Module 7: Building a Responsible AI Framework

Lightweight governance: realistic for small teams

Description

Frameworks come from corporations. Startups have 5-50 people. Enterprise governance at a startup = a bottleneck, abandonment, frustration.

This capsule teaches you lightweight governance: the minimum viable version that actually works on a small team.

By the end you'll be able to:

  • Design a governance structure proportional to your team size
  • Avoid the enterprise overhead that kills startups
  • Scale governance gradually as the company grows

The problem with enterprise governance

Standard governance models assume:

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

For a startup: literally impossible. Even harmful:

  • It slows down critical shipping
  • It creates "compliance theater" (process, no substance)
  • It demotivates the team
  • It misses the point: governance should enable, not block

Three principles of lightweight governance

Principle 1: Concentration of responsibility

On a team of 5-15, one person takes on the "AI Steward" role. Usually the Tech Lead.

Their responsibilities:

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

Benefit: clarity about who decides. Risk: a bottleneck if they're overloaded. Mitigation: clear escalation paths.

Principle 2: Embed it in existing processes

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

  • Sprint planning → reserve 1 hour quarterly for an ethics review
  • Code review → include verification of the ethics checklist
  • Release planning → the ethics review is a release gate
  • Post-mortems → include a question about the ethical implications

Benefit: no extra calendar burden.

Principle 3: Documentation as code

Reviews, ADRs, and risk acceptances live in Git alongside the code. Changes via PR. Visible. Versioned.

Benefit: developers already know how to use Git. No new tool to learn.


Implementation: the 3-tier model

Tier 1: Just the Tech Lead (team < 10 people)

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

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

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

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

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

Pro: distributed work. Con: it still concentrates power.

Tier 3: A 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
└── An external advisor (optional, for a diverse perspective)

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

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


Approval flows: keep them simple

Different decisions, different approval levels:

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

The flow is proportional to the risk.


Escalation paths

When something is ambiguous:

Engineer
  ↓ (concern)
Tech Lead / AI Steward
  ↓ (if unresolved)
CTO or CEO
  ↓ (if there are external implications)
Board / legal counsel

The key: each level has 24-48 hours to respond. If it takes 2 weeks, it's broken.


Meeting cadence (realistic)

MeetingFrequencyDurationAudiencePurpose
AI Steward office hoursWeekly30 minEngineersAd hoc questions
Sprint review (ethics agenda item)Sprint cadence5-10 minThe teamA quick ethics check on the sprint
Quarterly governance reviewQuarterly2 hoursThe Council or stakeholdersA comprehensive review
Annual deep auditAnnually1-2 daysEveryoneThe full M8 audit

For a 10-person startup: only the weekly + quarterly are needed.


Training and onboarding

For new hires:

Onboarding (week 1):

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

Quarterly refresher (30 min):

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

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


Common traps

Trap 1 — Copying the enterprise structure. "Microsoft has a Chief AI Officer; we need one too." For a 10-person team, no. The Tech Lead suffices.

Trap 2 — Process for process's sake. "Let's have monthly ethics meetings." With no agenda, it becomes a ceremony nobody attends.

Trap 3 — Hiring a "Chief Ethics Officer" too early. Expensive, the role is unclear on a small team, and it often becomes a blocker rather than an enabler.

Trap 4 — Governance that slows everything down. Does every PR require an ethics review meeting? No. Code reviews check ethics during the normal PR flow.

Trap 5 — No clear escalation paths. An engineer encounters an issue and doesn't know who to ask. They report nothing.

Trap 6 — Owner overload. The AI Steward gets every question. They burn out. Distribute via co-stewards or training.


When to scale up governance

Signals that it's time to expand the governance structure:

  • You hit 30+ engineers → consider a Tier 3 council
  • A new regulated market (healthcare, finance) → a dedicated DPO is needed
  • A major incident → strengthen IR + governance
  • Multiple clients with audit requirements → consider ISO certification (M6)
  • A public regulator inquiry → a formal structure is needed

Exercise

For your organization:

  1. Which tier (1, 2, or 3)?
  2. Define the AI Steward role + responsibilities
  3. Define the approval flows for 5 typical decisions
  4. Define the escalation path (3 levels)
  5. The meeting cadence (be realistic)
See the solution (example: a team of 15)

Tier: 2 (between solo and a committee)

AI Steward: the Tech Lead (Mike)

  • Owns the framework
  • Triggers the reviews
  • Final sign-off on routine decisions
  • Escalates significant decisions to the CTO

Co-steward: a Senior ML Engineer (TBD hire)

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

Approval flows:

  • Bug fix: PR review (no special process)
  • A new AI feature: Tech Lead review + the checklist (item by item)
  • A new use case: Tech Lead + an ethics review
  • A new AI system: CTO + Tech Lead (treat it as significant)
  • A 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 the regular planning
  • Quarterly review: 2 hours
  • Annual ethics audit: 1 day

Summary

You learned:

  • ✅ The 3 principles: concentration, embedding, documentation as code
  • ✅ The 3-tier model based on team size
  • ✅ Approval flows proportional to risk
  • ✅ Escalation paths with 24-48h SLAs
  • ✅ A realistic meeting cadence (weekly + quarterly is enough for a small team)
  • ✅ The traps: copying the enterprise model, process theater, premature hires

Checkpoint: if your governance can be executed without killing your shipping speed, you're ready.


Next capsule

08 — Project: The complete Responsible AI Framework consolidates all of M7 into a Framework deliverable.


Resources

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