Using AI at work: judgment calls
- judgment
- privacy
- verification
Using AI at work well comes down to a plan you write before you start: what you paste, what never, what you verify, what you disclose. Four lines, reused for every task. The plan is personal, but it answers to three masters: your employer’s rules, your readers’ trust, and your own name.
The four lines
- Paste — the classified-safe version: scrubbed data, public material, your own text. What not to paste set the buckets; the plan names them once so you stop re-deciding per task.
- Never — secrets, personal data, confidential figures, and untrusted content with a tool-enabled assistant (the prompt-injection boundary, drilled in the injection lesson).
- Verify — every number, name, quote and claim against its source before it appears in anything sent or published. Treat output as a draft with your name pending — the five-question check is the standing pass.
- Disclose — follow your workplace’s rules: some require labeling AI-assisted content, some restrict it for certain tasks entirely. When no policy exists, say so and set a local norm with your team — the absence of a rule is not a license; it is a gap you just found.
The meeting test
Before AI-assisted output goes into something consequential, three questions:
- Can I explain where each fact came from?
- Can I defend this text if someone questions it?
- Would I be comfortable if the whole process were visible?
Three yeses — ship it. Any no — fix that step first. The test is quick, and it encodes everything above in one pass.
Judgment calls that recur
- Tone about people: AI drafts, you decide — reviewing performance, saying no, delicate news. The draft can help; the choice is yours.
- Policy gray zones: ask early, not after. One message to your manager beats one embarrassed meeting.
- Your accountability: whatever the tool contributed, the signature is human. That is not a formality — it is why the other three lines exist.
Worked example
Task: a quarterly report section with customer examples. Paste: the outline, my own paragraphs, dashboard structure with fake sample values. Never: customer names and contract figures, raw exports, unreleased numbers. Verify: every figure against its dashboard; every example fictionalized or cleared; numbers re-typed, not copied. Disclose: change-log note that AI assisted the draft, plus a flag to the manager that no team policy exists yet.
Practice
Plan your usage before you start
You are about to use AI to draft a quarterly report section. It will include customer examples and figures from three internal dashboards. Your workplace has no explicit AI policy yet — and the section goes out under your name.
Write your usage plan before you start: what you will paste, what stays out, what you will verify, and what you will disclose. Then self-check it against the rubric and compare with a model answer.
A model answer (self-assessment — not machine-graded)
Plan: 1. Paste: the section outline, my own draft paragraphs, and the three dashboards' structure with fake sample values. Customer examples become 'Client A / Client B, typical usage ranges'. 2. Never paste: customer names, emails or contract values; raw dashboard exports; anything marked confidential; unreleased numbers. 3. Verify: every figure checked against its dashboard before it reaches the report; every customer example fictionalized or cleared for use; numbers re-typed by me into the final document, not copy-pasted from a draft. 4. Disclose: since no policy exists, I will note in the document's change log that AI assisted the draft, and flag the missing policy to my manager so we set a team rule. 5. Human checkpoint: I review the full section against the source dashboards line by line before it goes anywhere — the signature stays mine.
Hint
Four lines: paste / never / verify / disclose.
What does 'under your name' change about the final step?
Why this is the answer
A complete plan answers all four lines and keeps the checkpoint explicit: scrubbed inputs, named exclusions, figure-level verification, honest disclosure of both the assistance and the missing policy. The last line is the one people drop — and the one the signature depends on.
Transfer
- New job, new tool: rewrite the four lines for that context in ten minutes.
- Freelance or client work: version of the plan per client, agreed in writing.
- Teams: share your plan as a starting point — local norms beat silent variation.
Key takeaways
- Four lines before you start: what you paste, what never, what you verify, what you disclose.
- Verify every figure, name and claim against its source — the signature stays human.
- No workplace policy is not a license: ask early and set a local norm with your team.
Next
Next: Does ChatGPT train on your data? — the policy layer under these habits: what the controls actually control, dated and sourced.