Using AI Tools at Work Without Getting Into Trouble

Most offices now have someone who uses AI for everything, someone who refuses to touch it, and a lot of people in between quietly pasting things into a chatbot and hoping it's fine.

Used well, AI assistants save real time on drafting, summarising, analysing and coding. Used carelessly, they create two kinds of problems: data that goes where it shouldn't, and mistakes that look polished enough to get through.

This guide is about getting the first without the second.

Before anything else, find out:

• Is there an approved tool? Many companies provide an enterprise version of ChatGPT, Claude, Gemini or Microsoft Copilot. These usually come with agreements that your data won't be used to train models and with admin controls.
• What data is allowed? Some policies forbid customer data entirely. Others allow internal documents in the approved tool only.
• Are there banned tools? Browser extensions, meeting recorders and "free AI" sites are common problem areas.

If there's no policy, ask your manager or IT team. "I'd like to use AI for drafting reports, is that okay and which tool should I use?" is a sensible question, and it protects you if something goes wrong later.

Use your work account in the approved tool for work. Pasting company material into a personal free account is the most common way people break policy without meaning to.

Even in an approved tool, think twice. In an unapproved or personal one, never paste:

• Personal data about customers or employees: names with phone numbers, Aadhaar or PAN details, addresses, health information, salary data
• Credentials: passwords, API keys, tokens, database connection strings. These often hide inside config files and logs people paste to "debug"
• Unreleased financial results, deal terms, pricing strategy, M&A information
• Confidential source code, unless your company has explicitly approved the tool for code
• Legal documents under negotiation, or anything covered by an NDA

A good habit is to anonymise before you paste. Replace "Rahul Mehta, account 4471-XXXX, owes ₹3.2 lakh" with "Customer A owes amount X". The AI can help with the structure of your problem without seeing the sensitive parts.

India's data protection law, the Digital Personal Data Protection Act, 2023, makes careless handling of personal data a bigger issue for companies, so expect policies to get stricter, not looser.

Language models generate text that sounds right. Most of the time it is right. Some of the time it's confidently wrong, and the wrong answers look exactly like the right ones.

Common failure modes:

• Invented facts: statistics, quotes, court cases, research papers or product features that don't exist
• Wrong numbers: arithmetic slips, especially across many steps
• Outdated information: tax rules, prices, laws or software versions that have changed since the model was trained
• Plausible but broken code: code that runs but mishandles edge cases, or uses a library function that doesn't exist
• Missing context: advice that's generally correct but wrong for your company, your client or Indian regulations

Before anything AI-assisted goes out, ask yourself:

1. Facts and numbers: Did I check each one against a real source?
2. Names and references: Do these people, documents, laws and links actually exist?
3. Code: Did I read it, run it and test it, including the edge cases?
4. Tone: Does this sound like me and fit the person reading it?
5. Accountability: Would I be comfortable defending every line if asked?

The last question is the important one. Your manager won't accept "the AI wrote it" as an explanation. You sent it, so it's your work.

These are the tasks where most people get real value with low risk.

• First drafts of emails, reports, proposals and documentation
• Rewriting something to be shorter, clearer or more polite
• Turning rough bullet points into a structured update
• Checking grammar and tone, which is especially useful when writing to a senior stakeholder or a client in another country

• Brainstorming options before a meeting
• Asking "what am I missing?" about a plan
• Playing the sceptic: "Give me the three strongest objections to this proposal"
• Preparing questions for an interview, negotiation or review

• Summarising long documents you're allowed to share with the tool
• Explaining an unfamiliar concept, acronym or piece of code
• Turning messy notes into a table
• Writing spreadsheet formulas and explaining what an existing one does

• Boilerplate, test cases and documentation
• Explaining an unfamiliar codebase or error message
• Converting code between languages or frameworks
• Writing scripts for one-off tasks

Vague prompts get generic answers. Most improvement comes from giving the tool what a new colleague would need.

A useful structure:

1. Context: who you are, who it's for, what the situation is
2. Task: what exactly you want
3. Constraints: length, tone, what to include or avoid
4. Format: bullet points, table, email, code

Weak:

Write an email about the project delay.

Better:

I'm a project manager at an IT services company. Write a short email to our client's operations head explaining that the inventory module will be delivered on 20 November instead of 6 November, because their team's test data arrived late. Tone: professional and calm, not defensive. Take responsibility for keeping them updated, propose a 30-minute call this week, and keep it under 150 words.

The second prompt takes 40 seconds longer to write and usually saves several rounds of editing.

• Give examples. "Here's a status update I wrote last week. Match this style."
• Ask for options. "Give me three subject lines" beats "write a subject line".
• Iterate. Treat it as a conversation: "Shorter." "Less formal." "Add a line about the budget."
• Ask it to flag uncertainty. "If you're unsure about any fact, say so explicitly."
• Ask for its reasoning on analysis tasks, so you can spot where it went wrong.

Norms are still forming, but some guidelines are sensible:

• Follow your company's disclosure rules, if any.
• Be open when it matters. If a client or manager would care, for example in legal, academic, medical or highly sensitive work, tell them how AI was involved.
• Don't pass off AI analysis as research you did. "I used AI to summarise the 40 customer interviews, then checked the themes against the transcripts" is honest and still impressive.

The honest answer: AI is changing what jobs look like faster than it's removing them, and the people at most risk are those whose work consists only of tasks AI does well.

The practical response is to become the person who uses these tools well: faster at the routine work, and spending the saved time on judgement, relationships, domain knowledge and decisions, which AI still handles poorly. Someone who can do their job and use AI thoughtfully is more valuable than either alone.

It depends on the version and your company's agreement. Enterprise and team plans usually come with stronger data protections than free personal accounts. Use the version your company has approved, with your work account, and follow its data rules.

On company-provided enterprise tools, administrators may be able to review usage, depending on the setup. Assume anything you type into a work tool may be visible to your employer.

In workplaces, the concern is usually accuracy, confidentiality and honesty rather than plagiarism in the academic sense. In education and publishing, rules vary widely, so check them. In every setting, never present AI-generated facts or quotes without verifying them.

The best tool is the one your company has approved and that fits your task. The major assistants are all capable. Learning to prompt and verify well matters far more than which one you pick.

You can't eliminate it entirely, but you can reduce it: give it source material to work from instead of relying on its memory, ask it to quote or cite the parts of the source it used, ask it to say when it doesn't know, and verify anything important yourself.