Governance, risk, and the conversation you'll have with Legal

AI governance isn't a policy document — it's a set of decisions operationalized across procurement, security, legal, and HR. Here's what "doing the work" actually looks like in 2026.

The real risk categories

Skip the jargon. Five concrete risks to manage:

1. Data leakage. Employee pastes confidential info into a consumer AI tool; data trains a future model.
2. Hallucinated outputs acted on. Someone trusts an AI answer without verification; wrong decision ships.
3. Bias and discrimination. AI systems make decisions affecting people (hiring, lending, access) in biased ways.
4. Regulatory exposure. EU AI Act, sector-specific rules (healthcare, financial services), privacy laws. Non-compliance has real teeth.
5. Third-party dependency. A key workflow depends on a vendor's API; vendor changes, incident, or outage blocks work.

The acceptable-use policy you actually need

Most published AUPs are 20 pages of aspiration. The one that works is one page:

• What tools are approved for different data classifications.
• What never goes into AI (customer PII, regulated data, source code labeled confidential, unreleased financials).
• What requires review before acting on (anything customer-facing, anything with legal implications).
• Who to ask when something's gray.

Short, specific, memorable. Train on it. Revisit quarterly.

Data classification ≠ theoretical

Classify your data in three categories:

• Public: marketing copy, public filings. Any AI is fine.
• Internal: non-public but not sensitive. Approved AI tools only.
• Confidential: customer data, financial data, source code, strategic docs. Only enterprise AI with DPA + no-training guarantees.

If your organization doesn't know what category a dataset falls into, stop and answer that before AI governance matters.

Vendor evaluation

For every AI tool you adopt, check:

• Data handling. Where is data processed? Is it used for training? Can you opt out?
• DPA / BAA / SOC 2. Standard trust signals. Missing these = hard pass for regulated data.
• Sub-processors. What other vendors handle your data? Each is a new risk surface.
• Exit path. If you stop using them, is your data deleted? How long does it take?

Have legal + security review every new AI tool before broad rollout. Not one person — both.

The human-in-the-loop principle

For any use case where wrong AI output has real-world impact on customers or employees, the rule is:

• AI proposes, human disposes.
• The human is accountable for the output, not "the AI."
• Logs capture what AI suggested, what the human decided, why.

This isn't excessive process — it's the minimum legal and ethical bar.

Incident response for AI

When something goes wrong (wrong data sent, hallucination acted on, vendor breach):

1. Contain. Pause the affected tool or workflow.
2. Assess. What data was exposed? What decisions were made?
3. Notify. Regulatory reporting obligations under GDPR, state laws, etc. Your privacy team knows the deadlines.
4. Remediate. Fix the gap.
5. Post-mortem. Write up what happened; share internally.

AI incidents are new. Your org probably doesn't have playbooks for them. Build a template now, before you need it.

The conversation with Legal

Have legal review these three things before any AI rollout:

• Vendor contracts for DPA, warranties, indemnity.
• Employee-facing policy for AUP compliance.
• Customer-facing disclosure if AI is used in customer interactions (sometimes required by law).

Legal saying "yes" isn't a blessing. It's a checkpoint. Keep going from there.

The governance anti-pattern

The biggest failure mode: governance becomes a bottleneck that drives shadow IT. Engineers start using ChatGPT personal accounts because the approved path is too slow. This is worse than lax governance — it's invisible use of unapproved tools on company data.

The fix: make the approved path at least as easy as the unapproved one. If it's harder, governance has failed.