Systems & AdoptionNovember 7, 2026·5 min read

Your AI Policy Needs More Than 'Use ChatGPT Responsibly'

Five separate stories this fall all landed on the same gap: teams know they need an AI policy, but not what the document itself should actually say. Here's what to put in it.

By Patin Team · Examples are illustrative composites

If your company's AI policy is a sentence in the employee handbook telling people to "use AI responsibly," you already know it doesn't survive contact with a real question — which tool is actually approved, what happens when someone pastes the wrong thing into it, who signs off on the next one. The fix isn't deciding you need a policy. Every team already half-knows that. The fix is knowing what actually has to be written down.

Five stories, one missing document

This fall produced an unusual run of AI governance news, and almost none of it told anyone what to write — just that the gap existed. Sam Altman and Dario Amodei briefed the UN Security Council on treating AI failures as reportable incidents, which only works if a team can already tell a minor mistake from a serious one. DHS banned DeepSeek-based products from federal use the same week OpenAI and Anthropic rewrote their own usage rules, which matters only to a team that knows which model sits behind each tool it already uses. Altman and Hassabis both admitted, days apart, that AI security can't be patched into existence — only audited for, tool by tool. And one piece here already covered the question of who should own that decision. None of the five told anyone what the document itself needs to contain.

What the document actually needs

A real AI usage policy isn't a values statement. It's four specific answers, written down before someone asks:

  1. Which tools are approved, and which model actually runs behind each one. Not the vendor's marketing page — the real answer, confirmed with support or the contract. A DHS ban that reaches third-party tools nobody knew were built on the banned model is the reason this line exists now.
  2. A severity scale, not a vibe. What a minor AI mistake looks like in your team's actual work, versus one that gets escalated and logged. Without this line, "review before sending" is unenforceable advice.
  3. A named owner and a recheck date. Vendor terms change on their own schedule — a policy that was accurate in June can be wrong today without anyone on your side changing anything.
  4. An escalation trigger: the one question that moves something from "handled quietly" to "someone above me needs to know."

Renata: writing the document, not just deciding to

Renata runs people operations at a 70-person professional services firm. For two years, the firm's AI guidance was a line in the handbook: "use discretion." After the DHS order made her reconsider what was actually running behind the firm's resume-screening assistant, she spent one afternoon writing an actual policy — not a philosophy, four sections. Which three tools are approved and what model each runs on, confirmed by email from each vendor. A severity scale specific to hiring: a tool suggesting a slightly generic interview question is routine; a tool scoring a candidate lower for a protected characteristic gets logged and escalated same-day. Herself as owner, with a date on her calendar every quarter to recheck vendor terms. And one escalation line: anything touching a hiring decision for a named candidate gets a human sign-off before it's final. The document is two pages. It replaced two years of "use discretion."

Marcus: the smaller version that still counts

Marcus runs a 12-person accounting practice and assumed a real policy was overkill at his size — no compliance department, no legal team to review it. But his staff already use three different AI tools for draft client correspondence, and nobody had ever asked which ones train on input data. He didn't write a long document. He wrote one page: the two tools the practice actually approves, a one-line rule that any AI-drafted number or figure gets checked against the source file before it goes to a client, and his own name as the person who updates the page when a vendor's terms change. It took forty minutes. It's the first time anyone at the practice could answer, in writing, what AI is and isn't allowed to touch.

The one thing

A values statement about using AI "responsibly" answers nothing when a regulator, a client, or your own team asks a specific question. The document that holds up names the tools, scores the failures, assigns an owner, and sets a recheck date — before the next outside story forces the question instead of you.

Reading about it only gets you so far

Patin turns this into five-minute drills that score what you write and tell you why. It's in closed beta — join the waitlist and we'll email you when your cohort opens.

Just want the writing? .