Most professionals have tried an AI tool at work. Fewer have actually changed how they work because of one. AI adoption happens when a tool handles a specific, high-frequency task reliably enough that the habit forms without effort. The gap between trying and adopting is larger than usage statistics tend to suggest.
According to Gallup's workplace research, 45% of US employees now use AI at least a few times a year, but only 10% use it every day. Among people who do use AI tools at work, just 16% strongly agree those tools are actually useful for their job.
There is enough evidence that AI works when used well, but there’s a deeper question about why so few people have reached that point.
For a B2B sales rep or account manager whose output runs through their inbox, the gap between trying an AI tool and actually changing how they work is especially significant.
Starting with the wrong question
For sales reps and account managers managing high email volume, many AI rollouts share a recognizable format. For instance, someone senior will decide the company needs to act on AI, so new tools get procured, and the practical question of what employees should actually do with them gets left unanswered. In this case,usage usually stays low.
Gallup's research points to the most common reason: the main barrier is not technical complexity or employee resistance, but an unclear value proposition. People don't always know what they're supposed to do with these tools, or why it matters for their specific job.
That is a strategy problem, rather than a technology one. Tom Lorimer, co-founder and CEO of AI laboratory Passion Labs, puts it directly:
"Efficiency and productivity are means to an end. What is the end state? Walk backward from your core mission. When you can do that, you'll have a much clearer idea of what your AI strategy should look like."
Without that clarity, tools get adopted in name only.
When everyone is using AI differently
Even in organizations that have made a genuine effort, a structural problem often goes unaddressed. Different people use different tools in different ways, with no consistency across teams. One person prompts ChatGPT for meeting notes. Another has a Notion integration nobody else touches. A third tried something six months ago and stopped.
This fragmentation is part of why AI adoption has not moved macro productivity numbers the way many predicted. Research on firm-level AI adoption is instructive here: a study published in the Journal of Economic Behavior and Organization found that companies using AI intensively sold on average 13.9% more than non-users, but that effect was tied to sustained, integrated adoption across the business, not to individual or ad hoc use. Individual experimentation does not accumulate into organizational performance, and the gains generally stay personal and uneven.
Tom describes what a working alternative looks like:
"All you're getting is basically a bunch of people all using chatbots differently. Whereas something like Fyxer is deeply embedded within email and consistent across an organization."
That consistency is what produces compounding value. It is also what most approaches to productivity at work miss.
Why team-wide AI adoption is different from personal AI adoption
When an individual tries an AI tool and stops using it, the cost is generally pretty low. But when an organization deploys one to hundreds (or thousands) of employees and sees low uptake, the reasons are worth understanding properly, because they’re usually not the same reasons.
Research on workplace AI deployments consistently identifies three specific barriers that go beyond the "unclear use case" problem. A study of a Digital Productivity Assistant rolled out across 28 knowledge workers by Cranefield et al. (2023) found that even employees who could see the value of the tool in principle did not adopt it in practice. The barriers were concrete: the tool's model of how work is organized did not match how workers actually experienced their days.
One example from the study is worth noting. The tool categorized all time as either "focus time" or "collaboration time." For academics, this was meaningless. Their research was collaborative. Their teaching involved focus. The binary made the tool's outputs feel irrelevant to them, and so they stopped consulting it. One participant described the tool as giving her a reflection of her Outlook calendar, which she said was not a reflection of her work life. The DPA showed her 68% of her time as available for focus work, while she experienced her days as overcommitted. That gap between what the tool reported and what she lived made engagement with it feel like extra work rather than a reduction of it.
The Microsoft 365 Copilot study across 357 employees at a European multinational found a parallel issue. Employees who used the tool less than they might have did so partly because they were uncertain what data was safe to input. As one respondent put it, they defaulted to a "rather safe than sorry" approach, inserting only information they were confident was not sensitive. Uncertainty about the tool's boundaries was producing conservative use, which meant the tool delivered less value, which reinforced the sense that it wasn't worth engaging with more fully.
These are not problems that more training necessarily resolves. They reflect a mismatch between the tool's assumptions and the user's reality. A tool that requires workers to adapt their behavior significantly in order to use it is asking for a high upfront cost in exchange for uncertain returns. Most people, reasonably, decline that trade.
The cost that doesn’t appear on any report
One reason AI adoption is easier to deprioritize than other investments is that the problem it addresses is largely invisible. Organizations can see software costs, headcount, and office space. The time their employees spend managing email and meeting follow-up is rarely tracked as a cost at all.
The research on this is more specific than the general conversation about inbox overload tends to be. A study by Mark et al. (2016) tracked 40 information workers over 12 working days using computer logging and biosensors. Workers spent an average of nearly 90 minutes per day in their email client alone, checking it an average of 77 times a day. Longer time on email was directly correlated with lower perceived productivity and higher measured stress. A separate field study found that knowledge workers opened their email inbox on average every 5.3 minutes during the working day, and that four out of six participants spent more time in their email application than any other single tool on their desktop.
The cost is real, but it doesn't appear in any budget line. According to the Fyxer Admin Burden Index 2026, avoidable admin costs UK and US organizations $954 billion every year). IIt shows up in slower responses to higher-value work and in the friction that builds when communication management competes with the work it's meant to support.
The effort problem with general-purpose AI tools
There is a failure mode that looks like progress. Someone starts using a general-purpose AI tool, invests time learning to prompt it, builds a few use cases, and finds they are spending more time managing the tool than they save with it.
General-purpose tools ask you to supply the specificity yourself. You translate your problem into a prompt, evaluate the output, and iterate. For complex or one-off tasks, that investment can pay off. For the repetitive, high-frequency work that fills most professional days, it adds friction rather than reducing it.
Saving ten minutes on an email draft is only valuable if those ten minutes go toward something useful. For that to happen consistently, the time-saving itself has to be reliable and low-maintenance, not something that requires active management to sustain. According to Fyxer's 2026 Admin Burden Research, 2 in 3 employees describe their AI tools as partial, ineffective, or insufficient. The ambition to use AI is there. A tool specific enough to deliver consistently, for most people, is not.
What genuine AI adoption looks like in practice
The evidence on what AI adoption produces, when it works, is reasonably clear. A randomized controlled trial at Microsoft tested M365 Copilot across 310 subjects on standard information worker tasks: retrieving information from files, emails, and calendars; catching up after a missed meeting; and drafting content. Users with access to the AI tool completed tasks 26.6% faster than the control group, with no significant difference in accuracy.
Willingness to pay was also higher among users who had tried the tool than among those who had only heard about it, suggesting the experience exceeded prior expectations rather than falling short of them.
The Cranefield et al. study of Digital Productivity Assistants points to what separates people who reach that outcome from those who do not. Among the 28 knowledge workers in that study, a subset engaged fully with the tool and reported genuine improvements to their working patterns.
What they shared was a working understanding of how the tool collected and categorized its data. One participant, working in IT management, described it as a “dumb tool” he could nonetheless extract real value from, because he understood what it was measuring. He adjusted his calendar practices to match, and used the tool’s output to establish a team norm around protected focus time, making those blocks visible to colleagues so meetings wouldn’t be booked across them. The tool didn’t produce that outcome on its own. It was his understanding of how it worked that made it possible.
Fyxer’s consumer data reflects a consistent pattern. Professionals who describe AI as genuinely useful concentrate their usage on email and meeting follow-up; the tasks where time is spent in the largest and least-tracked quantities. The tool runs without requiring ongoing configuration or maintenance. That low-maintenance quality is not incidental to adoption; it is a significant part of what makes it last.
Tom describes this as the difference between AI as a product and AI as a utility:
"It's fundamentally a utility, in the same way that you would put wiring and pipes through a building. It needs to have that same embedded format within an organization."
When it works that way, you notice the outcomes, not the tool. Your morning starts with an organized inbox and draft replies already waiting. The meeting summary is done. The noise has been filtered out.
Why understanding the technology reduces resistance
A significant part of why people do not adopt AI is uncertainty rather than skepticism. They’re not sure what the tool is doing, not sure whether to trust its outputs, and not sure what to do when it gets something wrong. In that state, most people default to not relying on it.
Some of that uncertainty clears when people understand roughly how the technology works, not at an engineering level, but enough to know what it is good at and where it falls short. Tom's experience rolling out AI across organizations points consistently to the same finding:
"When people understand how these models are trained, that it's not magic, it's machine learning, they appreciate not only how brilliant they are, but also their limitations."
The Copilot research supports this. Employees who understood enough about the tool to know what questions to ask of it, and what data it was drawing on, were significantly more likely to engage with it and report value. Those who remained uncertain about its scope tended to use it less, and therefore got less from it. Understanding is not a prerequisite for adoption, but it tends to accompany it.
The difference between trying AI and adopting it
Trying an AI tool and adopting one are different things. Trying it means you have seen what it can do. Adopting it means it has changed how you work in a way you would not want to give up.
Most professionals are still on the trying side of that line. Getting across it comes down to one thing: finding a tool built for a specific, high-frequency problem that fits into how you already work rather than requiring you to change. That is harder to find than the volume of available AI tools would suggest. But it is the condition that produces lasting change.
