AI at Work: Practical Uses, AI Agents and What Changes for People

An AI tool can draft a client update in seconds. Someone still has to find the right information, check the figures and decide whether the message is ready to send. If those steps take longer than writing the update yourself, the quick draft has saved you very little.

That is the practical challenge of using AI at work. Its value depends on the whole task, including the effort needed to prepare inputs and correct mistakes. A useful result must be accurate enough for its purpose and worth the time and cost of producing it.

For employees and team leaders, the next step is to identify work that AI can help complete reliably. That means understanding where it saves effort, how AI agents handle several steps, and what people need to retain control over.

The short version:

• Start with recurring work you know well enough to spot a bad result: drafting from approved notes, summarising, comparing documents.
• Count the whole task. Briefing, reviewing and correcting the AI's work all come out of the time it saves.
• An AI agent can carry out several steps towards a goal, but only within the access and limits you give it.
• Research points to jobs changing more often than disappearing, so look at the tasks inside a role, not the job title.
• Introduce AI with a measured trial, a named reviewer and training built on the team's real work.

Where AI can save time at work

Writing, summarising and comparing information are sensible places to test AI. They have visible outputs that a knowledgeable person can check. The best starting point is often a recurring task you understand well enough to recognise a bad result.

Writing and communication

AI becomes more useful when you give it material to work from. For a client update, that might include completed tasks, confirmed results, unresolved issues and the next deadline. Those inputs allow it to produce a draft with substance. Our guide to writing client status updates covers what a good one contains.

A vague request to "write a professional update" leaves too much room for generic language or invented detail. State the purpose, audience and facts the message must preserve. Make clear that missing information should be flagged.

Editing is another practical use. AI can shorten a long email, explain technical wording or identify passages that repeat the same point. The person sending the message still needs to decide whether it says what they mean.

Research and document review

Comparing three supplier proposals can involve repeated searching for prices, exclusions and delivery dates. AI can assemble a comparison and link each entry to its source, making it easier to examine the differences.

The quality of that comparison depends on what it preserves. A low price may exclude implementation. A delivery date may depend on a condition buried elsewhere in the document. Ask the system to identify missing answers and qualifications as well as headline terms.

Source links make checking easier. Open them: a citation can point to a real document without supporting the claim beside it. Our routine for checking an AI answer before you rely on it takes a few minutes.

Choosing a worthwhile use case

Before introducing AI, ask four questions:

1. Does this task consume enough time to justify changing it? Frequent work lets small savings accumulate, though a substantial one-off task can also be worthwhile.
2. Are the inputs available and suitable for the tool? Scattered records, missing context and restricted data can make preparation expensive.
3. Can someone check the result efficiently? A summary of familiar material is easier to assess than an unfamiliar technical recommendation.
4. What happens if the output is wrong? The consequences should determine the review and approval required.

Some repetitive work is better handled by a formula, template or fixed automation. If a process follows exact rules, test those simpler options before adding a system that generates variable answers.

Does AI actually improve productivity?

Research shows why the answer needs context. A study published in The Quarterly Journal of Economics in 2025, Generative AI at Work (opens in a new tab), examined 5,172 customer-support agents. Access to an AI assistant increased issues resolved per hour by 15% on average. Less experienced workers benefited more; the most experienced workers saw small speed gains and small quality declines.

Results can also disappoint. METR's study of 16 experienced open-source developers (opens in a new tab) found that early-2025 AI tools increased completion time by 19% on tasks in repositories the developers knew well. This was a narrow setting, not a result for all software development.

The date matters too. In a February 2026 follow-up (opens in a new tab), METR said newer tools likely provided greater benefits, but participation and measurement problems prevented a reliable estimate of the current effect.

These findings support a practical rule: measure results in your own work. Include briefing, review and correction time. Track quality alongside speed, and repeat the comparison when the tools or tasks change substantially.

A faster first draft is useful only if it helps produce a better or faster finished result.

How AI agents change everyday work

An AI agent can use connected tools to pursue a goal across several steps. Depending on the system and its permissions, it might find documents, compare their contents, create a briefing and revise it after checking for gaps.

That changes what you delegate. You can describe a finished outcome and let the agent organise some of the steps required to produce it. Its usefulness still depends on access to the right information, reliable tools and clear limits on what it may do.

A worked example: preparing for a client meeting

Consider an account manager preparing for a monthly client review. Relevant information sits in recent emails, meeting notes, a project tracker and an approved performance report.

The following is an illustrative workflow, not a measured customer case study. With the necessary connections and access, an agent could:

1. Retrieve the latest approved report and relevant client records.
2. Compare outstanding actions with the project tracker.
3. Identify missed deadlines, conflicting information and unanswered questions.
4. Draft a one-page briefing with links to the supporting records.
5. Prepare an agenda and follow-up email for the account manager to review.

The instruction needs boundaries. Specify the client, reporting period and approved sources. Require the agent to flag uncertainty, preserve conflicting information and leave messages unsent. Give it read-only access where editing records is unnecessary.

The account manager then checks the figures, resolves conflicts and adds context from conversations that were never recorded. The agent's own checking can help, but it does not independently establish that the briefing is correct.

Count the checking time

Suppose manual preparation takes 45 minutes. Using the agent requires five minutes to brief it, ten minutes to review its output and five minutes to correct it. That is 20 minutes of human effort, saving 25 minutes per meeting. Across eight meetings, the saving would be 200 minutes, or three hours and twenty minutes.

Those are illustrative assumptions. Tool costs, setup and maintenance still need to be counted, and waiting time may affect deadlines even when the person can do other work.

Now suppose the agent uses an outdated report and correction takes 30 minutes. Human effort rises to 45 minutes, removing the saving. A polished briefing can still create extra work.

Before expanding the process, test whether it consistently selects the right records and makes errors easy to find. An agent that cannot show where its answers came from is harder to supervise.

Where Plexii fits

Several of the checks in this article are built into PlexiiDesk, the app this blog belongs to, and into Plexii, its assistant:

• It shows what it read. Above each answer, the trace lists what Plexii searched and read, and numbered citations link each statement to its source, which opens with one click.
• It proposes; you decide. Changes arrive as cards with a reason and a preview, and you apply them, or not. Assistant autonomy sets how far that goes: Suggest only, Ask before acting, or Act on low-risk, ask for the rest.
• Its agent works within limits. The Autonomous agent takes a whole goal and works through it in rounds on your desk. Whatever your setting, some actions always wait for you, including drafting an email or chat message and putting time in your calendar. Everything one run applied is a single step you can undo.
• It separates what was agreed from what was said. After a recorded meeting, PlexiiDesk lists the commitments it found, each with the line of the transcript it came from. A commitment it cannot match to a line is marked for you to verify, and nothing is filed until you confirm it.

Two more that fit this article: when Plexii researches on the web, its findings go onto a desk as a table, with every value exactly as on the page, and in documents it drafts or rewrites a passage you preview before it is inserted. For more on the assistant, see the Plexii assistant.

Will AI replace jobs?

The International Labour Organization's 2025 assessment of generative AI and jobs (opens in a new tab) estimated that one in four workers worldwide held an occupation with some exposure to generative AI. Exposure describes potential effects on tasks; it is not a forecast that a quarter of jobs will disappear. The report identified changes to jobs as the most likely overall outcome because many tasks still require human input.

That does not guarantee employment security. If fewer people can complete the same workload, an employer may reduce hiring or staffing. Another may use the capacity to serve more customers or improve service. The outcome depends partly on demand, costs and management decisions.

For an individual worker, examining the tasks within a role is more useful than treating the job title as a single unit. Which activities can be completed with less effort? Which require context that is missing from the records? Where would an error be difficult to detect?

What people should retain control over

Claims that AI will never perform judgment or relationship-based work are too broad. AI can already contribute suggestions, explanations and supportive language. That alone does not establish that it should make consequential decisions without oversight.

Organisations should assign a named person responsibility for decisions that affect employees, customers or significant commitments. That person needs enough knowledge, time and authority to challenge the output. An approval button provides little protection if the reviewer cannot assess what happened. Our post on approving AI changes before they land looks at what a useful approval step shows.

Human contact also matters where someone needs a discussion, an explanation or a way to challenge a decision. Teams should consider the expectations of the people receiving the service before replacing that contact.

Keep a way for people to learn

Junior employees often develop expertise by doing the drafting, research and routine analysis that AI can assist with. If those activities disappear entirely, how will newcomers learn to recognise an error?

Training should preserve opportunities to work through problems, compare their reasoning with AI output and receive feedback from experienced colleagues. Faster production needs to leave room for developing the people who will review future work.

How to introduce AI into a team

A useful starting point is one recurring process with a clear owner and an observable result. Record the existing completion time and quality, then compare several similar tasks with AI assistance. A single successful demonstration gives limited evidence about everyday reliability.

Set the acceptance criteria before the trial. For a client briefing, these might include correct figures, traceable sources, no missed commitments and lower total preparation time. Record failures as well as successes. Stop or revise the process if checking repeatedly consumes the saving or important errors survive review.

Teach people how to check the work

Training should use examples from the team's actual responsibilities. Ask participants to find an unsupported claim, check a calculation and spot a summary that changes the meaning of a source. Practise deciding when to reject an answer or complete the task another way.

People also need to know what information the tool can access. A useful exercise is to compare an answer based on a small document set with one based on the approved current records. The difference makes missing context easier to recognise.

Make the rules usable

An AI policy should answer the questions employees face during a task:

Review the guidance when a tool gains new access, a mistake reveals a gap or the team's work changes. Explain the practical effect of each update and provide training where needed.

Leaders should also explain the purpose of the trial, how results will be measured and what remains undecided about staffing or workloads. Avoid promises about job security that the organisation cannot support. Employees need a way to report poor results without being treated as resistant to change.

A practical next step

Choose one task you expect to repeat this month. Record how long it takes and what a good result must contain. Test AI on comparable work, including the time needed to prepare information and fix mistakes.

Keep the approach if it produces dependable value. Revise or stop it if it adds effort or makes errors harder to detect. That small test gives you evidence for deciding where AI belongs in your working day.

Common questions

How can I use AI at work? Start with recurring work you know well enough to judge: drafting from approved notes, summarising meetings, comparing documents and explaining spreadsheet formulas. Give the AI the facts it must use, and check the result before it goes anywhere.

What is an AI agent at work? An AI agent pursues a goal across several steps using connected tools, such as finding documents, comparing them and drafting a briefing. What it can do depends on the access and limits it is given, and a person still checks the result.

Does AI really make people more productive? Sometimes. A 2025 study of 5,172 customer-support agents found 15% more issues resolved per hour on average, while a 2025 study of experienced developers found tasks took 19% longer. Measure the whole task, including checking, in your own work.

Will AI replace my job? The International Labour Organization found one in four workers worldwide in an occupation with some exposure to generative AI, and judged changes to jobs the most likely outcome. That is not a guarantee: staffing also depends on demand, costs and management decisions.

What should a person keep control of when using AI? Decisions that affect employees, customers or significant commitments, with a named person who has the knowledge, time and authority to challenge the output. Keep human contact where someone needs a discussion, an explanation or a way to challenge a decision.

How do I introduce AI into my team? Pick one recurring process with a clear owner, record its current time and quality, set acceptance criteria, and compare several similar tasks with AI. Train people to check the work using the team's own examples, and write rules that answer the questions people face mid-task.

Sources

1. Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work (opens in a new tab). The Quarterly Journal of Economics, 140(2), 889–942. Figures above use the published study rather than the earlier working paper.
2. METR (10 July 2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (opens in a new tab).
3. METR (24 February 2026). We Are Changing Our Developer Productivity Experiment Design (opens in a new tab).
4. International Labour Organization (20 May 2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure (opens in a new tab).