline tests run on a curated set you chose; shadow mode runs on the live, real input distribution, which contains the rare and adversarial inputs no test set imagined, and it gives you the real base rate of the failure that you set your thresholds from. Shadow mode is not a repeat of the test; it is the first contact with reality at zero user cost, and it is the cheapest place to learn the feature is not ready.
- "Rollback is the engineering team's job; governance just writes the policy." The person who can flip the flag at 2 a.m. without a meeting is a governance decision made concrete: who is authorized, on what pre-registered trigger, following which runbook. Leaving "who pulls it and when" to be sorted out during the incident is how a tested rollback still arrives too late. Ownership of the trigger is as much a part of the plan as the flag itself.
Questions people ask
- What is rollout and rollback plan?
- The one to two page artifact this topic produces for a shipped AI feature. It names the staged exposure (per segment), the monitoring signals with thresholds and a named watcher, the pre-registered rollback triggers and responses, a dated record that the rollback was tested, and one honest line on what a rollback cannot recall. It is the first link in the Module 3 shipping chain.
- What is shipping as a governance act?
- The reframing at the heart of this topic. Exposing an AI feature to real users is a decision to subject real people to a system whose behavior on their specific inputs cannot be fully predicted, so it needs an accountable owner, does not end at launch (launch begins the watching), and ships with its own instruments and off switch.
- What is exposure ladder?
- The ordered stages of a staged rollout: shadow mode, internal use, canary, percentage ramp, cohort or segment, general availability. A feature climbs the ladder only as evidence earns each rung, and the first rung is set by the consequence and reversibility of the worst output.
- What is shadow mode (dark launch)?
- The rung where the feature runs on real inputs and produces real outputs that no user ever sees; the outputs are logged and compared, offline, against reality or a human's judgment. It is the safest place to discover that a feature fabricates, because the failures land in a log, not on a user.
- What is canary?
- A small fixed slice of real users (often 1 percent or less) who get the feature before everyone else, named for the canary carried into a mine to detect danger first. If the canary shows trouble, only that small slice was exposed.
Keep going
This lesson builds AI resilience, continuity and exit planning, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
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 →