── ── AI & agents

AI Risk & Compliance Baseline

Small teams rarely get hurt by exotic AI risk. They get hurt by basic gaps: client data pasted into a tool that may train on it, AI output reaching customers with nobody accountable for it, an agent that reads untrusted email and can also send money, vendor terms nobody read. None of these needs a governance department to prevent —…

How it works

Output: AI Use Register + Baseline Checklist

1. Inventory. For every AI tool/agent: purpose · users · input data types · where outputs go · whether it can take actions. Gate: unregistered "shadow AI" use exists → register it first; stop. 2. Walk the 7-item baseline (green/amber/red each): - Data privacy & consent: the purpose and consent for collecting personal data cover AI processing; handled under the applicable regime (e.g. GDPR where it applies). - Confidential data: client/trade secrets go only into tools that do not train on inputs and have data-processing terms; minimize and redact inputs. - Intellectual property: you have the right to use the inputs; the copyright status and permitted use of outputs are understood (check the relevant copyright authority's current position). - Customer disclosure & claims: disclose AI interaction or AI-generated content where the applicable rules require it (e.g. EU AI Act transparency obligations); do not overstate AI capability in marketing; treat what a customer-facing AI says as the company's own statement. - Security: untrusted content must not directly drive tool calls (prompt injection) · least privilege · no secrets in prompts · outputs are not executed directly (see OWASP Top 10 for LLM Applications). - Vendor terms: retention period, training use, sub-processors, data location, incident notification, deletion on exit — read for the tier you actually use. - Record keeping: use register, key outputs, review records and incident records are retrievable. Gate: any red item touching client data or external output → pause that use until it is amber. 3. Assign owner and deadline. For every red/amber item: remediation · owner · due date. 4. Set re-checks. Re-run immediately on a new tool, a vendor-terms change, or an incident; plus a fixed periodic review. 5. Route every figure. Any threshold, deadline, fine or applicability question that arises → record the governing instrument and "verify current text"; do not fill it in. Gate: a number or date appears in the register without a cited current source → remove it. 6. Stop-rule: if the AI handles only public information and nothing it produces leaves the company, compress to step 1 plus the security item.

Output template: Tool · Purpose · Data types · Output destination · Can act? · Privacy · Confidential · IP · Disclosure/claims · Security · Vendor terms · Records · Red/amber fixes + owner + due date · Next re-check

When to use it

• user says 'what are our compliance risks from using AI?', 'can we feed client data into AI?', 'do we have to tell customers we use AI?', 'who owns the copyright on AI output?', 'how do we stop prompt injection?', 'a client asked how we manage AI risk', or a small team is starting to use AI tools/agents in the business and needs a minimum baseline

Worked example

Moffatt v. Air Canada — The Chatbot That Was "a Separate Legal Entity" (2024)

A small-claims decision with a small award became one of the most cited AI-liability cases of 2024, because it answered a question every team deploying a customer-facing bot eventually faces: when the AI says something wrong, whose statement is it? The decision is Moffatt v. Air Canada, 2024 BCCRT 149, issued by the British Columbia Civil Resolution Tribunal (member Christopher C. Rivers) on February 14, 2024.

Related mental models