The fifteen-second check before you hit send

Automated redaction finds formats. This is the pass that finds meaning, and it is the one that catches what the rules structurally cannot.

It takes about fifteen seconds on a normal message. The order below is deliberate: most-commonly-missed first.

1. Names of people

The most frequent identifier in ordinary business text and the least detectable. Clients, colleagues, candidates, the person whose ticket you are summarising.

Ask whether the answer changes if the name is replaced with "a customer". If it does not, replace it. Usually it does not.

2. Numbers that are not public yet

Revenue, growth rate, headcount, runway, pricing under negotiation, conversion figures. A model cannot tell a published number from an embargoed one, because they look identical.

The test is whether the figure has appeared in something external: a filing, a press release, a public page. If not, it is unreleased, whatever the internal norms around discussing it.

3. Project and product code names

Code names exist to be spoken safely in public, which is exactly what makes them undetectable. "Falcon" is a bird, a car, and a spacecraft.

If your organisation gave it a code name, someone considered it sensitive enough to need one. Replace it with a generic description.

4. Anything under an agreement

Contracts, terms under negotiation, legal matters, anything covered by an NDA. The obligation lives in a document, not in the text, so nothing about the words will trigger a rule.

This is also the category with consequences beyond embarrassment, since a confidentiality obligation can be breached by disclosure to a processor.

5. Other people's details

Everything above concerns your own organisation's information. This one covers people who never agreed to any of it: the customer whose transcript you are summarising, the colleague in the email thread, the candidate in the interview notes.

They are not in the room and will not be asked. That is a good reason to hold this category to a higher standard than your own data.

The order matters

If you only have time for one, do the first. Names appear in more messages than the other four combined and survive every automated pass.

If you have already sent something

Different categories need different responses, and only the first is urgent:

In every case the failure mode is silence. The exposure has already happened; what remains controllable is the response, and that only starts once somebody knows.

Where the tool fits

The rules handle the mechanical formats: keys, tokens, addresses, hostnames, the things you skim past at the end of a long day. This checklist handles the meaning, which no tool reaches.

Together they cover the ground. Either alone leaves a gap, and the gap left by trusting the tool completely is the more dangerous of the two, because it feels like safety.