Locate the work
Choose the stage that describes the decision being made now.
03 · The operating system
Nine stages.
One quality conversation.
Every stage connects a purpose, AI contribution, human responsibility, artefact, checkpoint, quality gate, risk, and measure.
Use it in five minutes
Read the boundary
See what AI can accelerate and what the designer must own.
Run the checkpoint
Apply the quality gate before the work moves forward.
Quality operating system
Quality travels with the work.
The guardrail layer connects intent, evidence, trade-offs, accountability, and monitoring across all nine stages.
01Goal
02Risk level
03Quality dimensions
04Evidence
05Trade-offs
06Impact
07Human sign-off
08Monitoring
Stop / rework Unacceptable risk, weak evidence, or unresolved critical failure.
Quality drift Re-enter when behaviour, content, models, or outcomes change.
Connect design to measurable user and business change.
Make tasks understandable, efficient, and recoverable.
Support abilities, contexts, languages, devices, and access needs.
Make purpose, choices, status, and consequences visible.
Align states, channels, content, components, and agents.
Communicate capability, limitations, uncertainty, and provenance.
Enable edit, pause, reject, undo, exit, and recovery.
Reduce privacy, security, bias, manipulation, and misuse risk.
Design for errors, latency, interruption, and model failure.
The process
Choose where you are.
The process is a continuous loop. Stage 09 feeds validated learning into the next Stage 01 charter.
01Opportunity Framing02Intelligence Gathering03Insight Generation04Concept Exploration05Experience Architecture06Experience Crafting07Experience Validation08Build Enablement09Continuous Learning
01Formerly RequirementsOpportunity Framing
Establish what problem is actually being solved, for whom, and under what constraints, before any research spend begins
+
Opportunity Framing
Establish what problem is actually being solved, for whom, and under what constraints, before any research spend begins
Human owns
Facilitation of the live workshop; reading organizational politics and unstated authority; all commercial/contractual decisions; final charter sign-off
AI accelerates
Structured intake at scale (parallel async interviews instead of serial scheduled ones); first-pass synthesis and conflict-flagging so the workshop starts at decision-making, not information-gathering
- Inputs
- Prior RFP/SOW/contract language, existing product documentation, stakeholder roster, past project retrospectives, any existing analytics or support-ticket access
- Deliverables
- Stakeholder interview synthesis, conflict log, draft-then-approved project charter (objectives, scope boundaries, constraints, success metrics)
- Human checkpoint
- (1) Workshop output reviewed against AI-drafted brief, not rubber-stamped. (2) Charter formally signed off by accountable human before Stage 02 begins, this is the Framework's hardest gate; nothing downstream is trustworthy if this is skipped
- Gate to advance
- Charter has a named accountable approver, explicit success metrics, and every flagged stakeholder conflict has a documented resolution (not just a note that it exists)
- Risks
- Treating the AI-drafted brief as final instead of a workshop starting point; AI missing context that only exists in someone's head and was never written down or said in an interview; stakeholder data itself being sensitive in regulated verticals
- Measures
- Time from kickoff to signed charter; number of stakeholder conflicts surfaced pre-workshop vs. discovered mid-project (lower post-charter discovery = better framing); charter revision rate during Stage 02–03
- Enterprise outputs
- Signed project charter, the artifact that appears in the SOW and gets referenced in every later stakeholder conversation about scope
02Formerly ResearchIntelligence Gathering
Build a verified fact base, user behavior, market landscape, competitive structure, grounded in the charter's success metrics from Stage 01
+
Intelligence Gathering
Build a verified fact base, user behavior, market landscape, competitive structure, grounded in the charter's success metrics from Stage 01
Human owns
All live ethnographic research (AI cannot read micro-expressions, tone, or cultural subtext); persona/JTBD validation against real interviews before anything is trusted as fact; ethical data curation; strategic differentiation calls
AI accelerates
High-volume ingestion and clustering that would take a human team weeks; first-draft persona/JTBD synthesis for human validation; competitive feature mapping at a breadth no human team matches in the same time
- Inputs
- Signed charter, support tickets, app store/NPS reviews, product analytics access, competitor product list, prior research repository
- Deliverables
- Thematic insight clusters (with outliers flagged, not discarded), draft personas + JTBD statements, competitor feature/interaction matrix, research repository entries
- Human checkpoint
- Every AI-drafted persona is validated against real research before use, never presented to a client as fact until validated. Ethnographic research is scoped and conducted by humans, full stop
- Gate to advance
- At least one persona/JTBD statement per primary user segment has been validated against real interview data (not just AI-synthesized); competitor matrix reviewed for strategic relevance, not just completeness
- Risks
- Treating AI-generated personas as ground truth without validation, the single most-cited 2026 failure mode in this space; AI cannot observe behavior or feel frustration, only synthesize what's already been said. Outlier-discarding if the agent isn't explicitly instructed to flag rather than drop anomalies
- Measures
- % of personas validated against real research before use; breadth of competitive coverage vs. time spent; ratio of AI-surfaced themes that survive human review unchanged vs. corrected
- Enterprise outputs
- Research repository, validated persona set, competitor landscape deck, the artifacts that ground every later stage's decisions and that clients expect to see in a discovery readout
03Formerly Synthesis / Problem FramingInsight Generation
Convert the Stage 02 fact base into a defined, ownable problem statement. Promoted to its own stage because problem framing is the highest-leverage human-judgment step in the process and was previously buried as a sub-bullet
+
Insight Generation
Convert the Stage 02 fact base into a defined, ownable problem statement. Promoted to its own stage because problem framing is the highest-leverage human-judgment step in the process and was previously buried as a sub-bullet
Human owns
Final authorship and ownership of the problem statement, non-delegable. Judging which candidate framing captures real user empathy vs. which merely reads well
AI accelerates
Generating a wide candidate set of problem framings so humans are choosing from options, not starting from a blank page; checking each candidate framing against the evidence base for support
- Inputs
- Validated personas/JTBD, thematic insight clusters, competitor matrix, signed charter's success metrics
- Deliverables
- Problem statement (human-authored), 3–5 candidate HMW statements with the selected one marked, evidence trace linking the statement back to Stage 02 data
- Human checkpoint
- The final problem statement must be authored, not merely selected and left verbatim, by a human. Verbatim adoption of an AI-drafted statement without editorial ownership is a checkpoint failure, not a shortcut
- Gate to advance
- Problem statement is specific enough to be falsifiable (a bad concept in Stage 04 should be visibly bad against it), and traceable to at least two independent evidence sources from Stage 02
- Risks
- Selecting the best-worded AI candidate instead of the best-fit one, named explicitly in the Foundation doc as this stage's complacency risk; false precision (a sharp-sounding statement that isn't actually falsifiable)
- Measures
- % of Stage 04 concepts that can be scored as clearly on/off problem statement (a vague statement produces ambiguous scoring); stakeholder recall of the problem statement weeks later (a sign it's memorable and owned, not just filed)
- Enterprise outputs
- The problem statement itself, this becomes the north-star artifact quoted in every subsequent stakeholder review, and the yardstick Stage 07 validation is measured against
04Formerly IdeationConcept Exploration
Generate a high volume of structural and conceptual directions against the Stage 03 problem statement, then narrow to a defensible shortlist
+
Concept Exploration
Generate a high volume of structural and conceptual directions against the Stage 03 problem statement, then narrow to a defensible shortlist
Human owns
Final concept selection, taste, brand alignment, emotional intuition are non-delegable; ethical veto power, especially in sensitive contexts (healthcare, financial hardship, etc.); making the genuine innovation leaps AI can only interpolate toward
AI accelerates
Producing volume and structural diversity beyond what a human team generates in the same time; applying frameworks consistently as ideation prompts rather than replacing human creative judgment
- Inputs
- Approved problem statement + HMW, validated personas, competitor matrix (for differentiation, not imitation)
- Deliverables
- Concept set (volume, sketched), shortlist with selection rationale, anti-solutions log (what was deliberately rejected and why)
- Human checkpoint
- Concept selection is a human decision recorded with rationale, not just a pick; any concept touching a sensitive context gets an explicit ethical review before shortlisting
- Gate to advance
- Shortlist has documented rationale traceable to the Stage 03 problem statement, and at least one anti-solution was considered and explicitly rejected (proof the exploration was genuinely divergent, not a single idea dressed multiple ways)
- Risks
- Mistaking volume for progress; shortlisting the most-generated pattern because it appeared often, not because it's right; AI's structural interpolation reads as "safe" and can quietly narrow the option space toward convention
- Measures
- Volume-to-signal ratio (# concepts shortlisted / # generated) tracked over time, a flat or declining ratio means the agent isn't learning the team's taste and prompting needs recalibration; time from problem statement to shortlisted concept
- Enterprise outputs
- Concept board with selection rationale, the artifact presented in a concept-review stakeholder session
05Formerly Information ArchitectureExperience Architecture
Structure the shortlisted concept into navigable, cognitively sound information hierarchy
+
Experience Architecture
Structure the shortlisted concept into navigable, cognitively sound information hierarchy
Human owns
Cognitive-load validation by feel, not just by metric; localization and cultural adaptation; the specific discipline of removing steps, AI tends to map existing logic faithfully rather than challenge it
AI accelerates
Fast first-draft structure generation from standard mental models; consistent labeling suggestions; quantifiable load simulation as a starting diagnostic, not a verdict
- Inputs
- Shortlisted concept, personas/JTBD, any existing taxonomy/content inventory
- Deliverables
- Sitemap/nav tree, taxonomy/labeling doc, cognitive-load simulation report, draft microcopy
- Human checkpoint
- A structure that scores well on AI's click-count metric still gets a human "does this actually feel simple" pass before sign-off
- Gate to advance
- At least one full simplification pass has removed structure the AI proposed as-is (evidence the human step wasn't skipped); taxonomy reviewed for the target market's language and cultural context, not just English-default output
- Risks
- A structure that's technically efficient (few clicks) but cognitively unintuitive; AI reinforcing existing complexity because it maps reality rather than questioning it; taxonomy that reads correctly in the source language but breaks in localization
- Measures
- Clicks-to-goal for top user tasks; rework cycles between IA and later stages; % of AI-proposed structure that survives the human simplification pass unchanged (a high % may mean the pass isn't rigorous, not that the AI got it right)
- Enterprise outputs
- Approved sitemap and taxonomy doc, referenced directly in the dev handoff package at Stage 08
06Formerly Interaction & Visual DesignExperience Crafting
Apply behavior, motion, and visual language to the architected structure, the concept becomes a specific, branded, functioning experience
+
Experience Crafting
Apply behavior, motion, and visual language to the architected structure, the concept becomes a specific, branded, functioning experience
Human owns
Bespoke motion choreography, the specific arc/timing/easing that creates brand feel, which AI cannot originate; edge-case design, which requires empathy for user frustration AI doesn't have; cultural sensitivity review of generated imagery; final taste call on visual direction
AI accelerates
Mechanical application of design-system rules at scale and with near-100% consistency; rapid generation of styling variations for human curation; standard interaction pattern implementation, freeing human time for the bespoke moments that matter
- Inputs
- Approved IA, brand/design system, interaction requirements from the concept
- Deliverables
- Design-system-compliant high-fidelity screens, functional interactive prototype, documented edge-case states, motion spec for signature interactions
- Human checkpoint
- Every generated carousel/auto-advancing/timed interaction gets an accessibility check before it reaches Stage 07 testing, this is where several documented 2026 accessibility complaints originate. All imagery/iconography clears cultural-sensitivity review before client presentation
- Gate to advance
- Design-system token/component compliance verified (not manually eyeballed); every primary flow has a defined error, empty, and offline state, not just the happy path; at least one signature interaction has human-authored choreography, not AI-default timing
- Risks
- AI-default micro-interactions creating accessibility friction in contexts the agent wasn't told to consider; visual variation volume creating decision fatigue instead of clarity; cultural imagery risk in generated illustrative assets, especially for multi-market work
- Measures
- Design-system compliance rate of AI-generated screens (target near-100%); # of edge-case states defined per flow; accessibility issues caught at this stage vs. caught later at Stage 07 (earlier catch = cheaper fix)
- Enterprise outputs
- High-fidelity, interactive, design-system-compliant prototype, the artifact used in stakeholder demos and Stage 07 usability sessions
07Formerly Testing/PrototypingExperience Validation
Validate the crafted experience against real users and objective heuristics before build commitment
+
Experience Validation
Validate the crafted experience against real users and objective heuristics before build commitment
Human owns
Live moderation where reading body language, tone, and hesitation matters; root-cause interpretation of failure (the "why," not the "that"); final call on what gets fixed, balancing AI-reported heuristics against real constraints
AI accelerates
Mechanical, exhaustive, and fast heuristic/accessibility scanning; pre-interview volume via AI-moderated or synthetic sessions to focus scarce human moderation time on the sessions that need it most
- Inputs
- Interactive prototype, defined edge-case states, problem statement (as the validation yardstick)
- Deliverables
- WCAG audit report, heuristic findings log, usability session synthesis (human-moderated + AI-moderated combined), prioritized fix list
- Human checkpoint
- Any high-stakes test (regulated-industry flows, first-time-user critical paths) is human-moderated, not delegated to synthetic/AI-moderated sessions alone. Findings prioritization is a human call, explicitly weighing business/technical tradeoffs the agent doesn't have visibility into
- Gate to advance
- Zero unaddressed WCAG failures at the level committed to in the charter; every "user failed task" finding has a documented root-cause interpretation, not just a fail count; fix-priority list has explicit human sign-off
- Risks
- Over-trusting synthetic user testing as a proxy for real user reaction, explicitly flagged in current 2026 UX research practice as the most common misuse; treating a clean heuristic scorecard as proof of a good experience rather than proof of absence of *known* mechanical issues
- Measures
- % of accessibility/heuristic issues caught pre-human-testing (should trend toward "all the mechanical ones"); task success rate and time-on-task from human-moderated sessions; ratio of AI-flagged vs. human-discovered findings (a healthy practice finds real issues in both lanes, not just one)
- Enterprise outputs
- Usability test report and remediation plan, often the artifact a regulated-industry client's compliance/legal stakeholders review directly
08Formerly Handoff/DeliveryBuild Enablement
Translate the validated design into developer-ready code, specs, and assets without losing design intent
+
Build Enablement
Translate the validated design into developer-ready code, specs, and assets without losing design intent
Human owns
Pixel/motion-nuance QA, AI translation reliably misses spacing and timing nuance a trained eye catches; technical negotiation between design intent and backend reality; security/privacy review, applied with *more* scrutiny to AI-generated code, not less
AI accelerates
Fast, consistent code generation and spec-tagging at a volume that would otherwise consume significant engineering time; asset optimization and multi-resolution export
- Inputs
- Validated prototype, remediation-complete design, design system
- Deliverables
- Semantic code (HTML/CSS, SwiftUI, Jetpack Compose as applicable), tagged design spec, optimized asset package, component unit tests
- Human checkpoint
- Every generated-code artifact passes the same security/privacy review as human-written code, explicitly never skipped because "the AI wrote it." Visual QA against design intent before dev merge
- Gate to advance
- Security/privacy review signed off; visual QA confirms spacing/motion/timing match design intent within defined tolerance; any design-backend contradiction has a documented resolution, not a silent workaround
- Risks
- Complacency, the single named risk for this stage in the Framework: reviewers who skip scrutiny of code specifically because it looks clean and AI-generated code often does look clean; silent design-backend contradictions that get "resolved" by an undocumented workaround instead of a real negotiation
- Measures
- Design-QA rework cycles between generated code and final dev handoff; time from validated design to dev-ready spec package; security review findings rate on AI-generated vs. human-written code (should converge, not diverge, over time)
- Enterprise outputs
- Dev-ready spec + code + asset package, the artifact that closes the design phase of the SOW and opens the engineering handoff
09Formerly AnalyticsContinuous Learning
Monitor the shipped experience continuously and feed findings back into Stage 01 of the next cycle. Decoupled from Stage 07 because this is an always-on loop, not a one-time gate
+
Continuous Learning
Monitor the shipped experience continuously and feed findings back into Stage 01 of the next cycle. Decoupled from Stage 07 because this is an always-on loop, not a one-time gate
Human owns
Root-cause investigation, an anomaly may be an external event, not a design flaw, and only a human has the context to tell the difference; the Engagement Trap guardrail, explicitly preventing AI from optimizing toward addictive patterns just because the data says they "work"; final shipping decisions, balancing data against brand promise and long-term trust
AI accelerates
Continuous, tireless monitoring at a scale no human team sustains manually; statistical rigor on significance testing; first-pass anomaly flagging that focuses human attention
- Inputs
- Live product telemetry, A/B test configuration, shipped design's success metrics (from the original Stage 01 charter)
- Deliverables
- A/B test results with significance reporting, anomaly log with root-cause notes, quarterly (or cadence-appropriate) learnings digest feeding the next Stage 01
- Human checkpoint
- No variant ships to 100% without a human decision. Every flagged anomaly gets a human-authored root-cause note before it's closed, not just an "acknowledged" status
- Gate to advance
- Risks
- The Engagement Trap, optimizing for a metric that technically improves while harming user trust or well-being (e.g., infinite scroll, dark-pattern-adjacent nudges); anomaly-flagging blind spots, the agent flags what it's tuned to notice, "no anomaly flagged" is not the same as "nothing to review."
- Measures
- Time from anomaly detection to human root-cause triage; % of quarterly learnings that actually get written into the next Stage 01 charter (the real test of whether this is a loop or a report); shipped-variant business outcome vs. predicted outcome (calibration check on the whole framework)
- Enterprise outputs
- Live performance dashboard + quarterly learnings digest, the artifact that proves the Framework compounds over cycles instead of resetting to zero each project