The Order of Operations for Safe AI at Work

In maths, order of operations is why 2 + 3 × 4 is 14, not 20 — same numbers, same operations, and the sequence alone decides whether you're right. Safe AI use at work has the same property, and almost nobody has noticed: most AI incidents aren't people doing wrong things. They're people doing right things in the wrong order.

Verified the output — after the client saw it. Considered data sensitivity — after pasting. Tested the automation — after it ran on real invoices. Right actions, wrong sequence, real damage. Here's the sequence.

Step 1: Protect — Before Anything Enters

The first operation always happens before the prompt: what am I about to disclose, and to what?

• Client names, personal data, credentials, unreleased financials — swap for placeholders ("Client A", "[product]") or strip entirely. The structure of your question survives; the sensitive payload doesn't.
• Know your tool's tier: a public free chatbot, a paid account with data controls, and a company-approved deployment are three different disclosure environments. What's fine in one is an incident in another.
• When unsure, the sanitised version costs you thirty seconds. The unsanitised version can't be un-pasted — treat every prompt as a postcard.

Why this is step 1 and not a general guideline: it's irreversible. Every later step can be redone. This one can't.

Step 2: Verify — Before Anything Leaves

Between "the AI produced it" and "it goes onward" sits the verification gate — targeted, not total:

• Always check: numbers, names, dates, quotes, citations, legal or medical claims, anything client-facing or irreversible once sent.
• Skip freely: brainstorms, internal drafts, phrasings — work where an error costs a shrug.
• The mechanism to remember: models generate the plausible, which overlaps with the true only most of the time (why, in ten minutes). Fluency is not evidence.

Step 3: Trust — Earned Incrementally, Never Granted

Only after protect and verify are habits does the third operation apply: extending trust to AI for repeated or autonomous work.

• Pilot on low-stakes volume first. The automation earns the invoice run by surviving the newsletter run.
• Keep spot-checks forever — trust means checking less, never zero.
• Define the escalation rule before deployment: what the system must hand back to a human, written down, in advance.

The classic failure is running this backwards: trusting first (it demoed well!), verifying after the first incident, thinking about data exposure last. That's 2 + 3 × 4 = 20 — same operations, wrong answer.

Why Sequences Beat Checklists

A checklist says what; a sequence says what, when — and in the moment, when is what fails. The discipline of ordering-under-constraints is old and trainable: it's literally the logic of the 1,200-year-old river-crossing puzzle — some things can't be left alone together, some things must cross first, and the boat is small.

We built that training loop directly: River of Reasoning is a free 3D puzzle where you ferry AI tools, work assets and risks across the river in the right order — protect your data, verify outputs, don't sink the boat. Ten minutes of play installs the sequence as an instinct, which is where it needs to live (the broader case: The Missing AI Skill Is Judgement).

The Pocket Version

Protect → Verify → Trust. Nothing sensitive goes in unsanitised; nothing consequential goes out unchecked; nothing runs unsupervised until it's earned it. Three operations, one order, and the order is the safety.

Practise it as a puzzle →