Using AI at Work

Part 1

The line that must not be crossed

Almost every serious incident is the same mistake: information that belonged to someone else went into a tool the employer had no agreement with. Learn this list as a list.

• Customer or client names, contact details, account numbers, or anything identifying an individual.
• Patient information of any kind. Even a de-identified case can be re-identified from small details.
• Employee records: reviews, salaries, disciplinary notes, medical or immigration paperwork.
• Source code, if your employer treats it as proprietary. Many do, and this is the classic incident.
• Unreleased products, pricing strategy, financials before publication, or anything under NDA.
• Login credentials, API keys, or connection strings. This happens constantly and is worse than it sounds.
• Anything from a third party you signed a confidentiality agreement to protect.
• Legal advice you received, which can waive privilege over the whole matter.

Part 2

How to get the value anyway

The restrictions above rule out much less than people assume, because the highest-value uses do not require sensitive data at all.

1. 1

   Redact, then paste

   Replace real names with Customer A, real figures with round placeholders, real addresses with a city. Get the draft back, then substitute the real details yourself. Nearly everything you actually wanted still works.

2. 2

   Describe rather than paste

   'Write a firm but polite follow-up to a client who is sixty days late on a five-figure invoice, where we want to keep the relationship' needs no real data and gets you 90% of the way.

3. 3

   Use it on your own words

   Your rough notes, your bullet points, your voice memo. That is your own material, and turning a mess into clear prose is the single biggest daily time saving available.

4. 4

   Use it to prepare, not to send

   Rehearse a difficult conversation, anticipate objections in a proposal, generate questions before a meeting. No sensitive input, direct benefit.

5. 5

   Use it to understand

   Paste a public regulation, standard, or manual and ask what it requires of you. Public documents carry no confidentiality problem.

Part 3

What it is genuinely good at, by job

Concrete rather than aspirational, from the kinds of work we see across DFW.

• Office and admin: turning meeting notes into minutes with actions, drafting recurring correspondence, writing the awkward email.
• Sales: first-draft proposals, follow-up sequences, rehearsing objections, summarizing a long RFP down to what it actually asks.
• Trades and field work: writing up an estimate clearly, explaining to a customer why the repair costs what it does, standardizing job sheets.
• Retail and food: rewriting menu or product descriptions, drafting policies, turning a supplier email into a plain summary.
• Bookkeeping: explaining an unfamiliar term, drafting the chase letter, writing procedure notes so the work survives you being off.
• Management: structuring a review from your own notes, planning an agenda, writing the first version of a process that has only ever lived in someone's head.
• Anyone who writes: getting past a blank page. Editing something mediocre is dramatically easier than starting.

Part 4

Disclosure, and being straight about it

The question people worry about most, usually more than they need to.

1. 1

   Follow the written policy first

   If one exists, it settles the matter. If it does not, ask in writing and keep the answer.

2. 2

   Nobody expects disclosure for shaping your own work

   Using a tool to tidy your own email is like using spellcheck. Announcing it is not required and mostly reads as odd.

3. 3

   Say so where authorship is the point

   Performance reviews you write about someone, condolences, references, anything where the value is that a person sat down and thought about it. Using AI silently there is a trust problem rather than a policy one.

4. 4

   Always disclose where accuracy is claimed

   Research, analysis, legal or medical summaries, anything someone will rely on. Say what was AI-drafted and what you verified.

5. 5

   Never let it speak as you in a matter of judgment

   Approvals, sign-offs, testimony, disciplinary decisions. If your judgment is the product, it must actually be your judgment.

Part 5

A policy for a small business

If you employ people and have no policy, you have a policy — it is 'everyone guesses'. This is a workable one you can adapt in an afternoon.

1. 1

   Name the approved tools

   List which tools are allowed and for what. Provide at least one, because a blanket ban with no alternative gets ignored quietly, which is worse than a rule people follow.

2. 2

   Define the data that never goes in

   Use the list above, edited for your business. Be concrete — 'sensitive information' means nothing, 'customer names, card numbers, and anything from a patient file' means something.

3. 3

   Require a human check on anything that leaves

   Nothing AI-drafted reaches a customer, a regulator, or a public channel without a named person reading it. Put the responsibility on a role, not on 'the team'.

4. 4

   Say who decides the edge cases

   One named person to ask. Ambiguity is what produces the bad decision, and staff should never feel that asking is a career risk.

5. 5

   State the amnesty explicitly

   Anyone who realizes they pasted something they should not have must be able to say so immediately without being punished for the disclosure itself. You want to know in the first hour, and fear is what stops that.

6. 6

   Review it twice a year

   The tools change quickly. Diary it, or the policy quietly becomes wrong and everyone knows it.