Governance & responsible AI

RACI. Clear ownership: who approves new models, data sources, and customer-facing claims; escalation to exec risk committee for novel harms.

Policy library. Living documents for acceptable use, data classes, retention, human-review thresholds—version-controlled like code.

Inventory. Central registry of all models (vendor + fine-tuned), data flows, and DPIAs—no shadow fine-tunes.

Rhythm. Monthly risk review with metrics; quarterly tabletop exercises for incident response.

Enablement. Self-service checklist for product teams reduces ‘ticket to nowhere’ friction while keeping gates.

flowchart TB
  P[Policies] --> R[Risk reviews]
  R --> I[Inventory]
  I --> O[Ops audits]
            

Tiers. Classify use cases by materiality—customer-facing credit decisions demand deeper controls than internal email assist.

Documentation. Model cards + validation reports + monitoring plan + contingency procedures for each tier.

Independence. Validation team separate from builders for high-tier models; periodic third-party review for critical paths.

Monitoring. Drift, misuse, and outcome fairness metrics with board reporting.

Limitations. Honest residual risk statements with compensating controls, not checkbox ‘approved.’

Triage. Sev levels by impact; immediate disable feature flag or model route if ongoing damage possible.

Forensics. Immutable logs with trace ids; legal hold; PR+Legal joint comms review before external statements.

Customer care. Scripted responses, credits policy, regulatory notification timelines where required.

Root cause. Distinguish model regression vs data poisoning vs prompt injection vs deployment bug—different fixes.

Follow-up. Public postmortem summary when appropriate; update eval harness with new regression tests.

Definition. Agree with legal/compliance on protected attributes and fairness metrics relevant to jurisdiction—avoid hand-wavy ‘de-bias the AI.’

Slicing. Automated eval dashboards per demographic proxy slices where ethically usable; watch disparity in escalation or task success.

Mitigations. Retrieval filters (inclusive sources), calibration layers, or human review for sensitive cohorts.

Feedback. User appeals channel with SLA; corrections feed training governance.

Limits. Document where automation is inappropriate—some decisions remain human-only by policy.

Measurement. GPU-hour telemetry × region grid factors; separate training burst from steady inference burn.

Efficiency. Routing, caching, quantization, and right-sized models beat greenwashing offsets alone.

Procurement. Prefer low-carbon grids or PPAs for training clusters when brand promises exist.

Transparency. Publish methodology; avoid precise claims you cannot defend in audits.

Product. Offer ‘eco mode’ with slightly reduced quality if customers value sustainability KPIs.

Contracts. Customer agreements specify who owns prompts, outputs, derivative fine-tunes; default posture explicit.

Copyright. Avoid ingesting unlicensed scraped content where jurisdiction unsettled; maintain takedown process.

Outputs. Disclose limitations on enforceability; indemnity posture aligned with legal appetite.

Open weights. Track license chains; attribute and comply with share-alike terms.

Process. Legal review for high-risk ingestion pipelines—not post-hoc panic.

Guardrails. Block paste of secrets into unapproved tools; provide sanctioned alternative with logging.

Training. Quarterly security modules with realistic exfil scenarios—not punitive-only tone.

Detection. DLP signals on sensitive repos; compassionate enforcement with manager coaching first strike.

Productivity. Approved internal copilot should be better than shadow ChatGPT to win adoption.

Labor. Discuss impacts on roles openly with HR; reskilling budgets tied to automation wins.

Security. SOC2, pen test summaries, data processing agreement, subprocessors, incident history.

Resilience. SLA credits, multi-region posture, published uptime honest with exclusions.

Model lineage. Training data claims, eval transparency, known limitations documented.

Exit. Data export, model portability, notice period for breaking changes.

Scorecard. Quantitative rubric compared across vendors—avoids politics-only decisions.

TT L tiers. Hot logs days, warm aggregates months, cold archives years only if legally necessary—defaults aggressive trim.

Redaction. Structured PII scrub on ingest; optional no-content logs (metadata only) modes.

Regional. EU vs US retention schedules differ; automate geo-routed pipelines.

Erasure. Tie into subject-rights workflows; prove deletion including embeddings referencing PII.

Audit. Retention config in terraform; drift detection if someone widens TTL silently.

Scope. Legal defines which decisions require notices—credit, housing, employment varies by law.

Content. Provide principal reasons in plain language plus human appeal path; avoid raw logits as ‘explanation.’

Evidence. Tie decision to retrieved or structured inputs auditable later—no vibes.

Human. Override workflow with logging when agent disagrees with model suggestion.

Testing. Compliance simulations before launch in each regulated jurisdiction.

Regional cells. Inference + storage + logging co-located; control plane automation enforces routing pins.

Model allowlists. Some weights disallowed in regions—enforce in gateway, not tribal knowledge.

Transfers. SCCs, DPAs, impact assessments when analytics must aggregate globally—document residual risk.

Feature gaps. Product surfaces honest ‘unavailable here’ states; sales cannot promise parity.

Updates. Legal monitors regulatory diffs; engineering ships config templates not emergency hacks.

Contract. Liability caps aligned with realistic loss scenarios; carve-outs for gross negligence or IP infringement as counsel advises.

Cyber insurance. Disclose AI-assisted services to carrier; coverage for privacy and outage events.

Vendor pass-through. Map vendor indemnity for model IP claims vs your obligations to customers.

Disclaimers. UX copy matches legal stance—marketing cannot outrun contract.

Scenarios. Tabletop loss estimating with finance annually.

Scorecards. Few KPIs: incidents, near-misses, fairness disparities, spend trajectory, key model changes.

Context. Compare quarter-over-quarter; spotlight emerging regulations affecting roadmap.

Materiality. Escalate new use cases crossing risk tier thresholds before press learns.

Education. Short primers on limits of metrics—boards need intuition not false precision.

Cadence. Pre-read + focused Q&A; avoid 200-slide dumps monthly.

Evidence packets. Prebuilt exports: data maps, model inventory, access logs sample, change management records.

Reproducibility. Frozen configs for production models with checksums; who approved promotions logged.

Drills. Internal mock audits quarterly; time-to-produce answers measured.

Ownership. Named audit DRI per domain—no scramble spreadsheets.

Gaps. Track findings in risk register with remediation SLAs.

Risk tiering. Air-gapped open weights may be only option; cloud APIs forbidden—document formally.

Supply chain. Verify checksums, provenance, and license; block unknown download mirrors.

Maintenance. OS model patching is your problem—budget SRE + ML ops accordingly.

Red teams. Additional testing for unknown fine-tunes merged into community checkpoints.

Hybrid. Allow APIs for dev, weights for prod in defense sector patterns—consistent enforcement.