Your team already uses AI for its mental wellbeing. Is that safe?

n 1966 Joseph Weizenbaum, a computer scientist at MIT, built ELIZA, one of the first chatbots. Its best-known script cast it as a psychotherapist, and its trick was modest: it turned whatever the user typed back into a question. People confided in it all the same. Weizenbaum's own secretary, who knew exactly how the program worked, once asked him to leave the room so that she could talk to it in private.

Sixty years on, the machines are far cleverer and the confidences far more common. Artificial intelligence no longer just drafts emails and tidies spreadsheets; it fields questions about stress, sleep and how to get through a difficult week. It is open at 3am, never fully booked and never raises an eyebrow.

Our Work-Life Wellness 2027, which polled 5,021 employees in ten countries, found that half of the most AI-fluent workers ("Pioneers") already use AI regularly for their own wellbeing. Some 85% of Pioneers have tried it at least once. Outside the office, estimates of how many AI users turn to it for mental health range from 27% to 63%.

That does not mean AI is becoming a therapist. The same technology employees turn to for relief is also adding to the strain: 30% of workers say AI has increased their workload, and 47% of Pioneers say it has reduced meaningful interaction with colleagues (54% in Britain). 

For HR leaders, the question is not whether staff should use AI for their mental health. Many already do. It is whether the support around them is good enough.

Convenience is the point

Not every wellbeing question needs a professional. A breathing exercise before a difficult meeting, or a plan for winding down before bed, hardly justifies an appointment. AI answers instantly, and adds nothing to the to-do list.

When a professional is needed, the queue is often long. According to the World Health Organisation, between a third and a half of people with mental-health problems in rich countries do not get timely treatment. 

The UK does better than most at the front door: 85% of referrals to NHS Talking Therapies were seen within six weeks in June 2026. But six weeks is a long time when a chatbot replies in ten seconds. In a poll by KFF, 65% of those using AI for health questions cited speed as the main reason; 19% said they could not afford a professional.

In other words, people are not always choosing between AI and professional care. Often they are choosing between AI now and something else later, if it comes at all. 

If the company's own support requires two approvals, an eligibility explainer and a three-week wait, the chatbot wins by default. "Patients like the convenience, and they like the feeling of being heard," says Sabine Wilhelm, a psychologist at Harvard Medical School, in Harvard Medicine. "We have to meet them where they are."

We recommend you reading:AI is making work lonelier for the people who use it most

Feeling supported is not the same as being helped

Convenience brings a subtler problem: a pleasant exchange is easily mistaken for good care. A survey by the American Psychological Association of more than 1,200 psychologists found that 77% have clients who have discussed their use of AI, and a third say some clients treat it as an extra mental-health professional.

What makes a chatbot easy to talk to is not always what makes therapy work. General-purpose models tend to agree with their users. "Technology was built to optimize engagement, not mental health," says Nina Vasan, a psychiatrist at Stanford University. Therapy needs a little friction: a clinician who questions an interpretation or spots a pattern, rather than merely keeping the conversation comfortable. 

Among APA psychologists whose clients had formed a relationship with a chatbot, 36% noticed signs of dependence.

Nor is satisfaction proof of efficacy. In Wellhub's survey, 30% of employees said they have received an unhelpful or harmful response from an AI tool while seeking wellness support. "It made me feel better" and "it gave me reliable help" are not the same thing. 

Those keenest on the technology are also the least likely to question it: 34% of Beginners say they would not be comfortable using AI for wellbeing support, against just 3% of Pioneers.

Part of the problem is which tool people pick. In YouGov, 74% of those who had used AI as a therapist had used ChatGPT. Just 12% had used one designed for mental health. Purpose-built tools, trained and tested with clinicians, look more promising: in a trial published in 2025, Therabot, developed at Dartmouth College, cut depression symptoms by 51% on average compared with no treatment. As Ms Wilhelm puts it, clinical tools and general-purpose chatbots "are two very different things".

You might also find interesting:Employee happiness: the business metric most UK employers still don't measure

From insomnia to something worse

A wellbeing conversation can change course without warning. 

An employee asks how to sleep before a big presentation; three exchanges later the talk has drifted to anxiety, and occasionally to something darker. Dr Torous, a psychiatrist at Harvard, offers an example that could come from any office: "Imagine someone writing a goodbye letter to colleagues. Is this a planned retirement or a suicide note? The chatbot doesn't know."

A study by J.M. Santos and colleagues, published in Practice Innovations, tested six large language models against simulated high-risk scenarios and concluded that none could be considered clinically safe by default. That's why UK employers need to make sure the mental-health resources it offers are clinically backed, such as the digital options in Wellhub's catalogue.

None of this makes every AI wellbeing chat dangerous. It means boundaries matter. AI seems most useful when the task is narrow: learning the basics, practising exercises already worked on in therapy, or organising questions before an appointment. The trouble starts when it becomes the only place someone goes for a diagnosis, a treatment or a crisis. Employers need not police which category each question falls into. They do need to make sure staff know what to do when AI is not enough.

Private is not the same as protected

Employees also turn to AI because it feels "private". Asking a chatbot is easier than admitting to your manager that you are struggling. But feeling private and being protected are different things. In KFF's poll, 41% of those using AI for health had uploaded personal medical information, such as test results, while 77% of the public say they are worried about the privacy of such data.

In the UK,  health information is "special category" data under the UK GDPR, which can be processed only under strict conditions. That leaves employers in a delicate spot, and not only because of the tools their staff use. Some vendors propose using AI to combine surveys, smartwatch data and even emails to spot signs of stress. Even enthusiasts see the limit. "Programs that prioritize personalization, privacy and voluntary participation generally achieve higher engagement levels," writes Anita Gupta, a physician and health-care executive, in Forbes. 

The line between caring and monitoring is thin, and employees notice where it falls.

The better course is to set clear rules before anyone needs them: which tools are approved, what happens to information entered into them, when a general-purpose chatbot is the wrong place to share health data, and what protections the company's own wellbeing services offer. Nobody should have to read the terms and conditions to trust the system.

A first step, not a dead end

The most important question is what employees do after asking AI for help. 

KFF found that 58% of those who used it for mental-health information did not then speak to a doctor or other professional. Not all of them needed treatment. But the figure shows why employers should not assume AI leads to better support. Sometimes it is a bridge. Sometimes it is where the journey ends.

Demand for professional help is not the problem. Almost half of global employees say they currently see a therapist or counsellor in some form. The challenge is to make sure the chatbot conversation leads towards that help rather than replacing it. Clinicians have adjusted their stance: the APA has published guidance for psychologists on discussing AI use with clients, rather than telling them to stop.

Employers can do something similar. Neither monitor every personal use of AI nor pretend it is not happening, but make the next step obvious. If an employee starts with a conversation about stress, do they know how to reach professional support? If the problem is workload, do they know who can change it?

Some problems a chatbot cannot fix

AI can suggest how to organise an overwhelming workload, but it cannot shrink it. It can help someone rehearse a difficult conversation with their boss, but it cannot make the boss listen. If someone is exhausted because priorities change every week, a breathing exercise will not get them far.

That is why managers remain part of the infrastructure of wellbeing. They need not become therapists or diagnose anyone. They do need to make it safe for people to put their hands up, to know where professional help is, and to fix what they can about the work itself. 

Our report finds a 29-point gap here: 60% of Pioneers say their manager actively supports their wellbeing, against 31% of Beginners. The aim of automation should be to strip away the admin around human support, not to replace it.

Your wellbeing programme is competing with friction, not with AI

AI reveals what people want when they need support: quick answers, privacy, few obstacles and the freedom to ask without being judged.

A wellbeing programme can offer excellent therapy, exercise and mindfulness, but if employees do not know what is available, or must navigate a maze of portals to reach it, the quality of the offer matters less. The data bear this out: only 26% rate their company's wellbeing programme as good, down from 29% last year and 41% two years ago, and 46% say their company does not give them the time and space in the workday to look after their wellbeing.

It also helps to offer several front doors, rather than expecting people to know in advance what they need. Stress can stem from workload, sleep, movement, connection or mental health, and is usually a mix. Wellhub brings gyms, studios, classes, online therapy, mindfulness, nutrition and sleep into a single membership, with options that fit the gaps in a real working day.

At companies that offer it, 71% of employees say wellness is genuinely built into how their organisation operates, compared with 32% of those with another programme. AI can help someone work out what they need before taking the first step. It should not be the only door.