Keeping it safe

Data handling, IP, disclosure, and the failure modes that end careers.

Most AI safety failures at work are not exotic — they are a confidential document pasted into a personal account, a customer's details in a prompt, an AI draft that is secretly a competitor's copy. The safety competency is knowing the rules, the reasons behind them, and what to do in the first hour when something goes wrong.

1. Stage 1 · E1 FOUNDATIONAL

   Learn the one big rule

   Sensitive data goes only into approved tools. Customer names, account details, employee records, anything marked confidential: use the company-approved AI environment with appropriate data handling, or strip identifying details first, per your policy — or do not use AI on it at all.

   "Paste it and ask the AI to forget it afterward" is not a thing. Neither is using a personal account "so it is not linked to the company." If you do not know your company's policy, finding out is the first safety skill.

   Try it this week

   Find and read your organization's AI usage policy (or the closest data-handling policy). Write down which tools are approved for which data classes — from memory, then check yourself.

2. Stage 2 · E2 PROFICIENT

   Handle IP and disclosure like a professional

   When an AI draft comes back suspiciously close to a competitor's published copy, rework it in your own structure and voice and check it is not substantially copied — a low plagiarism-checker score is not a defense for structural imitation.

   Where a client's policy requires disclosure of AI use, disclose. The output quality is not the point; the relationship, and sometimes the contract, is. Rewriting 10% by hand does not change what the work is.

   Try it this week

   Draft your personal disclosure sentence — the honest, unembarrassed line you would use when a client asks how AI was involved in your work. Having it ready removes the temptation to dodge.

3. Stage 3 · E3 DISTINGUISHED

   Respond to incidents, not just rules

   When exposure happens — a teammate pasted a client document into a personal account and the summary is circulating — the first hour matters: stop further spread, inform the right owner (manager, security, privacy) promptly, and do not quietly fix it. Confidentiality obligations and disclosure duties, not personal discomfort, drive the response.

   Handle the person proportionately: assume ignorance, not malice, and fix the system gap (approved tools, training) that made the mistake easy. Cover-ups compound; candor contains.

   Try it this week

   Write the first three actions you would take, in order, for that exact scenario at your company — with the actual names or roles you would contact. Vague plans fail under adrenaline.

4. Stage 4 · E4 EXCEPTIONAL

   Anticipate the newer failure modes

   Content a model reads can carry instructions planted for the model to follow — a vendor document whose summary suddenly urges you to expedite payment deserves suspicion, not compliance. Treat content-borne instructions as untrusted, check the source yourself, and flag it.

   The exceptional operator also knows where AI should not decide alone: consequential judgments about people, legal exposure, anything irreversible. They can articulate the line and hold it under deadline pressure, which is when lines get crossed.

   Try it this week

   List the three decisions in your role where AI input is welcome but AI decision is not. For each, write the one-line reason — you will be asked someday, possibly by an exam.

Ready to see where you stand? The free check scores all six competencies in about fifteen minutes.