Published: 2026-07-24
Analysis & perspective

Redact Sensitive Files Before AI Sees Them: The 'Start With the Job' Method

Chapters / key moments (click to jump — plays here on the page)

"Don't paste sensitive info into AI" is advice that stops too soon — the contract still needs a risk review, the plan still needs analysis. Nate B Jones walks through Airlock, a tool he built to strip PII and confidential data from a document before handing it to a frontier model, while keeping the context the task actually needs. The core discipline: start with the job, not the file, and decide what the model needs to see before you decide what to remove.

Source video

"I Deleted 5 Things From This File Before ChatGPT Saw It. It Still Found The Problem." by Nate B Jones — Watch on YouTube →

Key Takeaways

Sensitive work is still work. Warning people off AI just hands the problem back to them. The goal is to redact without destroying the context that makes the model useful.
Start with the job, not the file. Define what the model needs to answer your specific question (e.g. "what assumptions could break this plan?"), then remove everything that only came "along for the ride" — customer names, home addresses, private medical notes, API keys, unreleased prices.
Define protected terms up front. An internal codename like "Project Lantern" doesn't look private to a machine but is highly confidential. Context often lives only in your head, so you have to supply it to any redaction tool.
Default to hide; keep on purpose. Anything uncertain is hidden by default — you have to deliberately decide to keep it, not the other way around.
Rebuild, don't black-box. Word files quietly retain comments, track changes, author names, old edits, and external relationships. A safe approach outputs a fresh document from the approved material and leaves the original untouched on your machine — not black rectangles over the old file.
The same fact can be essential or irrelevant depending on the question. A negotiated price matters if you're evaluating pricing; it's noise if you only want the operational assumptions. Intent decides.
Why it's newly urgent: AI 100×'d the data we want to move, and sensitive info is intermixed with the useful context. Verizon telemetry cited in the video: corporate-device AI use rose 15%→45%, two-thirds via non-company accounts, with source code the most-submitted material. "Security fatigue" means the easy path wins — so privacy has to live in the same path as convenience.
Know when redaction is the wrong tool. If removing the sensitive data removes the task itself (e.g. a full medical history that is the analysis), the work belongs in a governed environment — or not with AI at all.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

1. Start with the job, not the file
Decide what you are asking the model to do and what minimum context makes the answer useful. The same detail can be essential for one question and irrelevant for another: in his demo a price was dropped because the question was "which assumptions could break this operating plan", but would stay if the model were asked to evaluate the pricing.
2. Define protected terms
In his tool, Airlock, you first enter customer names, project names and internal code names, because a string like "Project Lantern" does not look private to software. The tool then combines its own pattern checks with your list on one screen.
3. Default to hiding
If there is any uncertainty the item is hidden; keeping it is a deliberate choice. In the demo he removed a customer name, a home address, an email, a private medical note, an API key and an unreleased price, and kept the facts the model needed (three warehouses move in the same week, an ERP integration must be ready before the first billing cycle, training happens during normal shifts).
4. Rebuild, don't blank out
It writes a new Word document rather than drawing black boxes over the old one. Word files can carry comments, tracked changes, author names, old edits and external relationships that survive a clean-looking page. The original stays untouched on your computer; you read the smaller copy and send only that to a frontier model.
5. Ask the model the real question
He asked the model which assumptions were most likely to break the plan before launch, why, and what mitigation to use. It flagged the all-at-once warehouse migration, the ERP readiness assumption and the expectation that training would reduce throughput, and suggested staging the rollout, proving the integration before the first billing cycle and building in training time.

Gotchas

Not for every document. If the sensitive history is the thing you want analysed (a full medical record), redaction removes the task; use a governed environment built for that data or keep AI out of it.
The demo used a synthetic pricing plan made for the camera.
Figures quoted from Verizon telemetry (as he reports them): corporate-device AI use at least every 15 days rose from 15% to 45%, two-thirds of those users were on non-company accounts, and source code was the most common material submitted. He also cites NIST's term "security fatigue". Unverified here.
The link to the tool is only in the video description, and the transcript does not say how it is installed or run.