The confidentiality drill: the gate that outranks craft

The Samsung story is the pattern to memorize: in spring 2023, Samsung permitted ChatGPT internally, and within about twenty days employees had pasted confidential source code (twice) and an internal meeting transcript into it - by May, generative AI was banned on company devices. Permitted tool, helpful intent, three leaks in twenty days. Nobody was malicious; everybody was productive. That is exactly why the rule has to live in your hands, not in a policy document.

The subtle row in the taxonomy is people. Everyone knows not to paste "Elena Voss"; fewer notice that "the CRO of our financial-services prospect, who joined from a payments company last year" is the same disclosure wearing a trench coat. Identification is a property of the COMBINATION, not the tokens: role + industry + timing can single out one human as surely as a surname. The working test: search your anonymized sentence in your head - how many real people could it describe? If the answer is "one," it is a name. This is the same logic privacy engineers call re-identification, and consultants get no exemption from arithmetic.

The move: replace identity, keep the analytical shape. "Meridian Bank's $4.2M contract" becomes "a major financial-services prospect in late-stage talks (contract in the low millions)" - the size, stage, and stakes your pricing analysis needs all survive; the identity does not. Then run the two tests: could a rival identify the party from your anonymized version? Could your analysis still run on what remains? Fail the first, you leaked; fail the second, you over-scrubbed.

Role-based references: "the client's CRO" becomes "a senior commercial stakeholder" when even the role identifies - and at a company with one CRO, it always does. Numbers: round or band them when the exact figure is itself identifying; a $4.2M contract is findable, "low millions" is not, and your margin math almost never needed the second decimal anyway.

The second layer is architectural: enterprise-tier tools. The clean rule from the course map: consumer tiers of the major vendors may train on your conversations unless you opt out; commercial, enterprise, and API tiers do not train by default - a contractual guarantee, not a vibe. And then anonymize anyway: defense in depth, because terms change, tabs get mixed up, and screenshots outlive settings pages.

Individual reflexes decay under deadline pressure; shared artifacts do not. The team version of this session is a forbidden-list per engagement, written at kickoff, taped to the wall (or pinned in the channel): every client name, person, code name, contract figure, and internal artifact that must never enter a prompt - with the approved anonymization alongside each. Two virtues: writing it forces the team to decide what is sensitive BEFORE anyone is in a hurry, and checking a draft against a list is mechanical enough to do at 11pm, which is when leaks happen. The lab's gate is exactly this - a forbidden-list with a scorer attached. Steal the pattern.

The four steps, in order: (1) stop - no more material into that session, no cleanup prompts, no "please forget that"; (2) tell the engagement lead - immediately, before the awkwardness compounds; (3) assess what tier and tool received it - enterprise with a no-training contract, or a consumer tier that may train on it; (4) document - what entered, when, which tool, which tier, who has been told.

One honest paragraph, because severity differs and pretending otherwise breeds hiding: a leak into an enterprise no-training tool is a policy breach to log - the contractual layer held, the behavioral layer failed, fix the habit. A leak into a consumer tool is an incident to escalate - the material may now sit in training pipelines you cannot recall. Different severity, same reporting duty. A team that punishes the report gets fewer reports, not fewer leaks - and the leak the lead never hears about is the one that ends up in the news.