AI at work: what it is really used for right now

The honest picture of AI at work is smaller and more useful than the marketing suggests. People paste a rough paragraph and ask for a cleaner one. They drop a 40 message email thread in and ask what was actually agreed. They translate a supplier's message, get a first pass at a spreadsheet formula, clean up a messy column of addresses, or let an assistant write up the meeting they just sat through. That is the bulk of it, across almost every office job. If you were promised an autonomous digital colleague and you are wondering where it is, the boring version arrived first, and it lives inside documents, inboxes and chat windows.

The six tasks that make up most real use

Drafting and rewriting is the largest by a distance. Not "write my report" but "make this three paragraphs shorter", "turn these bullet points into an email to a customer", "give me three subject lines". Narrow requests like those are also how you get a usable answer first time. The model is doing the first 70 percent and you are doing the last 30, which is the part that carries your judgment.

Summarizing comes second: long threads, call transcripts, policy documents, a 30 page tender you need the gist of before a meeting at two o'clock.

Translation is third, and it is where the quality jump has been most obvious to ordinary users. It is good enough for internal comprehension and routine supplier correspondence, and still not good enough for contracts, marketing copy or anything where tone carries money.

First pass code and formulas sit fourth. A regular expression, a spreadsheet formula with four nested conditions, a script that renames 400 files. People who are not developers use this more than developers expect.

Data tidying is fifth: reformatting dates, splitting names, matching two lists that nearly agree, turning a pasted table into something structured.

Meeting notes are sixth and growing fastest, because the tool joins the call and produces something without anyone changing their habits. That convenience is exactly why it needs rules, which is the subject of AI meeting notes and consent.

Why the pilot never became a habit

Most organizations that say "we tried AI" ran a demonstration, not a trial. Someone showed a model writing a product description, everyone nodded, and then people went back to work where nothing had changed. Habit forms around a specific task that a specific person does every week, not around a general capability.

The second reason is access friction. If the tool sits behind a separate login, a request form and a two week approval, nobody will use it for a five minute job. The tools that stick are the ones already inside the thing people have open: the document editor, the ticketing system, the inbox.

The third is that nobody said what was allowed. When the rules are unclear, cautious employees do nothing and confident ones do whatever they like. Both outcomes are bad, and the second creates a data trail the company cannot see.

Shadow AI is a data problem, not a discipline problem

Shadow AI means staff using personal AI accounts for work tasks, usually on their own phone, usually because the company has not provided anything. It is the same pattern as people emailing files to personal addresses before file sharing was sorted out: a reasonable person routing around an obstacle.

The risk is not that they are lazy. It is that customer names, salary figures, draft contracts and patient details end up in a consumer account with no business agreement behind it, no retention control, and settings that may allow the provider to use the content to improve its systems. Your company then cannot answer basic questions if something goes wrong: what was shared, by whom, when, and can it be deleted. The mechanics of who processes what are set out in what happens to the data you put in.

The fix is unexciting: sanction one tool, on a business plan, with training data use turned off, and tell people plainly which categories of information must never be pasted in. That structure is covered in writing an AI policy people will follow.

Finding your own first use case

1. For one week, keep a note every time you write, rewrite, summarize or reformat something. Just the task and rough minutes.
2. At the end of the week, circle anything that appears three or more times.
3. Apply the three tests. Is it repeated? Is it mostly text rather than judgment? Can you verify the output faster than producing it yourself?
4. Pick the one that passes all three and is lowest risk if it is wrong. An internal summary beats a customer email for a first attempt.
5. Run it for two weeks with a real deadline attached, and keep the same note: minutes saved, and how often you had to redo the work.
6. If the rework rate is high, the task is a poor fit. Try a different one rather than a different tool.

The third test is the one people skip, and it is the one that decides whether you save time. If checking takes as long as doing, you have added a step. That trade off is unpacked in what AI really saves you at work.

What stays your job

The model does not know what is confidential, what your regulator expects, what you promised a customer in March, or which of two plausible numbers is the real one. It produces fluent text regardless of whether the underlying claim is true, which is why anything factual needs checking against a source you trust rather than against how confident the answer sounds.

Accountability does not transfer either. If a quote goes out with the wrong price, the fact that a model drafted it is not a defense to the customer, to your manager or to a regulator. Treat output as a draft from a fast, tireless, occasionally confident colleague who has never seen your files.

What to check first

Before you use AI for anything at work this week, find out two things: whether your employer has an approved tool, and what you are not allowed to enter into it. If the answer to the first is no, ask for one and name the task, because a specific request ("summarizing support tickets") gets approved far more often than a general one. If nobody can answer the second, write down your own red lines now: no customer identifiers, no payroll data, no unreleased financials, no health information, no credentials. Those five cover most of what would actually hurt.