Módulo 7: Building a Responsible AI Framework

Governance light: realistic for small teams

Descripción

Frameworks come from corporates. Startups have 5-50 people. Governance enterprise in startup = bottleneck, abandonment, frustration.

Esta cápsula te enseña governance light: mínimo viable que funciona realmente en team chico.

Al terminar vas a poder:

  • Diseñar estructura de governance proporcionada a tu team size
  • Evitar enterprise overhead que mata startups
  • Escalar governance gradualmente con company growth

El problema con enterprise governance

Standard governance models asumen:

  • 200+ employees
  • Multiple roles (CISO, Chief AI Officer, dedicated compliance team)
  • Months-long approval cycles
  • Multiple committees

Para startup: literally imposible. Hasta dañino:

  • Slows shipping critical
  • Creates "compliance theater" (process, no substance)
  • Demotivates team
  • Misses point: governance debe enable, no block

Tres principios de governance light

Principio 1: Concentration of responsibility

En team de 5-15, una persona asume "AI Steward" role. Generalmente Tech Lead.

Sus responsabilidades:

  • Owns the Responsible AI Framework
  • Triggers review processes
  • Final sign-off on ethics decisions (with escalation if uncertain)
  • Liaison con DPO/legal counsel

Beneficio: clarity de quién decide. Risk: bottleneck si overloaded. Mitigación: escalation paths claros.

Principio 2: Embed in existing processes

Don't create new meetings. Add ethics to existing processes:

  • Sprint planning → reserve 1 hour quarterly para ethics review
  • Code review → include ethics checklist verification
  • Release planning → ethics review is a release gate
  • Post-mortems → include ethics implications question

Beneficio: no extra calendar burden.

Principio 3: Documentation as code

Reviews, ADRs, risk acceptances viven en Git alongside code. Cambios via PR. Visible. Versioned.

Beneficio: developers ya saben use Git. No new tool to learn.


Implementación: 3-tier model

Tier 1: Solo Tech Lead (team < 10 personas)

Tech Lead = AI Steward (everything)
├── Drives framework
├── Reviews checklists
├── Signs off
└── Escalates to founder/CEO if material decision

Pro: minimal overhead. Con: single point of failure.

Tier 2: Tech Lead + Junior steward (team 10-30)

Tech Lead = AI Steward
├── Strategic ownership
└── Final sign-off

Senior Engineer = Co-steward
├── Day-to-day execution
├── Coordinates reviews
└── Escalates to Tech Lead

Pro: distributed work. Con: still concentrates power.

Tier 3: Light committee (team 30-100)

AI Ethics Council (3-5 people, meets quarterly)
├── Tech Lead (chair)
├── Senior ML Engineer
├── DPO or Legal Counsel
├── Product Lead
└── External advisor (optional, for diverse perspective)

Day-to-day: AI Steward (Tech Lead)
Strategic: Council quarterly

Pro: more perspectives, diverse input. Con: scheduling overhead (manage with discipline).


Approval flows: keep simple

For different decisions, different approval levels:

Decision typeApproval neededTime
Bug fix (no ethical implications)None (PR review only)minutes
New AI featureTech Lead review1 day
Significant model changeTech Lead + checklist1 day
New use case (existing AI)Tech Lead + ethics review1-3 days
New AI systemCouncil review (or equivalent)1 week
Significant compliance changeDPO + Legal + Tech Lead1-2 weeks
Strategic ethics directionCEO/founderas needed

Flow es proportional to risk.


Escalation paths

Cuando algo es ambiguous:

Engineer
  ↓ (concern)
Tech Lead / AI Steward
  ↓ (if not resolved)
CTO or CEO
  ↓ (if external implications)
Board / legal counsel

Key: each level has 24-48 hours to respond. Si lleva 2 weeks, broken.


Meeting cadence (realistic)

MeetingFrequencyDurationAudiencePurpose
AI Steward office hoursWeekly30 minEngineersAd hoc questions
Sprint review (ethics agenda item)Sprint cadence5-10 minTeamQuick ethics check on sprint
Quarterly governance reviewQuarterly2 hoursCouncil or stakeholdersComprehensive review
Annual deep auditAnnually1-2 daysAllFull M8 audit

Para startup de 10 personas: only weekly + quarterly are needed.


Training y onboarding

For new hires:

Onboarding (week 1):

  • 1-hour intro: ethics framework
  • Read AI Use Policy
  • Awareness of escalation paths

Quarterly refresher (30 min):

  • Updates to framework
  • Discussion of recent decisions
  • Q&A

Ad-hoc: when significant changes (new regulation, new feature with implications)


Trampas comunes

Trampa 1 — Coping enterprise structure. "Microsoft has Chief AI Officer; we need one too." For 10-person team, no. Tech Lead suffices.

Trampa 2 — Process for process's sake. "Let's have monthly ethics meetings." Without agenda, becomes ceremony nobody attends.

Trampa 3 — Hiring "Chief Ethics Officer" too early. Expensive, role unclear in small team, often becomes blocker rather than enabler.

Trampa 4 — Governance que slow everything. Each PR requires ethics review meeting? No. Code reviews check ethics during normal PR flow.

Trampa 5 — Sin escalation paths claros. Engineer encounters issue, doesn't know who to ask. Reports nothing.

Trampa 6 — Owner overload. AI Steward gets every question. Burns out. Distribute via co-stewards or training.


Cuándo escalar governance

Signals to expand governance structure:

  • Hit 30+ engineers → consider Tier 3 council
  • New regulated market (healthcare, finance) → dedicated DPO needed
  • Major incident → strengthen IR + governance
  • Multiple clients with audit requirements → consider ISO certification (M6)
  • Public regulator inquiry → formal structure needed

Ejercicio

Para tu organization:

  1. ¿Qué tier (1, 2, or 3)?
  2. Define AI Steward role + responsibilities
  3. Define approval flows for 5 typical decisions
  4. Define escalation path (3 levels)
  5. Meeting cadence (be realistic)
Ver solución (ejemplo team de 15)

Tier: 2 (between solo and committee)

AI Steward: Tech Lead (Mike)

  • Owns framework
  • Triggers reviews
  • Final signoff for routine decisions
  • Escalates significant decisions to CTO

Co-steward: Senior ML Engineer (TBD hire)

  • Conducts checklist reviews
  • Tracks risk register
  • Escalates to Tech Lead

Approval flows:

  • Bug fix: PR review (no special process)
  • New AI feature: Tech Lead review + checklist (Item-by-item)
  • New use case: Tech Lead + ethics review
  • New AI system: CTO + Tech Lead (treat as significant)
  • Compliance change: + Legal Counsel

Escalation path:

  1. Tech Lead (engineering)
  2. CTO (technical/strategic)
  3. CEO + Legal Counsel (external implications)

Meeting cadence:

  • AI Steward office hours: weekly 30 min
  • Sprint planning ethics check-in: 5 min in regular planning
  • Quarterly review: 2 hours
  • Annual ethics audit: 1 day

Resumen

Aprendiste:

  • ✅ 3 principles: concentration, embed, doc as code
  • ✅ 3-tier model según team size
  • ✅ Approval flows proportional to risk
  • ✅ Escalation paths con 24-48h SLAs
  • ✅ Meeting cadence realista (weekly + quarterly suficiente para small)
  • ✅ Trampas: enterprise copy, process theater, premature hires

Checkpoint: si tu governance puede ejecutarse sin matar shipping speed, estás listo.


Siguiente cápsula

08 — Proyecto: Responsible AI Framework completo consolida M7 entero en un Framework deliverable.


Recursos

  1. Microsoft Responsible AI Council.
  2. SaaStr — startup operations — for scaling lessons.
  3. Holacracy — alternative organizational model.
  4. GitLab handbook — example transparent governance.