Responsible AI Examples: What It Looks Like at Work

“Responsible AI” gets talked about as a company-level policy question — a framework, a committee, a set of principles on a slide. For most people at work it is smaller and more concrete than that: four habits, repeated every time you open a chat tool, that stop a convenience from becoming a problem. None of them require permission from anyone. They are things one person can just start doing today.

1. Remove confidential details before you paste anything in

The most common real-world failure is not exotic. Someone pastes a client contract, a patient note, or an unreleased product spec into a general chat tool to get help rewriting it, and that text has now left the building in a way it never did before. The fix is mechanical, not philosophical: before you paste, swap out names, account numbers, exact figures and anything else that identifies a real person or a real deal with placeholders — "Client A", "£X", "[date redacted]" — then ask your question about the redacted version. If the task genuinely needs the real numbers to work, that is a sign the task needs a tool with an actual data agreement behind it, not a workaround.

“Confidential” is wider than personal data. An unreleased pricing change, a draft redundancy list, a term sheet still being negotiated — none of these name a person, but all of them cause real damage if they turn up somewhere they should not. The same redact-first habit applies: strip the specific figure or the specific name of the deal, keep the structural question (“how do I phrase this clause more clearly”), and paste that instead of the whole document.

2. Check what actually happens to what you paste

This is not the same question as "is the company allowed to use this tool." It is "where does my input go." Several providers let you turn off the setting that uses your conversations to improve their models — Claude’s own privacy settings are one example of where that control lives — and a free consumer tier and a paid business tier from the same company can have different defaults. Five minutes checking the actual setting, once, is the whole task. Guessing is not a substitute for checking.

3. Verify before it leaves your desk

A draft is not a fact until someone checks it against a real source — the mechanism behind a confidently wrong answer does not announce itself, so the check has to happen on purpose, every time, not just when something feels off. A worked method for checking a summary against its source is the specific version of this habit worth building first, because summarising something you already have is the single most common task these tools get wrong in a way that is also the easiest to catch.

4. Keep a person deciding on anything that cannot be undone

Drafting a reply is reversible — nobody has read it yet. Sending it, approving a payment, or changing a live record is not. Separating what a model can draft from what still needs your approval is the fuller version of this split, and it is the one habit on this list that protects everyone else, not just you, when the task touches another person’s money, data or time. A refund approval and an apology email look similar on the surface — both are a short piece of text responding to a complaint — but only one of them moves money, and that is the one that needs a person reading it before it goes anywhere, every single time, not just on the cases that look unusual.

A worked example: a client email with real numbers in it

A client emails asking for an update on an account with specific figures in it — a balance, a fee, a date. Before pasting the email into a chat tool for help with the reply, replace the real balance and fee with placeholders and strip the account number entirely, keeping the tone and the actual question intact. Ask for a draft reply addressing the question the client actually asked. Once the draft comes back, put the real numbers back in by hand, and read the whole reply once more specifically checking that every figure matches your own records before it goes out — not a glance, an actual comparison.

That sequence — redact, draft, restore, verify — takes under two minutes longer than pasting the original email in directly. It is the whole method. Nothing confidential left your hands unredacted, and nothing shipped unchecked. The same four steps work whether the task is a client email, a short internal memo with a number in it, or a reply to a colleague about a deal that is still being negotiated — only the thing being redacted changes.

Where this goes wrong

• —Redacting the obvious things (a name) and missing the less obvious ones (a reference number that is unique enough to identify the same person once combined with other details).
• —Treating "the company has a policy" as the same thing as "I checked today’s setting myself" — policies lag behind product changes, and a setting can default differently than the policy assumes.
• —Confusing a certificate with a practice. AI ethics certification is about what a credential actually represents, which is a different question from whether today’s redact-draft-verify habit happened on a real document this week.
• —Skipping the verification step specifically on drafts that already look polished — a well-formatted reply with a wrong figure in it reads exactly as convincing as a correct one.

Responsible use is not a policy document. It is the same four small checks, done on purpose, on the one email or spreadsheet actually in front of you right now.

What to do Monday

Pick the task you already do most often with a chat tool and run the four habits on it once, deliberately, rather than assuming they already happen by default. If that task involves another person’s data by nature — a client file, a patient record, a customer ticket — how to use AI as a lawyer and how to use AI as a banker both cover what confidentiality means concretely in professions built around exactly this kind of material, and customer support automation covers the same redaction habit applied to a support inbox rather than a single email. If the task runs through a spreadsheet rather than a chat window, Excel accounting software covers the same sign-off discipline for figures that end up on a filing.

None of this requires waiting for a written AI policy to exist first — AICPA & CIMA’s small-firm AI policy template and the monitoring habit in NIST’s AI Risk Management Framework are both useful once a team writes one down, but one person can start the four habits above alone, today, on the next thing they paste into a chat tool. That judgement is close to the skill behind the wage premium PwC has measured for workers with real AI skills: 62%, up from 57% the year before. Coursium teaches it directly, through ChatGPT Essentials. Stay ahead of AI by learning the tools on your phone.