{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-08-05T09:56:07.156Z"},"content":[{"type":"documentation","id":"d7147a76-fab7-4731-816e-b376bb33868b","slug":"ai-staged-rollout-user-research","title":"Staged Rollout for AI Features: Shadow Mode, Canary, and Kill Switches (2026)","url":"https://www.koji.so/docs/ai-staged-rollout-user-research","summary":"Staged rollout for an AI feature is a research design, not just a deployment tactic. Shadow mode can veto a launch but never approve one, because it contains no user response to measure. The canary ring is the first stage producing real evidence: acceptance and edit rates, escalation and abandonment, guardrail metrics, and interviews explaining why the numbers moved. Rollback thresholds must be pre-registered before launch. Under EU AI Act Article 14, high-risk systems must let overseers halt the system in a safe state; Article 72 requires documented post-market monitoring, with the Commission template due 2 February 2026.","content":"**Staged rollout for an AI feature is not a deployment tactic. It is a research design.** Each stage answers a different question, and the questions are not interchangeable. Shadow mode tells you whether the model produces defensible outputs on real traffic. A canary ring tells you whether real users *behave* differently. General availability tells you whether that behaviour survives contact with people who are not early adopters. Teams that collapse these stages into one do not ship faster — they just learn later, from users who are already angry.\n\nThe uncomfortable part is that the most important failures of an AI feature are invisible to the stage most teams over-invest in. Shadow mode, by construction, has no user in it. Every metric that depends on how a person responds to the output — do they accept it, edit it, ignore it, escalate, or quietly stop using the feature — is unmeasurable until a human is on the other end.\n\nThis guide covers the four gates of an AI rollout, the specific research question that belongs to each, the rollback criteria you should write down *before* you turn anything on, and why the kill switch has quietly become a legal requirement rather than an engineering nicety.\n\n## The four gates, and the question each one owns\n\n| Stage | What runs | The question it can answer | The question it **cannot** answer |\n|---|---|---|---|\n| **Shadow mode** | Model runs on live traffic; output discarded, never shown | Does output quality hold on real inputs, not curated test data? | Anything about user behaviour or trust |\n| **Internal alpha / dogfood** | Real feature, employees only | Does the end-to-end flow work? Are there obvious harms? | Whether a normal user understands it |\n| **Canary ring** | Real feature, 1–5% of real users | Do users accept, edit, abandon, or escalate? | Whether it holds at population scale |\n| **General availability** | Everyone, with a kill switch armed | Does it hold across segments and over time? | — (this is where drift lives) |\n\nThe most common rollout mistake is treating these as increasing percentages of the same experiment. They are four different studies with four different populations and four different instruments. A canary at 1% is not \"shadow mode with users\"; it is the first point at which the feature has a user experience at all.\n\n## Shadow mode: high confidence about the wrong thing\n\nShadow mode is genuinely valuable. It catches the failures that curated evaluation sets miss, because real production traffic contains input distributions your golden set never anticipated — truncated inputs, mixed languages, adversarial phrasing, and the long tail of formatting chaos that real users generate.\n\nBut shadow mode inherits a well-documented problem from the recommender-systems literature: offline and online evaluation frequently disagree. Studies contrasting the two have repeatedly found that the ranking of algorithms by offline accuracy can contradict the ranking produced by online tests with the same users, and a 2016 RecSys paper on exactly this contrast established the pattern as a standing methodological caution rather than an anomaly. The mechanism is straightforward: historical data was generated by the *old* system, so it encodes the behaviour the old system induced, not the behaviour the new one will induce.\n\nFor AI features this problem is sharper, because the output is not a ranked list the user scans — it is a claim the user has to decide whether to believe. Two models can have identical accuracy on a held-out set and produce wildly different user outcomes, because one hedges appropriately and the other states wrong answers with total confidence. Shadow mode scores both the same.\n\n**The practical rule:** use shadow mode to *veto* a launch, never to *approve* one. It can prove a model is not ready. It can never prove it is.\n\n## Internal alpha: the least representative sample you will ever use\n\nDogfooding is cheap and catches embarrassing failures fast. It is also the single most biased population available to you. Your colleagues know what the feature is supposed to do, know the vocabulary that makes it work, have a mental model of the underlying system, and — critically — are motivated to make it succeed.\n\nThat combination systematically hides the two failure modes that matter most: users who do not understand what the feature is for, and users who trust the output more than they should. An employee who spots a subtly wrong output and mentally corrects it has not found the bug; they have masked it.\n\nTreat internal alpha as a functional smoke test, not as evidence about users. If you want one honest signal out of it, run structured interviews with the employees who used it *least*, not the ones who used it most — non-adoption inside the building is a leading indicator of non-adoption outside it.\n\n## The canary ring: the first stage that produces real evidence\n\nThis is where research earns its keep. At 1–5% of traffic, you have real users, real stakes, and a small enough blast radius to reverse. Four families of signal are worth instrumenting before you open the ring:\n\n**1. Acceptance and modification.** What fraction of outputs are accepted as-is, edited, or discarded? Edit distance is one of the highest-signal, lowest-cost metrics in AI product work, and almost nobody logs it. A feature with 95% acceptance and a feature with 95% acceptance-after-heavy-editing are different products.\n\n**2. Escalation and abandonment.** How often does the user bail out to a human, to search, or to nothing? Rising abandonment with flat satisfaction scores is the classic signature of a feature people are too polite to complain about.\n\n**3. Guardrail metrics.** These are the metrics you are *not* trying to improve but refuse to damage — task completion, latency, support contact rate, unsubscribe rate. AI launches fail on guardrails far more often than on their headline metric.\n\n**4. What people actually thought.** Behavioural telemetry tells you *what* changed; it never tells you *why*. A drop in acceptance rate is consistent with \"the model got worse\", \"the UI made the accept button harder to find\", and \"users lost trust after one bad experience last Tuesday\". Only asking distinguishes them.\n\n### Pre-register your rollback criteria\n\nGoogle's People + AI Guidebook offers a template worth adopting verbatim: *\"If {metric} for {feature} {drops below/goes above} {threshold} we will {action}.\"* Fill it in before the ring opens, and get the owner to sign it.\n\nThe reason to pre-register is not process hygiene. It is that the moment a canary starts wobbling, every incentive in the room points toward explaining the number away — small sample, seasonal effect, one noisy cohort. A threshold agreed in advance, by people who did not yet know which way the data would break, is the only version of the decision you can trust.\n\nNote the sample-size trap here too: a 1% canary is often far too small to detect the effect size you care about. Work out the [minimum detectable effect](/docs/statistical-power-minimum-detectable-effect) for your ring *before* you promise stakeholders a read on the metric. A ring that cannot detect a 5-point drop should not be described as \"monitoring for regressions\".\n\n## The kill switch is now a legal requirement, not a nice-to-have\n\nFor high-risk AI systems under the EU AI Act, human oversight is not a design preference. Article 14 requires that systems be built so that people can effectively oversee them, and it specifies that overseers must be able to *\"intervene in the operation of the high-risk AI system or interrupt the system through a 'stop' button or a similar procedure that allows the system to come to a halt in a safe state.\"*\n\nTwo details are routinely under-read. First, Article 14 also requires that overseers remain aware of *\"the possible tendency to rely on the system\"* — automation bias is named in the legislation itself, which means \"our reviewer approved every output\" is evidence of a problem, not of diligence. Second, the halt must land in a **safe state**. A kill switch that drops the user into a broken screen with no path forward is not compliance; it is a second incident.\n\nBeyond the switch itself, Article 72 requires providers of high-risk systems to *\"establish and document a post-market monitoring system in a manner that is proportionate to the nature of the AI technologies and the risks\"*, actively and systematically collecting and analysing performance data across the system's lifetime. The Commission was required to adopt an implementing act establishing a template for the post-market monitoring plan by **2 February 2026** — so the paperwork phase of this obligation is no longer hypothetical.\n\nThe research implication: post-market monitoring that consists only of dashboards is incomplete. Telemetry shows you the metric moved. Only user evidence explains the movement, and \"analyse relevant data on performance\" is not satisfied by a chart nobody can interpret.\n\n## Why staging matters more in 2026 than it did in 2023\n\nDORA's 2025 *State of AI-assisted Software Development* report frames AI's effect on engineering organisations bluntly: AI's primary role is *\"as an amplifier, magnifying an organization's existing strengths and weaknesses.\"* DORA's earlier 2024 research had already found that AI adoption negatively affected software delivery stability and throughput — teams generate change faster than their review, testing, and QA capacity can absorb it.\n\nThat is the case for staging in one sentence. If your release process is disciplined, AI makes it faster. If it is not, AI makes the mess arrive sooner and in higher volume. Staged rollout is the mechanism that converts velocity into something survivable.\n\nThe macro numbers point the same way. Gartner projected that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, and unclear business value. MIT's Project NANDA study *The GenAI Divide: State of AI in Business 2025* — built on 150 executive interviews, 350 employee surveys, and 300 public deployments — reported that roughly 95% of generative AI pilots produced no measurable P&L impact. Most of those did not fail at the model. They failed somewhere between the demo and the user, which is precisely the territory a staged rollout is designed to survey.\n\n## What to ask users at each stage\n\n| Stage | Method | Sample | The question worth asking |\n|---|---|---|---|\n| Internal alpha | Short structured interview | 8–12 low-adoption employees | \"What made you stop using it?\" |\n| Canary, week 1 | In-product interview at the moment of use | 30–50 | \"What did you expect it to do, and what did it do?\" |\n| Canary, week 2–3 | Segmented interviews | 60–100 across segments | \"When did you decide whether to trust it?\" |\n| Pre-GA | Guardrail check | 100+ | \"What would make you turn this off?\" |\n| Post-GA | Continuous, triggered on drop-off | Rolling | \"What happened right before you stopped?\" |\n\nThe last row is the one teams skip and then regret. Drift is not just a model phenomenon; user expectations drift too. A feature that was delightful when it was novel becomes table stakes, and the bar for \"good enough\" rises underneath you without any change to the model at all.\n\n## How Koji fits into a staged rollout\n\nThe bottleneck in stage-gated research has never been knowing what to ask. It is that traditional research takes two to three weeks per round — recruit, schedule, moderate, transcribe, analyse — while a canary ring lives for days. So teams either stall the rollout waiting for insight, or ship blind and call the telemetry \"research\".\n\nKoji is built for the second timescale. Because interviews are AI-moderated and run around the clock, a canary cohort can be interviewed the same day the ring opens, and the analysis is finished when the last participant is. In practice that turns a stage gate from a three-week pause into an overnight one.\n\nSpecifically:\n\n- **AI-moderated interviews at the moment of use.** Trigger an interview when a user accepts, edits, or abandons an AI output — while the reasoning is still fresh, rather than reconstructed a fortnight later.\n- **Voice interviews** for the canary stage, where you want the unguarded version. People say things out loud that they will not type into a feedback box.\n- **Automatic thematic analysis.** Fifty canary interviews are coded and themed without anyone spending a weekend in a spreadsheet — the difference between a stage gate you can actually hold and one you skip under deadline.\n- **Structured questions alongside open conversation.** Koji supports six types — `open_ended`, `scale`, `single_choice`, `multiple_choice`, `ranking`, and `yes_no` — so one instrument produces both the quantitative guardrail reading and the explanation behind it. A `scale` question gives you the trend line across rings; the `open_ended` follow-up tells you why the line moved. See the [structured questions guide](/docs/structured-questions-guide) for how to combine them.\n- **A customisable AI consultant** you can point at the specific failure mode this ring is meant to surface, instead of running a generic satisfaction survey.\n- **Real-time reporting**, so the rollback decision is made against evidence that is hours old rather than weeks old.\n\nCompare that with the legacy pattern: a SurveyMonkey form fielded to a canary cohort returns closed-ended answers with no follow-up, which tells you the score dropped and nothing about the cause. The whole value of a canary is diagnostic, and a static form is the one instrument that cannot diagnose.\n\n## Common mistakes\n\n1. **Treating shadow mode as a launch approval.** It can veto, never approve.\n2. **Sizing the canary for comfort rather than power.** If the ring cannot detect the regression you fear, it is theatre.\n3. **Writing rollback criteria after the numbers arrive.** By then everyone has a story.\n4. **A kill switch that lands users nowhere.** Halting into a broken state is a second failure, and under Article 14 it is not a safe state.\n5. **Dogfooding as user evidence.** Your colleagues are the least representative sample you have access to.\n6. **Watching only the headline metric.** AI launches usually die on guardrails.\n7. **Stopping research at GA.** Post-market monitoring is a legal obligation for high-risk systems and a good idea for everything else.\n\n## Frequently asked questions\n\n### What is the difference between shadow mode and a canary release?\nShadow mode runs the new model on live traffic but discards the output — no user ever sees it, so it measures model behaviour only. A canary release shows the output to a small slice of real users, so it measures user behaviour. They answer different questions, and shadow mode cannot substitute for a canary because it contains no human response to measure.\n\n### How big should a canary ring be?\nBig enough to detect the regression you are worried about, which is usually larger than teams assume. Calculate the minimum detectable effect first: a ring sized at 1% of traffic may have a detectable effect of 10 percentage points or more, meaning a real 5-point drop would pass unnoticed. If the ring cannot detect the effect you care about, either enlarge it, extend the exposure window, or be explicit that it is a safety check rather than a measurement.\n\n### Does the EU AI Act require a kill switch?\nFor high-risk AI systems, effectively yes. Article 14 requires that human overseers be able to intervene or interrupt the system through a stop button or similar procedure that brings it to a halt in a safe state. It is framed as a human-oversight obligation rather than as a \"kill switch\" requirement, but the practical engineering consequence is the same. Most consumer AI features are not high-risk under the Act — but the design principle transfers regardless of classification.\n\n### What is post-market monitoring under the AI Act?\nArticle 72 requires providers of high-risk AI systems to establish and document a post-market monitoring system, proportionate to the technology and its risks, that actively and systematically collects and analyses performance data throughout the system's lifetime. The Commission was required to adopt a template for the monitoring plan by 2 February 2026. Dashboards alone rarely satisfy the analysis requirement — you need to be able to explain what the data means, which generally requires user evidence.\n\n### Should we run user research during shadow mode?\nYes, but not on the shadow output. Shadow mode is the right time to research the *decision* rather than the feature: interview users about the current experience so you have a genuine baseline, and pressure-test your acceptance criteria against what users say matters. What you cannot do is ask users about outputs they never saw.\n\n### How do we decide when to roll back versus push forward?\nDecide before you launch, in writing, using a template like \"if this metric crosses this threshold, we take this action\". Include at least one guardrail metric you are not trying to improve, and one qualitative trigger — for example, \"if more than two canary participants describe the feature as untrustworthy, we pause and investigate\". Pre-registration is what stops a wobbling metric from being argued away in the room.\n\n### Can AI-moderated interviews keep up with a canary that only runs for a week?\nThat is the specific problem they solve. A traditional interview round — recruiting, scheduling, moderating, transcribing, analysing — takes two to three weeks, which is longer than most canary rings exist. AI-moderated interviews run 24/7 and are analysed as they complete, so a cohort recruited on Monday can inform a Wednesday go/no-go decision.\n\n## Related Resources\n\n- [Structured Questions in AI Interviews](/docs/structured-questions-guide) — the six question types and how to combine quantitative and open-ended signal in one study\n- [Acceptance Criteria for AI Features](/docs/ai-feature-acceptance-criteria) — how to define \"good enough to ship\" before the rollout starts\n- [AI Failure Mode Analysis](/docs/ai-failure-mode-analysis) — an FMEA framework for anticipating what a rollout should watch for\n- [AI Incident Postmortems](/docs/ai-incident-postmortem-user-research) — what to do when a rollout goes wrong anyway\n- [Evaluation Datasets for AI Products](/docs/ai-evaluation-dataset-golden-set) — building the golden set that shadow mode runs against\n- [Statistical Power and Minimum Detectable Effect](/docs/statistical-power-minimum-detectable-effect) — sizing a canary ring so it can actually detect a regression\n- [The EU AI Act and User Research](/docs/eu-ai-act-user-research-compliance) — the compliance context for Articles 14 and 72\n- [User Research for AI Products](/docs/user-research-for-ai-products) — the broader methodology this fits inside\n\n---\n\n**Run your canary research overnight, not over three weeks.** Koji gives you 10 free credits to start — enough to run your first AI-moderated interviews with a canary cohort and see the analysis the same day.","category":"Research Methods","lastModified":"2026-08-05T03:24:45.421037+00:00","metaTitle":"Staged Rollout for AI Features: Shadow Mode, Canary & Kill Switches (2026)","metaDescription":"What shadow mode can and cannot measure, what to ask users at each canary ring, how to pre-register rollback thresholds, and why the EU AI Act made the kill switch a legal requirement.","keywords":["staged rollout ai","shadow mode ai","canary release research","kill switch ai","feature flag rollout","ai rollout plan","post-market monitoring ai act","progressive delivery ai"],"aiSummary":"Staged rollout for an AI feature is a research design, not just a deployment tactic. Shadow mode can veto a launch but never approve one, because it contains no user response to measure. The canary ring is the first stage producing real evidence: acceptance and edit rates, escalation and abandonment, guardrail metrics, and interviews explaining why the numbers moved. Rollback thresholds must be pre-registered before launch. Under EU AI Act Article 14, high-risk systems must let overseers halt the system in a safe state; Article 72 requires documented post-market monitoring, with the Commission template due 2 February 2026.","aiPrerequisites":["Basic familiarity with feature flags or phased releases","An AI feature approaching launch","Access to product telemetry"],"aiLearningOutcomes":["Choose the right research question for each rollout stage","Understand why shadow mode cannot approve a launch","Instrument a canary ring for acceptance, escalation and guardrail signal","Pre-register rollback criteria that survive contact with stakeholders","Meet EU AI Act Article 14 and Article 72 obligations with user evidence"],"aiDifficulty":"intermediate","aiEstimatedTime":"13 min"}],"pagination":{"total":1,"returned":1,"offset":0}}