AI in everyday life · Research checked 13 October 2026

At 9:07 a.m., Riya asks an AI assistant to make a customer email sound friendlier. It returns three polished paragraphs, a warm greeting and enough enthusiasm to open a theme park.

She deletes half of it. The remaining half is useful.

By 9:10, she has discovered a surprisingly good description of the AI economy: impressive assistance, followed by a human deciding what should actually leave the building.

Riya, her colleagues and their working day are fictional. The studies and deployments linked below are real. This story uses an imagined office to explain what the evidence says about AI in everyday work.

Riya works at a small business that sells office supplies online. Her day contains customer questions, a promotional campaign, a spreadsheet and a meeting that could probably have been a spreadsheet. She is a useful guide to a question bigger than any single chatbot: what is AI actually doing in the real world?

9:15 a.m.: AI has already joined the email thread

Riya's assistant helps her turn rough notes into a clear reply. She provides the order status, the delivery window and the shop's actual policy. The tool supplies the wording; she checks the promises.

This division matters. “We will look into it” and “Your refund is approved” may have similar word counts, but only one should appear when nobody has approved a refund.

There is evidence for this less theatrical version of AI adoption. Anthropic's June 2026 Economic Index report found that, in its analysed Claude chat and Cowork conversations, work-related outputs most often included documents and reports, explanations, email drafts, and analyses or summaries. Documents and reports accounted for 20% of work conversations; email drafts accounted for 7%.

Those are observations from one provider's products and classification system, not a census of every workplace. Still, the picture is recognisable: a technology discussed in civilisation-sized terms is also helping someone get through Tuesday's inbox.

Riya's next request is smaller: fewer exclamation marks. The customer ordered printer paper, not a surprise birthday party.

10:30 a.m.: the customer-support desk gets a useful second brain

A customer asks why an order appears twice. The AI helps Riya organise the checks: compare order numbers, inspect the payment record and check whether one entry is only a confirmation message. She checks the actual records before responding.

The appeal is obvious. A newer employee can use a reminder of what to look for without interrupting the busiest person in the office every six minutes. The experienced colleague might finally drink tea while it is still recognisably tea.

A 2025 paper in the Quarterly Journal of Economics (author's PDF) studied an AI assistant introduced to 5,172 customer-support agents at a business-software company. Access to the assistant increased issues resolved per hour by about 15% on average, with larger benefits for less experienced and lower-skilled workers.

The system suggested responses and relevant documentation. Human agents retained control over what to use. The rollout studied mainly occurred in 2020–2021, so this is evidence from an earlier deployment, not a test of today's newest models. It also concerns one firm and occupation; it does not establish a universal productivity increase.

For Riya, the useful idea is straightforward: put relevant knowledge within reach while keeping somebody responsible for the customer. A helpful suggestion becomes valuable when it leads to a correctly resolved problem.

12:00 p.m.: the paperwork has met its natural predator

The team finishes a meeting about a delivery promotion. Riya asks the assistant to turn the notes into an action list. It organises the decisions neatly. One item, however, turns a tentative suggestion into a firm deadline.

She checks the notes and corrects it. Apparently, confidence is available in unlimited quantities; approval still needs a name beside it.

Large organisations are exploring similar administrative uses. In a UK government trial reported in June 2025, more than 20,000 civil servants received Microsoft 365 Copilot for three months. Uses included drafting documents, handling communications and summarising information. Participants reported average savings of 26 minutes per day.

The measurement matters: those time savings were self-reported. The report also says the experiment could not establish how the saved time was spent. A shorter task is promising; it does not, by itself, show that public services improved by a matching percentage.

In our fictional office, Riya uses the time she saves to telephone a supplier about an unclear delivery commitment. The assistant has helped with the document. The unresolved conversation still needs a person.

2:30 p.m.: the developer discovers that “faster” needs a stopwatch

Across the room, Arjun is fixing the shop's checkout page. The AI produces code quickly. Unfortunately, part of it assumes the website works in a way the website has never agreed to.

Arjun checks the existing code, rewrites a section and runs the tests. The first draft arrived at remarkable speed. The finished fix took longer than the first draft would like to admit.

Research has found that this relationship can change with the task and the tools. METR's February 2026 update revisits an early-2025 experiment in which experienced developers took 19% longer on their own projects when allowed to use AI tools. In its follow-up, METR thought newer tools were probably providing more help, but said selection effects made the size of the benefit unreliable to estimate.

Some developers avoided participation or withheld tasks because they did not want to do that work without AI. The researchers were losing sight of exactly the people or tasks that might benefit most. Changes in pay and difficulties measuring concurrent agent work complicated the picture too.

That update deserves to travel with the old headline. The 2025 result is not a timeless verdict on AI coding, and the follow-up is not a clean measurement of a universal speed boost.

Arjun adopts a less glamorous metric: did the change work, did it pass the checks, and how much time did the entire process take? The marketing department may struggle to fit that on a sticker. The website appreciates it.

4:00 p.m.: the boss asks whether AI can do everyone's job

The manager sees the quicker email drafts and asks the question hovering over the whole afternoon: could the business eventually run with fewer people?

Riya looks at her day. Writing sentences is one part. Finding the correct order, understanding a frustrated customer, checking a supplier's commitment and deciding which promise the business can keep are other parts.

Turning several tasks into AI-assisted tasks does not tell us exactly what will happen to headcount. A company might serve more customers, change roles, reduce hiring or cut positions. These are different outcomes, and the evidence presented here does not predict which will happen at Riya's imaginary shop.

Even expectations need careful handling. Anthropic's June report includes a survey linked to about 9,700 Claude users' activity. It explicitly warns that this sample is not representative of the general population. A user's prediction about next year's job is a perception, not a recorded future outcome.

The useful management conversation is more specific: which tasks have changed, what quality is acceptable, who checks the result, and what happens to the time saved? “We installed AI” is a purchasing update. It still needs an operating plan.

Riya proposes a modest start: measure one repeatable task before redesigning the entire company around a particularly charming demo.

6:30 p.m.: AI clocks out of the office and enters the kitchen

At home, Riya asks for dinner ideas using the ingredients she already has. The assistant suggests a dish requiring an ingredient she definitely does not have. After a correction, it provides something workable.

This shift from office tasks to personal help also appears in Anthropic's usage data: personal conversations increased as a share of activity on weekends, while recipe requests became more common in the evening. AI use follows ordinary human schedules surprisingly well.

Riya closes the laptop and starts cooking. The machine has contributed a plan. The onions remain stubbornly committed to being chopped by somebody with hands.

What can you use AI for in your own working day?

Try a small assignment with an outcome you can inspect. The examples below are suggested starting points, not claims that every tool will do them well.

Use information you are permitted to share with the tool. Keep a record of the time spent preparing the request, checking the output and correcting it. A saved minute should survive the checking stage.

For a fuller measurement method, see our AI reliability research briefing and testing scorecard. If you are checking factual answers, use the AI answer fact-checking guide.

The questions people actually want answered

Is AI genuinely being used at work, or is this mostly advertising?

The support-company study, government trial and usage research document actual deployments or interactions. The harder question is how much useful work each deployment produces. Judge a claim by its setting, method and measured result.

Will using AI automatically make me faster?

No single result here establishes that. Try comparable tasks with and without assistance, include checking time and keep the quality requirement the same. Otherwise, you may be comparing a finished job with a very confident draft.

Do I need to become a programmer to benefit?

Document drafting, email editing and working through explanations do not require writing software. Knowing the subject well enough to recognise a bad answer is useful. Choose a task you already understand before delegating something you cannot check.

What skill should I practise alongside using AI?

Practise defining a clear result and evaluating whether you got it. Explain the audience, constraints and evidence needed, then inspect the output. For tools that can take actions, our AI agent guide explains how those systems differ from a simple chat assistant.

The real-world AI story is happening between the headlines

Riya's day ends with better drafts, some corrected mistakes, a completed customer reply and dinner. There is no dramatic announcement. Nobody unveils a chrome robot in reception.

The evidence gives us a more interesting story than a single promise of effortless productivity. Useful assistance exists. Benefits differ across people and tasks. Measuring them takes care, and the tools keep changing while researchers try to keep up.

For your next experiment, pick one recurring annoyance, decide what a good result looks like and check whether AI makes the whole job easier. If it helps, keep it. If it creates extra work, change the approach.

And if the customer email still sounds like a theme park opening, remove two exclamation marks. Progress comes in many forms.

Reporting note: Sources checked on 13 October 2026. This article combines original fictional scenes with published research; its characters, dialogue and office incidents are illustrative. Study dates and limitations are stated beside the findings. It was prepared with AI assistance, and the illustration is AI-generated. See our editorial policy or send a correction.