The Operator's Library · No. 10
For founders who don't have a compliance team
AI Safety &
Governance.

For small teams. The non-bureaucratic version of governance — the disciplines that fit on one page, get followed in practice, and stop the headlines you don't want to read.

AI Safety & GovernanceContents & Introduction
Contents & Introduction

Governance, the small-team version.


Most AI governance writing is for enterprises with compliance teams. This volume is for the founders, ops leads, and CTO-of-three-people running real AI systems with no formal programme. The good news: a small team can have excellent governance — if it's the right kind.

Contents.

01Why small teams need governance moreNo legal team to catch you. The case for upstream discipline.p. 03
02Least privilege & PIIWhat the model can see; what it must not. Done in a day.p. 04
03Audit logs as a productIf it's not logged, it didn't happen. Build the record first.p. 05
04Human-in-the-loop as designApproval not as a feature — as the architecture.p. 06
§The one-page policy & colophonp. 07

Three principles.

  1. Governance is a design choice, not a deliverable. Build the right gates into the system; don't write a 40-page policy no one reads.
  2. Logs are the product. The single most useful governance artefact is "what did the system do, and why." Build it first; debate everything else after.
  3. Default to friction. Whenever in doubt about a gate, leave it in. Removing one later is easy; recovering from a bad action is expensive.

Governance is what lets a small team trust the system it built. Without it, every Friday is a quiet risk; with it, every Friday is a working week.

02Contents
Ch. 01 · Why small teams need governance moreAI Safety & Governance
01
Chapter One

Why small teams need governance more.


A large company that screws up has lawyers, PR, insurance. A small company that screws up has a Tuesday. The asymmetry is the reason small teams should be more, not less, deliberate — though their governance looks different.

What can go wrong.

  • Customer PII leaks through a prompt into a log.
  • An agent sends an email that misquotes a price; the customer sues.
  • A model is asked something it shouldn't answer; produces something it shouldn't say; screenshot circulates.
  • An automated decision turns out to discriminate; a complaint is filed.
  • A connector with too much scope is exploited; data goes places.

What "good" looks like.

You can answer five questions, by close of business, on demand:

  1. What systems can the AI touch?
  2. What did it do yesterday?
  3. Where does PII go?
  4. Who approved every consequential action?
  5. How do we stop the system in 60 seconds if we need to?

The small-team advantage.

You have something a Fortune 500 doesn't: every person involved fits in one meeting. Governance can be a 20-minute weekly review, a one-page policy, and a Slack channel — not a department.

The small-team trap.

The other side of the coin: there is no one whose job it is to ask "are we sure?" Pre-commit to that role for someone — yourself, the CTO, a contractor. The role gets harder to add the longer you wait.

The reframe

Governance for a small team is not "we built a compliance programme." It is "one person owns the answers to the five questions, and the system was designed so those answers are findable."

A small team's governance is small, fast, and real. A large team's is large, slow, and on paper. Yours can be better — and is, if you build it.

This week

Pick the person who owns the five answers. Add a recurring 20-minute weekly review to the calendar. The role isn't a job change; it's a remit. Without it, no one's it.

03Chapter 01
Ch. 02 · Least privilege & PIIAI Safety & Governance
02
Chapter Two

Least privilege & PII.


Two disciplines that sound boring and are. Both can be implemented in one focused day each. Both pay back for the life of the system. Both are the difference between a small AI deployment and a small AI incident.

Least privilege — the practice.

  1. List every connector the AI uses. (Vol. 07 Ch. 02.)
  2. Name the minimum scope each one needs to do its job. Document the rationale per scope.
  3. Use OAuth and service accounts wherever possible. Avoid raw API keys; avoid acting-as-human accounts.
  4. Set token TTLs aggressively (30 days max for production; shorter for high-risk).
  5. Quarterly review of permissions. Anything not used in 60 days → revoke.

PII — the practice.

  1. Inventory. List the PII types in your business. Names? Email addresses? Payment info? Health info? Some categories carry legal weight.
  2. Map flows. For each PII type, draw where it enters the AI system and where it leaves. Most surprises live in the "leaves" arrow.
  3. Minimise. If the AI doesn't need the customer's full address to draft the email, strip it before passing. The most secure data is data the system never sees.
  4. Redact logs. Most accidental PII exposure is in logs. Redact at write time, not after.
  5. Retention. Define how long each type lives. Delete on schedule.

Default policies.

  • No PII in prompts unless the task requires it.
  • No PII in eval golden sets (use realistic but synthetic data).
  • No PII shared with external services without a DPA.
  • Customer-facing AI outputs reviewed for inadvertent PII echo.
Watch out

"Just feed the whole customer record into the prompt" is the line that ends in a privacy review. Pull only what the task needs; the discipline of scoping is part of the system.

Region awareness.

Different regions have different rules — GDPR (EU), CCPA (California), HIPAA (US health), etc. You don't need to be a lawyer; you do need to know which regimes you're under. Talk to one once a year.

PII protection isn't a feature you add. It is a default the system inherits from the way you designed it.

A weekend project

Day 1: connector audit + scope reduction. Day 2: PII inventory + redaction pass. By Monday you have a more defensible system than 80% of companies your size.

04Chapter 02
Ch. 03 · Audit logs as a productAI Safety & Governance
03
Chapter Three

Audit logs as a product.


If something goes wrong, the audit log is the only record. If it isn't there — to you, to your auditor, or to a court — it didn't happen. Building it first is not paranoia; it is the simplest single decision that turns AI from "story I tell" into "thing I can defend."

What every log row needs.

  • Timestamp — ISO 8601, UTC.
  • Actor — which agent, which user identity propagated through.
  • Action — the named tool call or skill that ran.
  • Inputs — what was passed (redacted of PII).
  • Outputs — what came back (redacted of PII).
  • Result — success / failure / partial.
  • Reason / model output — the model's chosen path, where applicable.
  • Cost — tokens used.

Three rules.

  1. Append-only. Rows never edited or deleted. Use a store that enforces this if your scale needs it.
  2. Queryable. "Show me every action this agent took for customer X last month." Should take 30 seconds.
  3. Retained per policy. A retention schedule per category. Default: 1 year minimum; longer for regulated work.

The killer query.

Pick a real action that happened in production last week. Try to answer: "What did the system do? In what order? Why? What returned what? Who could have seen this happen?" If it takes more than 5 minutes, your logging needs work.

Logs as debugging

The same logs that satisfy auditors are the logs that let you debug at 2 a.m. when something fires that shouldn't have. Build them once; use them twice.

Privacy in logs — the tension.

Detailed logs help governance. They also concentrate PII. Resolve the tension by structured redaction: store PII fields as hashed IDs; resolve to the real value only when authorised. Logs stay useful; PII stays minimal.

An untouched log is a kept promise. A missing one is a story you'll regret telling.

If logs don't exist yet

Make them this week. A simple append-only table in your existing database, with the eight columns above, beats nothing by infinity. Fancy can come later.

05Chapter 03
Ch. 04 · Human-in-the-loop as designAI Safety & Governance
04
Chapter Four

Human-in-the-loop as design.


The single most effective governance move is to design the system so a human signs off on every consequential action — until you have evidence they don't need to. This is not a feature. It is the architecture.

PATTERN 01 · PRE-ACTION

Approve before fire.

Draft is shown to a human; human approves; action fires. Default for: customer-facing communications, money movement, public posts, contracts.

Approval surface shows: what's about to happen, the chosen rule, a one-tap revoke.

PATTERN 02 · POST-ACTION SAMPLE

Review 1 in N.

The system acts; a human reviews a sample after the fact. For: high-volume / low-risk work — classification, summarising, drafting internal notes.

Sampling rate calibrated to stakes. Lower it as evidence accumulates.

PATTERN 03 · EXCEPTION-ONLY

Flag the weird ones.

Most runs go through unattended; the model self-flags low-confidence or rule-conflict situations to a human. The mature shape, earned over time.

Define what "flag" means before going live. Tune flag rate from a known baseline.

PATTERN 04 · KILL SWITCH

One command to halt.

Independent of patterns 1–3. A single, known action that stops all AI activity. Drill it quarterly. If you've never tested it, you don't have one.

Document. Practice. Make sure two people on the team know how to use it.

— APPROVAL FATIGUE —

The failure mode of pattern 01.

40 drafts a day; reviewer rubber-stamps; bad one slips through. Now you have the worst of both worlds.

Cut to one gate per workflow, late. Pair with confidence flagging.

— RULE OF THUMB —

Default to friction.

When in doubt about whether to add a gate, add it. Removing it later is one PR. Adding one after an incident is a quarter.

"We don't need a gate here" is something you earn with data, not something you assume on day one.

Approval gates are not bureaucracy. They are the design pattern that makes AI shippable. Build them in from the start; remove them with evidence.

06Chapter 04
Reference · One-page policyAI Safety & Governance
Reference

The one-page AI policy.


A template you can adapt today, tape to the wall, and actually follow. Long policies don't get read; this one fits. Customise the placeholders to your team; revisit quarterly.

[Company] — AI policy · v1

1 · What we use AI for. [Draft customer-facing copy, summarise meetings, triage support tickets, draft internal documents, code assistance.] All other uses require a 1-on-1 with [the AI lead] before proceeding.

2 · What we don't use AI for. Final decisions on hiring, firing, pricing, or legal matters. Generation of content presented as not-AI when it is. Anything involving customer PII without an explicit data flow design.

3 · Approval. Anything sent to a customer, posted publicly, or paid is reviewed by a human first. Internal-only outputs may be sampled. Exception-only review allowed only after 10 clean supervised runs.

4 · Data. No customer PII in prompts unless the task requires it. No client data shared with third-party AI services without a DPA. PII in logs is redacted at write time. Retention: [12 months default; per-category exceptions listed].

5 · Tools. Only AI providers approved in [registry]. New providers added after [the AI lead]'s review of security, DPA, and pricing. MCP connectors scoped to least privilege; quarterly review.

6 · Audit. Every consequential action logged with timestamp, actor, action, inputs (redacted), outputs (redacted), result. Logs append-only, queryable, retained per policy.

7 · Incidents. If something goes wrong, post to [#ai-incidents] within 1 hour. Owner: [name]. Kill switch: [link to the documented procedure].

8 · Review. This policy is read by every new hire on day 1. Reviewed by the team quarterly. Revised based on incidents and changes in regulation.

Signed: [the AI lead], [the CEO]. Last updated: [date].

Colophon

AI Safety & Governance for Small Teams · The Operator's Library · No. 10. Next: Vol. 11 — The One-Person AI-Native Business.

Scope the access. Log the actions. Gate the consequential. Drill the kill switch.

— END · OPERATOR'S LIBRARY NO. 10

07One-page policy