AI Product Operating Model
Design a company-wide operating model for AI in product when the question is bigger than one prompt, one feature, or one model choice.
This skill is for organizations that need to align:
- product AI feature design
- agent and tool-using workflow posture
- data boundaries across product, context, and analytics
- provider and deployment posture
- public-cloud handling of PI or sensitive data
- evaluation, rollout, and governance
Default posture for May 2026:
- external model APIs first
- central platform ownership before federated sprawl
- explicit data-plane separation
- provider abstraction instead of hard vendor lock-in
- risk-tiered controls for sensitive data
- evals, telemetry, and rollback before broad rollout
- model-lifecycle governance: provider model deprecations treated as planned product events, not surprises
- multi-model ops: running more than one frontier provider in production as a standing operating concern
ASCII Flow
company AI ambition
|
v
scope classification
product AI | internal copilots | agents | mixed platform
|
v
operating model
central platform + provider posture + data-plane boundaries + risk tiers
|
v
shared controls
evals + telemetry + rollout + privacy + security + ownership contracts
|
v
roadmap
sequenced workstreams with owners, gates, and adoption feedback
When to Use This Skill
Use this skill when the user asks for:
- an AI in product operating model
- a company AI platform strategy
- an LLM or agent governance model
- how product teams should work with central AI/data/platform teams
- how to handle PI or sensitive data with public-cloud model providers
- how to split analytics, runtime context, and operational truth
- how to sequence workstreams for AI platform foundations
Use Other Skills for Depth
- LLM architecture, provider choice, adaptation, eval design -> ../ai-llm/SKILL.md
- Agent architecture, MCP vs A2A, approval patterns -> ../ai-agents/SKILL.md
- Product integration, streaming UX, structured outputs -> ../software-ai-integration/SKILL.md
- App context layer, memory, retrieval, grounding ->
ai-context-layer
- Analytics semantics, marts, metric governance -> ../data-analytics-engineering/SKILL.md
- Product instrumentation and AI/agent telemetry ->
marketing-product-analytics
- Production controls, privacy, incidents, auditability -> ../ai-mlops/SKILL.md
- AppSec and application-layer security boundaries -> ../software-security-appsec/SKILL.md
- Platform infra, workload identity, CI/CD, policy-as-code -> ../ops-devops-platform/SKILL.md
- Agent eval harnesses and regression gates -> ../qa-agent-testing/SKILL.md
Quick Reference
| Need | Default move | Where to go next |
|---|
| Company wants "AI in product" direction | Start with central platform ownership + external APIs first | assets/company-ai-operating-model-template.md |
| Teams mix product AI, internal copilots, and agents | Split surfaces by workflow shape and risk tier | references/data-boundaries-and-risk-tiers.md |
| Data is used for both analytics and runtime AI | Separate operational truth, AI context, and analytics planes | ai-context-layer, ../data-analytics-engineering/SKILL.md |
| PI or sensitive data must go through public cloud | Apply minimization, redaction, tenant isolation, audit, and provider allowlists | references/data-boundaries-and-risk-tiers.md |
| Teams want agents everywhere | Default to non-agentic product flows until agent value is explicit | ../ai-agents/SKILL.md |
| Platform needs rollout control | Standardize evals, telemetry, canaries, and rollback before scale | ../ai-mlops/SKILL.md, ../qa-agent-testing/SKILL.md |
| Provider announces model deprecation | Treat as planned product event: migration runbook, eval-parity gate, cutover comms | Section 6 — Model-Lifecycle Governance |
| Running multiple frontier providers in production | Define routing policy, fallback chain, eval parity, and per-provider cost attribution at platform level | Section 7 — Multi-Model Ops |
| EU AI Act applicability check | Classify features against high-risk categories; assign compliance owner and verify phase-in date | Section 8 — EU AI Act Compliance Checkpoint |
| Leadership wants to report "AI adoption" | Separate usage/activation metrics from downstream cycle-time, cost, or quality outcomes with a pre-AI baseline | Section 9 — Adoption Signal vs Adoption Theater |
| Usage or approval-rate metrics look strong but value is unclear | Check for survivorship bias, automation bias, and Goodhart risk before trusting the number | Section 10 — Measurement Pitfalls Beyond the Eval Suite |
Default Workflow
- Classify the company scope: product AI only, internal copilots only, or mixed platform.
- Classify the current posture: greenfield, retrofit, or fragmented existing systems.
- Choose the org model: central platform first, hybrid, or federated.
- Separate the data planes:
- operational product truth
- AI context layer
- analytics and semantic metrics
- Choose the model posture:
- external APIs first
- managed private endpoints
- mixed by risk tier
- Define shared platform contracts:
ModelRequest
ModelResponse
ContextBundle
SafetyDecision
ToolInvocation
EvalRun
AuditEvent
- Define risk tiers and minimum controls for each tier before feature work spreads.
- Define rollout gates: split eval function into two distinct roles — capability evals (does the new model/prompt meet the quality bar?) owned by the platform/ML team on a per-change cadence, and regression evals (did we break anything already shipped?) owned by each feature team on every deployment. Different owners, different cadence, different suite.
- Define model-lifecycle governance: deprecation runbook, eval-parity gate, cutover comms, and rollback plan before the first provider model ID is hardcoded in production config.
- Assess EU AI Act applicability: classify features against high-risk categories (Annex III); assign compliance owner and applicable obligation date for any qualifying feature. Verify dates at https://artificialintelligenceact.eu/implementation-timeline/.
- Name the downstream metric each "AI adoption" initiative is meant to move (cycle time, cost per unit, defect rate, conversion) and its pre-AI baseline before rollout — not after. No named metric means the work is an experiment, not a rollout.
- Sequence workstreams: foundation -> data/context -> product AI features -> governance and reliability.
Canonical Operating Model
1. Central Platform Before Federated Sprawl
- A central AI platform function owns provider policy, model routing, prompt/config versioning, eval standards, telemetry, and approval patterns.
- Product teams own feature UX, business workflows, instrumentation, and outcome metrics.
- High-impact decisions stay centralized: provider allowlist, risk-tier policy, shared contracts, audit posture, and incident escalation.
- Org-design evidence grade (hedged): The enabling-team / platform-team pattern (Skelton & Pais, Team Topologies) is widely cited but evidence of superior outcomes in AI platform contexts is moderate and contested (grade C). Treat it as a useful default heuristic, not a settled best practice. Adjust ownership structures to your actual org scale and delivery rhythm.
2. External APIs First
- Do not make self-hosting the default planning assumption.
- Optimize first for provider abstraction, policy control, data handling, and rollback.
- Reserve stronger isolation paths for higher-risk tiers instead of forcing every workflow into one expensive posture.
3. Three Explicit Data Planes
- Operational truth: users, orgs, billing, permissions, product state, and other live business facts.
- AI context layer: retrieved evidence, derived memory, reusable context bundles, and task-scoped grounding.
- Analytics layer: event taxonomy, identity resolution, semantic metrics, marts, and BI or NLQ consumption.
Do not collapse all three into one warehouse, one vector DB, or one prompt log.
4. Workflow Shape Before Tooling
Classify every AI surface as one of:
LLM feature: request/response or streaming generation inside product UX
Tool workflow: bounded tool use with explicit control flow
Agent system: longer-running or ambiguous work where autonomy has measurable value
Default to the simplest shape that can work.
5. Risk Tiers Before Provider Selection
Use the tier model in references/data-boundaries-and-risk-tiers.md before choosing model providers or deployment posture. Provider choice follows data and action sensitivity, not the other way around.
6. Model-Lifecycle Governance
Frontier model deprecations are a planned product event, not an infrastructure surprise. The platform team owns the deprecation runway, not individual feature teams.
- Deprecation triggers: provider sunset announcement, capability regression on your eval suite, or model ID retiring in provider docs.
- Migration runbook: (1) identify all features calling the deprecating model ID, (2) select the candidate replacement, (3) run capability evals against your production eval suite on the replacement, (4) run regression evals to confirm no shipped behavior breaks, (5) canary-roll the replacement behind a feature flag, (6) execute comms to dependent teams with the cutover date.
- Eval-parity gate: block cutover until the replacement clears your capability eval threshold and your regression eval suite. "Roughly equivalent on public benchmarks" is not a gate — your task-specific evals are.
- Comms: treat a model migration as a minor product release — changelog entry, dependent-team notification, rollback plan documented before cutover starts.
- Current model IDs are a volatile fact: provider model namespaces churn quarterly. Always verify active model IDs from the provider's models page before hardcoding in platform config.
7. Multi-Model Ops
Running more than one frontier provider or model in production is now a standing operating concern, not an edge case. Address it explicitly in the platform design.
- Routing policy: define at the platform level — which features route to which model, and why (cost, capability, latency, data-residency, risk tier). Do not let routing decisions leak into feature code.
- Fallback chain: every Tier-1 and Tier-2 feature must have an explicit fallback (a secondary provider or a graceful degraded response). Define the trigger condition (error rate, latency threshold, provider outage signal) and the fallback target before first production traffic.
- Eval parity across providers: before routing a feature to a secondary provider, establish that it meets your capability eval threshold on your task-specific suite — not just on public benchmarks.
- Cost attribution by provider and model: per-feature cost telemetry must be tagged by provider and model ID so spend is visible and routing decisions are accountable.
- Provider contract hygiene: each provider in the allowlist needs a reviewed DPA, a known data-residency posture, and a verified data-handling tier. Audit when provider terms change, not only at onboarding. Note that the EU AI Act's controller/processor classification may require DPA amendments for AI providers — verify this classification for each provider in your allowlist.
- A2A protocol (agent-to-agent delegation): the Agent2Agent (A2A) protocol is an inter-agent task-delegation standard that complements MCP. Conceptual separation: MCP defines tool and context boundaries (what an agent can access); A2A defines how one agent delegates a task to another agent (agent-to-agent handoff). A2A moved to the Linux Foundation and reached a stable v1.0 (signed Agent Cards, multi-language SDKs) with adoption from major cloud platforms; treat it as production-viable for cross-vendor agent handoff, not an experimental spec. The canonical spec now lives at https://a2a-protocol.org/latest/specification/ (the earlier
google.github.io/A2A URL is stale) — verify the current version and adoption footprint before committing to it as a routing backbone, since agent-interop standards are still consolidating.
8. EU AI Act Compliance Checkpoint
The EU AI Act phases obligations in on a staged timeline. Verify the exact applicable date for your system at the official timeline (https://artificialintelligenceact.eu/implementation-timeline/) before committing to a compliance roadmap.
Verified milestones as of 2026-05-17 (source: artificialintelligenceact.eu):
- 2025-02-02: Prohibited AI systems (Article 5) and AI literacy requirements began applying.
- 2025-08-02: GPAI (general-purpose AI) model obligations and governance/penalties framework began applying.
- 2026-08-02: Main body of the Act applies — high-risk AI system obligations begin for new systems placed on market from this date. Systems placed on market before this date must comply if they undergo significant design changes.
- 2027-08-02: GPAI models placed on market before 2025-08-02 must achieve full compliance.
- 2030-08-02: High-risk AI systems used by public authorities must comply.
Digital Omnibus on AI — deadline extensions (formally adopted as of mid-2026):
EU lawmakers reached political agreement on May 7, 2026 to extend compliance deadlines for specific system categories; the European Parliament endorsed the package on June 16, 2026 and the Council gave final sign-off on June 29, 2026, so treat these dates as adopted law pending official-journal publication, not as a pending proposal. Qualify all 2026-08-02 roadmap commitments accordingly:
- Annex III high-risk systems (biometrics, critical infrastructure, employment/recruitment, credit, public sector): new or substantially modified systems receive a 16-month extension — new effective deadline December 2, 2027 rather than August 2, 2026.
- Annex I AI safety components (AI embedded in regulated products such as medical devices, lifts, radio equipment): 12-month extension — new effective deadline August 2, 2028 rather than August 2, 2027.
- AI-generated content transparency (Article 50(2) marking/detection duties): extension to December 2, 2026 for synthetic-content systems already placed on the market before August 2, 2026; systems placed on the market from August 2, 2026 onward must still comply from that date — this is a transition rule, not a blanket delay.
Verify the official-journal publication date and any last-mile textual changes before finalizing a roadmap — the package enters into force on the third day after publication, and implementation guidance (including the Commission's June 10, 2026 Code of Practice on AI-generated content transparency) continues to evolve. Sources: Gibson Dunn (https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/), White & Case (https://www.whitecase.com/insight-alert/eu-agrees-digital-omnibus-deal-simplify-ai-rules).
Operating-model checkpoints to add to your planning cadence:
- Classify all AI features against the high-risk categories (Annex III) — a feature that was low-risk at scoping may cross a threshold at scale or through a design change.
- If any feature qualifies as GPAI or high-risk, assign a named compliance owner and include the applicable obligation date in the roadmap, noting whether it falls under the extended or original timeline.
- Do not treat EU AI Act compliance as a one-time gate: design changes, new data classes, and new action types can change the classification.
- Verify exact dates against the official source — this is a staged regulation with multiple applicable dates and the enforcement window for your specific system depends on classification and market-entry date.
9. Adoption Signal vs Adoption Theater
Executive pressure to "show AI adoption" creates a predictable failure mode: usage metrics rise while delivered value stays flat. Distinguish the two before reporting either to leadership.
- Adoption theater markers: seat counts or license activations reported as the primary success metric; mandatory-use policies for AI tools with no measurement of what changed downstream; hackathon or pilot output celebrated without a path to production; a rising "% of engineers with Claude/Copilot/ChatGPT access" metric substituting for a rising output or quality metric.
- Real-leverage markers: cycle-time or lead-time change on shipped work, defect-escape rate, support-deflection rate, or revenue/cost outcomes tied to specific AI-enabled workflows — each with a pre-AI baseline and a comparable post-AI measurement window.
- The self-report trap: developer or agent self-reported "time saved" and satisfaction surveys are a weak proxy for delivered value and are prone to social-desirability bias; they are a useful leading indicator, never the headline metric in a board or leadership update.
- The novelty-decay trap: adoption and satisfaction curves for a new AI tool are almost always highest in month one and decay over the following quarter as users hit the tool's real limits; a rollout report that only captures week-one sentiment overstates durable value. Re-measure at 90 days before declaring success.
- Operating-model implication: require every "AI adoption" initiative to name the downstream business metric it is meant to move (cycle time, cost per ticket, conversion, defect rate) and the baseline before rollout — not after. If no such metric exists, treat the initiative as an experiment, not a rollout, and gate spend accordingly.
10. Measurement Pitfalls Beyond the Eval Suite
Passing the eval suite is necessary but not sufficient evidence that an AI feature or platform investment is working. Watch for these measurement failure modes specifically:
- Survivorship bias in usage data: usage or satisfaction telemetry from users who kept using the feature says nothing about the (often larger) group who tried it once and silently reverted to the old workflow. Instrument abandonment and reversion, not only active usage.
- Local optimization, global regression: a team-level win (faster ticket resolution, higher self-reported productivity) that offloads verification, review, or rework cost onto another team or onto production incident rates is not a net win — attribute cost as well as benefit across the full workflow, not just the AI-touched step.
- Automation bias in human-in-the-loop metrics: "human approval rate" on AI-generated output is not a quality metric once approvers learn to trust the tool and stop reading carefully — pair it with periodic blind spot-checks (compare a sample of approved output against an independent quality rubric) rather than trusting the approval rate alone.
- Goodhart risk on any single north-star metric: once "AI-assisted PR count" or "agent task completion rate" becomes a target tied to performance review or team funding, it will be gamed (trivial PRs, loosely-scoped "completed" tasks). Pair every adoption or throughput metric with a paired quality or outcome metric before it becomes a target.
Required Shared Contracts
Every company-wide AI platform should standardize these contracts before scale:
ModelRequest: caller, task type, risk tier, input schema, allowed providers/models
ModelResponse: output payload, schema validation result, citations or evidence refs where relevant
ContextBundle: live facts, memory, retrieved evidence, tenant scope, freshness, provenance
SafetyDecision: allowed, redacted, blocked, or escalate-to-human with reason
ToolInvocation: principal, requested tool, approved scope, result, side-effect status
EvalRun: eval type (capability | regression), suite version, judged dimensions, score, status, regression notes, owner team
AuditEvent: trace ID, actor, feature, provider, data class, approval event, retention class
Without these contracts, teams will re-implement incompatible local wrappers and the platform becomes impossible to govern.
Minimum Platform Controls
- Model/provider calls go through an abstraction or gateway layer with traceable policy decisions.
- Sensitive fields are minimized or redacted before model submission.
- Logs and traces mask raw PI by default.
- Retrieval and tools enforce tenant isolation and action scoping.
- Mutating tools require stronger approval than read-only tools.
- Every production feature has cost, latency, refusal, and fallback telemetry.
- High-risk launches require offline evals and staged rollout.
Known Traps
- Central platform ownership that becomes a product-team ticket queue instead of a standards, contracts, and guardrails layer.
- "One AI gateway for everything" designs that erase risk-tier differences between low-risk summarization, tool-using workflows, and high-sensitivity agent actions.
- Reusing the analytics warehouse as the live context plane for product features with no freshness, tenancy, or provenance controls.
- Declaring sensitive-data safety based on provider marketing language instead of verifying the actual request path, logging path, and retention settings.
- Shipping tool-using or agentic experiences before evals, telemetry, rollback, and human-escalation paths exist.
- Building a nominal provider-abstraction layer that still leaks provider-specific prompts, JSON schemas, and tool semantics into every product surface.
- Treating a provider's model deprecation as an infrastructure surprise rather than a planned product event with a migration runbook.
- Conflating capability evals (is this new model good enough?) with regression evals (did we break what ships?): they have different owners, different suites, and different cadences — collapsing them into one "eval run" hides both regressions and capability gaps.
- Running multiple frontier providers in production without explicit routing policy, eval-parity gates across providers, and per-provider cost attribution.
- Reporting seat counts, license activations, or week-one satisfaction as "AI adoption" success without a downstream cycle-time, quality, or cost metric and a pre-AI baseline.
- Treating "human approved it" as a durable quality signal once approvers have habituated to trusting the tool's output.
Common Anti-Patterns
- Federating ownership before shared contracts, evals, and policy boundaries exist.
- Treating prompt files as the primary governance mechanism for company-wide AI behavior.
- Combining operational truth, long-term memory, and analytics into one undifferentiated storage layer.
- Measuring only model quality while ignoring refusal rate, latency, cost, user outcome, and rollback behavior.
- Letting each product team choose its own provider auth, safety posture, and retention policy.
- Treating "agent" as the default upgrade path for features that should stay bounded request/response or tool workflow systems.
- Reporting adoption (seats, activations, self-reported time saved) as the success metric instead of a downstream business or quality metric with a pre-AI baseline.
Output Shape
When this skill is the primary skill, the default deliverable should include:
- target org model
- target data-plane architecture
- provider and risk-tier posture
- shared platform contracts
- rollout workstreams
- acceptance gates and residual risks
Use assets/company-ai-operating-model-template.md as the starting scaffold.
Navigation
References
Assets
Scripts
python3 scripts/contract_validator.py --input contract.json — Validates typed shared-contract JSON against required field schemas
python3 scripts/risk_tier_classifier.py --input feature-spec.json — Reads a feature spec and emits Tier 0/1/2 classification with minimum controls
Related skills:
Fact-Checking
- Known bugs, regressions, framework/compiler/runtime footguns, and version-specific crash or workaround guidance must be verified against current primary web sources before being treated as current fact.
- Verify current provider capabilities, retention defaults, training defaults, data-processing terms, controller-vs-processor posture, RBAC boundaries, regional controls, and managed isolation claims before final recommendations.
- Verify legal effective dates, scope boundaries, and enterprise-key or data-residency limitations before presenting governance advice as current.
- Prefer primary sources for vendor features, pricing, privacy posture, and regulatory guidance.
- If web access is unavailable, separate stable architecture guidance from unverified vendor-specific claims.
Learnings Loop
Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).
After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.