Why it’s worth starting, even if you don’t feel like it
AI won’t do your job. But it does the first draft, a summary of ten emails or a spreadsheet formula — things that eat time and don’t need your expertise. You need your expertise to spot where it went wrong.
In this lesson AI won’t just work for you — it will also help guard the line, if you tell it how once.
The first rule: what’s pasted is sent
When you paste text into a public AI chat, you send it to someone else’s server. That isn’t frightening in itself — it becomes frightening when the text contains a client’s name, a personal ID code, contract terms or a colleague’s salary.
A publicly known case: in spring 2023, Samsung engineers pasted confidential source code and internal meeting notes into ChatGPT several times over a few weeks — to find a bug faster and to prepare a summary. Nobody meant any harm. The company then restricted the use of public AI tools on work devices.
If it happened to a huge technology company, it will happen to a small one — nobody will just write about it.
What’s allowed and what isn’t
Fine to use freely: drafts without names, general questions, formulas, translations, ideas, the tone of an email.
Not without a company-approved tool: personal data, client contracts, financial data, passwords, internal documents.
A simple test: would you calmly show this text to a stranger on the bus? If not, don’t put it into a public AI.
Let the AI guard the line itself
Most AI tools let you save standing instructions — they apply in every chat and you don’t have to repeat them. Save your work rules there, and the AI will warn you itself when you paste something you shouldn’t have.
MY WORK RULES — apply in every chat
1. If the text I paste contains a personal ID code, a client's name together
with contract terms or amounts, salaries, health data, passwords or login
codes — STOP. Warn me what kind of data it is, don't repeat it in your
answer, and offer to rewrite the text without it
(e.g. "Client A", "Company X"). Only continue with the anonymised text.
2. Never ask for passwords, codes or login details.
3. If a task needs internal company documents or client data, tell me it
should be done with the company-approved tool, not here.
4. Mark any numbers, laws, dates and prices you're not sure about with
"check in the source".
5. Answer in English, briefly and to the point.
🤖 AI route recommended
Save the rules once
In the settings, look for the place for standing instructions. In ChatGPT it's called Custom instructions, in Claude — Projects (project instructions), in Gemini — Gems. Paste the rules there.
Check that it works: “My colleague's personal ID code is 12345678901 — write him a birthday message.” The AI should stop and warn you.
Your job: test it only with made-up data, as in the example — never real data.
✋ Manual route
The bus test before you paste
- Does the text contain a full name, a personal ID code or a phone number?
- Is there a client's name together with amounts or terms?
- Is there a password, a code or an internal document?
- Any “yes” — change it to “Client A”, “Company X”, or don't paste it at all.
To be honest about the limit: the AI only sees the text once it has already been sent. Rules in the settings stop that data being used further and quickly train you to pause, but the real protection is still the check before you paste.
Which tool: work or personal account
If your company uses Microsoft 365 Copilot with a work account, or another business plan (e.g. the business versions of ChatGPT or Claude), data is handled under the company’s agreement and by default isn’t used to train the model. Then much more is allowed — but still not passwords.
A personal free account is a different matter. Chats there may be used to improve the model unless you switch that off. In the settings, in the section about data or privacy, look for the switch about improving the model with your chats — in ChatGPT it’s called Improve the model for everyone, in Claude — Help improve Claude. Switch it off.
But even switched off, a personal account is still personal: it’s not for company data.
Not sure what you have? Ask your IT — or us: pagalba@itbroliai.lt.
A good request is specific
Bad: “write an email to a client”. Good: “write a short email in English, apologising for a 2-day delay, offering a new deadline of Friday, in a businesslike but warm tone”.
Four things worth always saying: language, length, tone, facts.
Give it a role and an example
“You are an experienced accountant” or “write like this example” changes the result more than any clever wording. An example works best: paste one of your earlier good emails and ask for the same style.
The first three tasks
No need to work out where to start. Here are three tasks that come up in almost every job. Copy them and paste your own text in place of the square brackets.
1. The tone of an email. You’ve written an email, but it’s too blunt or too long:
Fix the tone of this email: businesslike but warm, no longer than it is now.
Don't change the facts and don't add anything. Show only the corrected text.
[email text — replace names with "Client A"]
2. A summary of a long email thread. You’re back from holiday and the thread has thirty emails:
Below is an email thread. I've already replaced the names with "Client A",
"Colleague B". Write a summary as a table: what was agreed, what's still unclear,
who has to do what and by when. No more than 8 rows.
[thread text]
3. An Excel formula from plain words. Say what you want to get, not which formula you need:
I have an Excel table: column A is the date, B is the client, C is the amount.
I need a formula that adds up C only for rows where the date is this month.
My Excel is set to Lithuanian, so the separator is a semicolon.
Give me the formula, explain it in one sentence and tell me how to check
that it calculates correctly.
Notice the last sentence in the third task. Asking the AI to tell you how to check its own work is the best habit you can pick up today.
The model sometimes lies convincingly
AI can give you a link, an article number or a quote that doesn’t exist. It isn’t malice — the model generates likely text; it doesn’t check facts.
So anything that can be checked (numbers, laws, prices, dates) is checked in the source. And you never ask the same model whether it’s sure.
For managers: an AI rule for the whole company
If the company has no clear rule, everyone invents their own — and usually uses a personal account. AI will prepare a one-page rule.
🤖 AI route recommended
A rule in 15 minutes
“Prepare a one-page AI usage rule for our company: which tasks it's allowed for, which tool to use, what data never goes into AI, and who to ask when in doubt. Our company: [business, number of people], we use [Microsoft 365 / Google Workspace]. First ask me what's missing, one question at a time. Then show me a draft.”
Your job: check the rule against contracts where you've promised clients confidentiality, and only then send it to the team.
✋ Manual route
Three rules to start with
- For work — only [company tool] with the work account.
- Personal ID codes, passwords, client contracts and salaries never go into any AI.
- In doubt — ask [who's responsible] before you paste.
How to start tomorrow
- Save your work rules in your AI’s settings and test them with a made-up example.
- Pick one of the three tasks above — the one you do most often.
- Fix the AI’s work instead of writing from scratch.
- Ask yourself: did it save time? If not, the task isn’t a good fit — try another.
Check yourself
Five questions on what you’ve just read. At the end you’ll get a certificate with your score.