The most productive AI users are not simply getting more work done. Many are starting to prefer AI as a source of help, feedback and support. New research suggests there is a practical reason: asking AI carries almost none of the social cost of asking another person.
One reason AI is becoming a preferred work partner may have little to do with whether it is actually better than a human colleague.
It is simply easier to ask.
A 2025 survey of 2,500 executives, employees and freelancers from Upwork found that among employees reporting the largest productivity gains from AI, 67% said they trusted AI more than their coworkers and 64% said they had a better relationship with AI than with their human teammates. More than half, 54%, considered AI more empathetic, while 85% said they were more polite to AI than to the people around them.
The finding needs an important qualification. It does not describe every employee in the study. The comparison focused on full-time employees in the top quartile for self-reported AI-related productivity gains.
Even with that limitation, 67% is a striking number.
The interesting question is not whether workers have suddenly developed genuine friendships with software. It is why people who have become especially effective at working with AI appear so comfortable turning toward it instead of toward another person.
A growing body of workplace research points to a surprisingly ordinary explanation: human help comes with social friction. AI help often does not.
Asking a coworker has a cost that asking AI does not
Most people know the small calculation that happens before asking someone at work for help.
Is this person busy? Should I already know the answer? Will I look inexperienced? Am I interrupting them? Will I owe them something later? Have I asked too many questions this week?
None of those concerns requires a toxic workplace. They are normal parts of working with other people.
AI changes that calculation.
You can ask the same question five times. You can ask something basic without exposing what you do not know. You can send an unfinished idea without worrying that someone will judge it. There is no colleague to interrupt and no favor to return.
Researchers are now finding evidence that this lower social cost changes help-seeking behavior.
A 2025 Academy of Management study on AI and workplace help-seeking ran three preregistered experiments examining what happens when employees seek help from generative AI instead of colleagues. The researchers found that positive experiences asking AI for help, including lower feelings of indebtedness and less concern about appearing incompetent, reduced people’s willingness to approach coworkers for help later.
A separate experimental study on AI assistance and employee autonomy, accepted for publication in Computers in Human Behavior, reached a closely related conclusion. Across two experiments, employees receiving certain forms of help from AI reported greater autonomy and lower indebtedness than those receiving help from a human colleague. They were consequently more willing to seek that type of help from AI.
That gives the 67% finding a much more practical interpretation.
AI does not need to become emotionally meaningful for employees to prefer interacting with it. It only needs to become the easiest place to think out loud.
AI is becoming a place to think, not just a place to produce
The early workplace use case for generative AI was largely about output. Workers used it to draft emails, summarize documents, create presentations or write code faster.
That role is expanding.
The 2026 Work Trend Index analysis of more than 100,000 Copilot conversations found that 49% of classified user goals involved cognitive work such as analyzing information, solving problems, evaluating options and creative thinking. Another 19% involved work related to interacting with other people.
That matters because a tool used mainly to produce text is easy to think of as software.
A system used to test an idea, work through uncertainty, prepare for a difficult conversation or ask a question you are embarrassed to ask someone else occupies a different position in the workday.
It starts performing some of the informal functions previously handled by colleagues.
This does not require anyone to believe the AI has feelings. Workers only need to find the interaction useful, private and low pressure.
For companies, that creates an unusual situation. AI can improve an employee’s individual ability to solve problems while quietly reducing some of the interactions through which teams traditionally exchange knowledge.
Coworkers do more than answer questions
Replacing a five-minute interruption with a five-second AI response looks like an obvious productivity gain.
Sometimes it is.
The problem is that asking another person a question often produces more than the answer.
A junior employee asking a senior colleague about a client problem may learn why the problem exists, how the customer thinks and which previous decisions created the current situation. The conversation may expose another issue nobody had noticed. The senior colleague learns what the junior employee finds difficult. Both people build a little more understanding of how the other works.
AI can provide the answer without necessarily producing those secondary benefits.
This is especially important because help-seeking is one of the ordinary ways information moves through an organization. It creates relationships between teams, exposes gaps in knowledge and gives experienced employees opportunities to pass on context that may never have been formally documented.
If workers increasingly resolve routine uncertainty privately with AI, some of those conversations disappear.
The effect will not necessarily be visible on a productivity dashboard. The task still gets completed, perhaps faster than before.
What changes is the network around the task.
Research has already found a connection between AI interaction and loneliness
Concerns about workplace connection predate the latest generation of AI agents.
A 2023 workplace study covering 794 workers across four countries examined the consequences of employees interacting with AI in Taiwan, Indonesia, the United States and Malaysia.
The researchers found a complicated pattern. More interaction with AI was associated with a stronger need for social affiliation, which could encourage workers to help colleagues more. At the same time, greater AI interaction was also associated with higher feelings of loneliness, which were linked to poorer well-being outside work.
That nuance matters.
Working more closely with AI does not automatically make people antisocial. In some circumstances, feeling the absence of human interaction can actually increase the desire to reconnect with other people.
What the research does suggest is that AI interaction is not socially neutral.
Changing who employees talk to during the workday can change how connected they feel, even if their output improves.
The newer survey makes that question harder to ignore because the workers reporting the strongest preference for AI were also the workers reporting some of the strongest productivity gains.
High productivity and a healthy work experience are not the same thing
The same study behind the 67% figure found that the highest-productivity AI users were also experiencing significant strain.
Eighty-eight percent reported burnout, and they were twice as likely to consider leaving their organization. The research also found that this group was unusually disconnected from the company’s broader AI strategy, despite being among the people getting the most value from the technology.
Those findings are correlations from survey data. They do not prove that trusting AI causes burnout or that stronger relationships with coworkers would automatically solve it.
But they raise an important management question.
A company can look highly successful at AI adoption while its most capable AI users become increasingly self-contained.
Those employees may appear especially productive because they can solve more problems without asking anyone else. They may write faster, research faster and make decisions faster. Managers can easily reward that independence.
Taken far enough, however, independence can become isolation.
The person no longer needs to wait for a colleague, but they may also stop learning what colleagues know. They receive immediate feedback, but less of the contextual feedback that comes from someone who understands the company, customer and history of the work.
AI can make an employee less dependent on their team without making the team less important.
There may also be a feedback loop between AI use and coworker behavior
The relationship can move in the other direction too.
Employees may turn to AI because asking colleagues feels uncomfortable. Their visible use of AI may then change how colleagues treat them.
A 2025 study on how coworkers interpret employee AI use combined two experiments, a field survey and a field experiment. The researchers found that when coworkers interpreted someone’s AI use as an attempt to avoid effort, they judged that employee more negatively and became less willing to help them.
This does not mean coworkers generally punish people for using AI. The negative response depended on how they interpreted the reason for using it.
Still, it creates the possibility of a reinforcing cycle.
An employee uses AI because it feels safer than asking a colleague. Coworkers interpret extensive AI use as laziness or reduced effort. They become less willing to help. AI then becomes even more attractive as a source of support.
That possibility also connects with another problem Flexable has examined: many employees prefer not to reveal how much AI contributes to their work.
A workplace can therefore end up with two parallel systems of collaboration. One is visible, involving meetings, managers and coworkers. The other is private, happening between individual employees and AI systems throughout the day.
The second system can become important long before management has any useful way to measure it.
Trusting AI is not the same as AI being trustworthy
There is another reason to interpret the 67% cautiously.
People can feel comfortable with a system without being good at judging whether its answers are reliable.
Research increasingly shows that presentation itself can influence trust in AI.
An experiment involving 778 people collaborating with the same type of large language model found that presenting the system as an AI teammate rather than as a tool significantly increased trust and changed the tone people used when communicating with it.
Other experiments have reached related conclusions. A 2025 study of 1,155 participants examining humanlike AI teammates found that making AI appear more human could increase initial trust and willingness to work with it under some conditions.
Those findings do not mean people are easily fooled.
They show that trust is partly social.
Communication style, framing and human-like behavior can affect how we feel about an AI system independently of the quality of a particular answer.
That is normal human behavior, but it creates a workplace risk. An employee who says they trust AI more than coworkers may be describing convenience, predictability, privacy or interpersonal comfort as much as factual reliability.
Companies should not treat greater employee trust in AI as evidence that the technology deserves greater autonomy.
The attraction of AI is that it cannot embarrass you
The strongest explanation for AI’s appeal may also be the simplest.
Workplaces contain status.
A new employee asks a senior employee. A direct report asks a manager. Someone from finance asks engineering. A person who wants a promotion asks a question that might expose what they do not know.
Those interactions are rarely completely neutral.
AI strips much of that away.
It does not remember that you asked an elementary question last month and bring it up during promotion discussions. It does not get irritated because you interrupted deep work. It does not expect you to return the favor. It does not think less of you because your first idea was poor.
For many tasks, that is a feature.
A worker can use AI to prepare before speaking to another person, which may actually improve human collaboration. Someone can learn the basics privately, arrive at a meeting better informed and reserve colleague time for questions that genuinely require experience or context.
That is probably the healthier interpretation of AI as a work partner.
The concern begins when preparing to talk to people becomes a substitute for talking to people.
Freelancers offer an interesting contrast
The original survey also included 625 freelancers, and their relationship with AI looked different from that of the most productive full-time employees.
Freelancers reported productivity gains from AI too, but they were more likely to describe the technology as a learning partner. Ninety percent said a major advantage was learning faster, 42% credited AI with helping them specialize, and 88% said AI had positively affected their careers.
The researchers also found a greater sense of control among freelancers. Thirty-four percent strongly agreed that AI gave them an additional advantage in their work, compared with 28% of full-time employees. Meanwhile, 48% of full-time employees said their employer was asking too much of them when it came to integrating AI.
That does not prove independent work creates a healthier relationship with AI. The groups differ in many ways, and the evidence is based on self-reported survey responses.
But autonomy may matter.
A freelancer generally chooses when a tool is useful in the workflow and remains directly responsible for the client outcome. An employee may be using the same technology inside a system of management targets, corporate mandates and changing expectations.
As independent work becomes a larger part of the labor market, a trend covered in Flexable’s research on the global rise of freelancing, that distinction deserves more research.
The important variable may not be how much someone uses AI.
It may be how much control they have over the reason they are using it.
Companies may be measuring the wrong side of AI adoption
Most organizations still measure AI success through access, usage and output.
How many employees have licences? How often do they use the tools? How much time do they save? How much additional work can they complete?
Those metrics miss what happens between people.
A company that wants to understand the full effect of AI adoption should also know where employees now go when they are stuck. Are junior workers still asking senior employees questions? Are teams still exchanging informal feedback? Are people using AI before asking a colleague, or instead of asking one? Are managers seeing fewer questions because employees are becoming more capable, or because asking the machine feels safer?
There is nothing inherently wrong with AI becoming the first stop for routine questions.
In many organizations, that will be an improvement. Experienced workers should not need to answer the same basic question hundreds of times if a reliable internal AI system can handle it.
The value of the saved interaction should then be reinvested in better human interaction: mentoring, difficult judgment calls, feedback, collaboration and the exchange of knowledge that does not fit neatly into a prompt.
Otherwise a company can remove the small conversations without creating anything better in their place.
What the 67% figure actually measures
The headline statistic comes from research conducted between March 24 and April 9, 2025 across the United States, United Kingdom, Australia and Canada.
The full survey included 2,500 respondents: 1,250 C-suite executives, 625 full-time salaried employees and 625 freelancers. Respondents ranged from 18 to 78 years old and used a computer for work at least some of the time. Separate focus groups included 16 US business leaders and 24 US freelancers who used AI several times a month or more.
For the findings about trust, burnout and preference for AI over colleagues, researchers compared employees in the top quartile for self-reported AI-related productivity gains with other employees.
That wording matters.
The 67% figure should not be described as “67% of workers” or “67% of AI users.” It applies to the study’s highest-productivity employee subgroup.
The research also relies on self-reported productivity and relationships. It cannot establish that AI caused weaker coworker relationships, burnout or intentions to leave.
What it does reveal is still significant.
Among employees who believe AI is helping them the most, a large majority say they now trust the technology more than the people they work with.
Independent research on help-seeking offers a plausible reason why. AI is available immediately, does not judge the question, creates little sense of social debt and allows employees to work through uncertainty privately.
Those are real advantages.
They also make it possible for an employee to have an extremely productive day with very little need to speak to another person.
For the next phase of workplace AI, that may be the more important question to watch. Not simply how much work AI can do, but which human interactions disappear once it becomes easier to ask the machine.