AI Model Cards UI in Figma

Quick answer

AI model cards and system documentation UI tell buyers and users what a model can do, what it was trained on, where it runs, and what can go wrong—before they rely on outputs in production. Design a customer-facing model card with version, intended use, limitations, and subprocessors; an in-product system panel with latency, context window, and safety filters; and an internal compliance record linked to DPIA, ROPA, and launch gates. Start from the Figma guides hub and pair with AI training opt-out, automated decisions, trust center, launch gates, and Dev Mode handoff.

Who this is for

Product designers shipping copilots, classifiers, or generative features where model behavior must be explainable to customers and regulators.
Trust and compliance teams preparing EU AI Act documentation, enterprise security questionnaires, and public transparency pages.
Design system teams standardizing model version badges, risk-tier chips, and “out of scope” warnings across editor, API docs, and admin consoles.

AI system documentation hub (internal overview)

AISystemDocsHub — Acme AI Platform · 14 production models · 3 high-risk systems
├── Header: Last card refresh Aug 10 · 2 models pending legal review · 1 blocked at launch gate
├── Actions: [ New model card ] [ Bulk export for trust center ] [ Compare versions ] [ Audit trail ]
├── Tabs: Model cards · System records · Risk tiers · Evaluations · Incidents · Changelog
├── Alert: acme-fraud-v3 · High-risk · Conformity assessment due Sep 2026 · Owner @risk-ai
├── Filters: Risk tier · Region · Status (draft/live/deprecated) · Product surface
└── Link: [ROPA](/designing-records-of-processing-activities-and-data-mapping-ui-in-figma/) · [DPIA](/designing-privacy-impact-assessment-and-dpia-workflow-ui-in-figma/) · Trust center · Launch gates

Section	Purpose
Model cards	Public or customer-facing summaries per model version
System records	Internal technical + compliance dossier
Risk tiers	Minimal / limited / high-risk classification with controls
Evaluations	Benchmarks, red-team results, bias checks
Incidents	Safety failures tied to model version
Changelog	What changed between v2.1 and v2.2

Verdict: Model cards fail when marketing copy replaces tested limits—document what the model refuses, not only what it excels at.

Customer-facing model card (trust center)

ModelCard — acme-gpt-4o-mini · v2.4.1 · Live since Jul 2026
├── Summary: General-purpose text generation for drafting, summarization, and Q&A inside Acme workspaces
├── Intended use: Productivity assistance · Not for medical, legal, or sole credit decisions
├── Capabilities:
│   ├── Context: 128k tokens · Languages: EN, DE, FR, ES
│   ├── Modalities: Text in/out · Image input (beta) · No audio generation
│   └── Tools: Function calling · Retrieval over tenant index only
├── Limitations (tested):
│   ├── May hallucinate citations · Do not use as authoritative source
│   ├── Weak on tabular math >50 rows · Route to dedicated analytics
│   └── Blocked topics: self-harm instructions, malware generation (safety filter)
├── Data & privacy:
│   ├── Training: Customer content NOT used (zero-retain workspace default) · See [AI opt-out](/designing-ai-training-opt-out-and-model-data-usage-transparency-ui-in-figma/)
│   ├── Logging: 30-day abuse logs · Region EU-West
│   └── Subprocessors: Azure OpenAI EU · [DPA listed](/designing-data-processing-agreements-and-subprocessor-management-ui-in-figma/)
├── Performance (p50): 1.2s first token · 99.2% availability SLA · [Status page](/designing-customer-incident-status-page-and-communication-ui-in-figma/)
├── Risk classification: Limited risk · Transparency obligations met · Not high-risk under Annex III for this deployment
├── Version history: [ v2.4.1 changelog ] · Deprecation: acme-gpt-4o-mini v2.3 ends Nov 2026
└── Contact: security@acme.com · [Report unsafe output]

Element	Requirement
Version pin	Every card tied to immutable model ID, not marketing name
Intended use + out-of-scope	Explicit negations reduce misuse
Tested limitations	From eval suite, not engineer intuition
Data posture	Link training, logging, region to live policy
Deprecation timeline	API customers need migration runway

In-product system panel (editor / API dashboard)

SystemPanel — Copilot · Model acme-gpt-4o-mini v2.4.1
├── Badge: Limited risk · EU region · Zero-retain
├── Runtime:
│   ├── Context used: 4,210 / 128,000 tokens
│   ├── Safety: Standard filter · PII redaction ON
│   └── Fallback: Disabled (enterprise policy)
├── What this model is for: Drafting and summarizing your workspace docs
├── Not for: Automated hiring decisions · See [automated decisions](/designing-automated-decision-making-and-profiling-transparency-ui-in-figma/)
├── Known issues (v2.4.1): Table extraction errors on scanned PDFs · Fix in v2.4.2
├── Switch model: Grayed · Admin locked to approved list
└── [ Full model card ] · [ Report output issue ] · Opens trust center anchor

Surface known issues from the same changelog feed as the public card—hiding regressions erodes trust.

High-risk system record (internal compliance)

SystemRecord — acme-fraud-v3 · High-risk · Annex III §5(b) creditworthiness adjunct
├── Classification rationale: Scores transaction risk affecting payment approval · Human review required
├── Conformity: Technical doc v1.2 · Risk management file RM-044 · Pending notified body review
├── Data inputs: Transaction metadata, device signals, account age · No protected attributes in feature set
├── Human oversight: Scores >0.85 route to analyst queue · Override logged · SLA 15 min
├── Monitoring: Weekly bias report · Drift alert on geography skew · Last eval Aug 8
├── Incidents: INC-992 false decline spike · Model rolled back to v3.1.4 · PIR linked
├── Links: [DPIA addendum](/designing-privacy-impact-assessment-and-dpia-workflow-ui-in-figma/) · [ROPA PA-218] · [Launch gate FEAT-FRAUD-09](/designing-privacy-by-design-launch-gates-and-feature-privacy-review-ui-in-figma/)
└── Public summary: Abbreviated card without sensitive thresholds · Legal approved

High-risk systems need two views: full internal dossier and redacted customer summary.

Model version changelog UI

ModelChangelog — acme-gpt-4o-mini
├── v2.4.1 · Aug 10 2026 · Patch
│   ├── Fixed: citation hallucination on long PDFs (+12% accuracy in eval- cite-01)
│   ├── Unchanged: safety filter thresholds · context window
│   └── Migration: None · Auto-deployed
├── v2.4.0 · Jul 15 2026 · Minor
│   ├── Added: Image input beta · Disabled for HIPAA workspaces
│   ├── Deprecated: Legacy JSON mode · Remove Jan 2027
│   └── Notify: In-app banner + email to API admins · Notice v4.0 stored in [consent ledger](/designing-consent-records-and-preference-management-admin-ui-in-figma/)
└── [ Subscribe to RSS ] · [ Webhook for model updates ](/designing-webhooks-and-event-subscriptions-ui-in-figma/)

Breaking changes require re-notice when outputs materially affect downstream automation.

Comparison: documentation depth by audience

Audience	Show	Hide
End user	Intended use, limits, privacy, report link	Weights, exact thresholds
Admin buyer	SLAs, subprocessors, risk tier, deprecation	Internal eval raw scores
Regulator / auditor	Full system record, human oversight, monitoring	Unrelated product roadmaps
Developer (API)	Version IDs, rate limits, migration guides	Legal classification memos

One card rarely fits all—use tabs or expandable sections instead of a wall of jargon.

Handoff checklist (Dev Mode)

ModelCard — model_id, version, risk_tier, regions[], intended_use[], limitations[], subprocessors[], published_at.
SystemRecord — system_id, annex_iii_category, conformity_status, oversight_mode, monitoring_cadence.
ModelChangelogEntry — version, change_type, breaking_flag, eval_refs[], notify_channels[].
InProductPanel — context_used, safety_profile, fallback_allowed, admin_locked_models[].
KnownIssue — model_version, issue_summary, workaround, fix_target_version.
Accessibility — risk badges include text labels; changelog tables responsive; report links keyboard reachable.

Common mistakes

Mistake	Why it hurts	Fix
Marketing-only model page	Misuse and regulatory gaps	Add tested limitations and out-of-scope
No version pinning	Incidents untraceable	Immutable model_id on every surface
Hiding deprecations	Broken API integrations	Changelog + webhook + banner
Same copy for high-risk and chat	Wrong controls applied	Separate templates by risk tier
Stale subprocessors	Trust center audit failures	Sync with DPA hub
No link to human oversight	EU AI Act gaps for high-risk	Document review queue in system record
Burying “not for X”	Harmful reliance	Lead with intended use boundaries
Eval results never updated	False confidence	Tie card refresh to eval pipeline

Recommended workflow

Classify each AI system (minimal / limited / high-risk) with Legal before public copy.
Draft model card from eval results and data policy—not adjectives.
Build in-product panel with version, limits, and report path.
Create internal system record for high-risk with oversight and monitoring links.
Wire changelog to trust center, API docs, and webhooks.
Gate releases at launch review until card approved.
Publish summary on trust center with ROPA cross-links.

FAQ

Difference from AI training opt-out UI?

Training opt-out controls whether user content improves models; model cards describe model behavior, limits, and compliance either way.

Need a card for every fine-tuned variant?

Yes for customer-visible models; internal-only experiments can stay in draft until promoted.

Link to automated decisions?

When the model output materially affects users (scoring, ranking, denial), cross-link both surfaces.

EU AI Act high-risk?

System record holds conformity artifacts; public card shows transparency items without leaking sensitive thresholds.

API-only models?

Same card content in developer portal; in-product panel optional; changelog via webhook essential.

Next steps

Design AI training opt-out and model data usage transparency UI in Figma — data posture section of every card
Design trust center and security documentation UI in Figma — publish approved model cards
Design privacy by design launch gates and feature privacy review UI in Figma — block release without documentation
Design automated decision-making and profiling transparency UI in Figma — when models affect user outcomes
Design records of processing activities and data mapping UI in Figma — map each model to processing activities

Designing AI model cards and system documentation UI in Figma: capabilities, limits, and compliance