AI in the workplace: useful gains, real risks, and the gap between them

One employee uses AI to draft a customer reply. Another asks it to summarise a meeting. A third uses it to analyse a spreadsheet before the leadership call. None of those moments feels like a company-wide transformation. Together, they are changing how work gets done.

The opportunity is real: quicker first drafts, faster access to information and more capacity for work that used to wait. So are the risks: a confident wrong answer, private information in the wrong tool, a team expected to produce more without being given time to learn, or a decision made without anyone checking its basis.

What is happening inside companies now?

AI use is widespread, but uneven. In Gallup's May 2026 US workforce data, 52% of employees said they used AI in their role at least a few times a year. Thirty per cent used it a few times a week or more; 15% used it daily. Those figures describe US employees, not every workplace or country, but they show the gap between trying a tool and making it part of a working routine.

From the company side, McKinsey's 2026 global survey found that nearly nine in ten respondents reported regular AI use in at least one business function. Yet 44% said AI was scaling across their enterprise. That is a different claim from saying nine in ten companies have transformed how they operate.

In practice, many businesses are living in the middle: people use AI to write, search, analyse, code and handle routine requests, while their processes, permissions and measures of success are still catching up. McKinsey reports agents being scaled most often in IT, knowledge management and software engineering. The everyday use is already here; the operating model is still being worked out.

The good: more capacity where the work is clear

AI can make a useful colleague faster without pretending to replace their judgement. It can prepare a first draft, pull together a long history of notes, flag exceptions in a report, help a service team find a relevant answer, or give a developer a starting point for a routine task. The gain is strongest when the input is available, the output can be checked and someone knows what a good result looks like.

A well-known field study of 5,179 customer support agents found that access to a generative AI assistant increased issues resolved per hour by 14% on average, with a much larger gain for newer and lower-skilled agents. It is one setting, not a promise that every job will improve by 14%. It does show how AI can spread useful know-how through a team when the task and outcome are measurable.

Employees also report benefits beyond speed. In Microsoft's 2026 Work Trend Index, 66% of surveyed AI users said AI let them spend more time on high-value work and 58% said they were producing work they could not have produced a year earlier. That survey included workers already using AI, so it should not be read as a measure of all employees. It is still a sign of why people keep reaching for these tools.

The useful measure is not how many people have an AI licence. It is whether the work is better, faster and easier to trust.

The bad: speed can hide new costs

A polished answer can still be wrong. An AI summary can omit the exception that mattered. A draft can invent a source, misread a customer's tone or turn a tentative number into a confident statement. The US National Institute of Standards and Technology lists false or misleading output, data privacy and information security among the risks organisations should manage when using generative AI.

There is also a people cost when AI arrives as a productivity target before anyone explains how work should change. Microsoft's 2026 survey found that 65% of AI users feared falling behind if they did not adapt quickly, while only 26% said their leadership was clearly and consistently aligned on AI. Those are perceptions, but they point to a real management task: people need a safe way to learn and a clear account of what stays their responsibility.

Financial returns are less automatic than individual time savings. In the same McKinsey survey, 80% of respondents said AI improved their own productivity, but 37% reported a positive effect on their organisation's earnings before interest and taxes. Those self-reported measures are different, yet the gap matters. A faster draft has little value if it creates rework downstream or never changes a customer outcome, a decision or a cost line.

Jobs are changing too, but sweeping forecasts can be misleading. McKinsey found more respondents expected AI-related headcount reductions in the coming year than in 2025, while actual reductions during the previous year fell well short of earlier expectations. Leaders should talk honestly about roles and skills without presenting either mass replacement or effortless job creation as a settled fact.

What a sensible company does next

The best response is neither to ban useful tools nor to tell everyone to use AI more. Start with a piece of work where the benefit and the failure mode are both visible. Then design the process around the people who will use it and the people affected by its output.

1. 1. Choose one real workflow

   Pick work with a clear owner, a repeatable input and a result you can check. Measure time saved, errors, rework and the quality of the decision, rather than counting prompts or licences.

2. 2. Set a safe boundary for information

   Tell people which approved tools they can use, what customer or employee information may enter them, and what must stay out. Make the safe route easier than an unofficial workaround.

3. 3. Keep a person accountable

   Name who checks the output and signs off before it reaches a customer, changes a record or informs a consequential decision. Review should match the cost of being wrong.

4. 4. Train managers alongside teams

   Show managers how to evaluate a use case, challenge an answer and discuss changes to roles. Give employees time to learn and a way to report failures without penalty.

This is where many AI programmes become more grounded. Gallup found that only 25% of US employees said their organisation had communicated a clear AI plan in May 2026. Clear guidance is not bureaucracy for its own sake. It lets a team know where experimentation is welcome, what needs review and what success should look like.

AI in the workplace is no longer a future question. The choice now is whether companies let it grow as a collection of individual shortcuts, or turn the useful parts into a reliable way of working. That requires good data, clear ownership, human judgement and enough evidence to know when a tool is genuinely helping.