Two employees can use AI all day and be doing very different things. One dictates an email, checks the transcription and sends the words they intended to write. Another asks a system to choose a response to a customer complaint and accepts its recommendation. An activity report might call both employees frequent users. It tells us little about who did the thinking, who made the decision or what would happen if the output were wrong.
That distinction matters to the debate about the AI utilisation gap. In Workplace Insight, Richard Colwell recently explored why people use AI more broadly in their personal lives than at work. His argument raises a further question: what do employers actually want to learn when they measure use? Counting interactions can show whether a tool is being opened. It cannot, by itself, show whether a barrier has been removed or a responsibility has been delegated.
Accommodation makes this particularly clear. A person may use an AI-enabled writing tool to reduce the physical effort involved in getting their own words onto a screen. The assistance may be frequent precisely because the underlying barrier is frequent. Treating that activity as evidence that the person is surrendering more of their work to a machine would confuse access with authorship. Treating it as proof of advanced AI capability would be equally premature.
A useful assessment begins with the task. Transcribing a person’s spoken words, suggesting alternative phrasing, summarising a document and recommending a decision are different interventions. Even within a single application, those functions can be mixed together. The relevant question is what changed between the person’s input and the work that emerged.
Is it accurate?
Consider a dictated message. If the system transcribes it, the review concerns whether the words were captured accurately. If it also rewrites the message, the review must include meaning, tone and any added claims. If it invents a reason for declining a request, something more consequential has happened. The tool’s name and the number of times it was used do not reveal those differences. Nor does a broad label such as ‘AI-assisted’.
Decision authority is the next part of the picture. There is a difference between receiving a suggestion and allowing that suggestion to determine an outcome. A human approval step deserves scrutiny of its own: can the reviewer see the source material, recognise an error and reject the recommendation? Does the workflow leave enough time for that review? An approval button is a weak safeguard when the person pressing it cannot meaningfully assess what they are approving.
This shifts attention from how busy the software appears to the actual division of responsibility. An employee who uses assistance throughout the day may retain control over every consequential choice. Someone who uses it only once may allow a consequential recommendation to pass without examination. Frequency alone cannot distinguish them.
Reversibility adds another dimension. A rough paragraph in a private draft can be corrected before anyone relies on it. A message already sent to a client is harder to repair. A decision affecting another person’s opportunity may cause harm that a later correction cannot fully undo. Even a draft can carry consequences if confidential information has already been entered into an unsuitable service. Reversibility therefore concerns the whole process, including the input, rather than just whether there is an undo button.
These distinctions also suggest a more careful approach to accommodation privacy. Employers need enough information to manage the work and the technology responsibly. A general AI-use register need not become a place where colleagues explain medical histories. The operational record could describe the permitted function, information-handling restrictions and review responsibility, with any necessary personal information handled through an appropriate confidential process.
An employee could then explain that an approved tool supports transcription and that they check the final text, without having to explain a diagnosis to everyone who sees the record. This is a proposed design principle, not a substitute for the organisation’s specific responsibilities. Its value is that it separates the information needed to assess a workflow from personal details that a broad readership does not need.
Is it safe?
None of this makes accommodation use automatically safe. Accuracy, confidentiality and accountability still matter. Equally, those concerns should be examined through the actual task, rather than assumed from the fact that assistance was used. A blanket judgement in either direction leaves too much unexplained.
Employers will still need adoption data for practical reasons, including understanding whether people can access and use the tools provided. But that data needs interpretation. A more informative account would connect use to the task performed, the authority retained by a person and the consequences of error. Evidence of benefit would then come from the work: whether someone could complete it, whether the result was sound and whether mistakes could be caught before they mattered.
The workplace AI debate often asks how much further people can go with the technology. Before asking for more use, leaders should be able to explain what the existing use means. Sometimes the important change is that a person can get their own ideas into the world with less friction. That deserves to be understood on its own terms.
Morissa Schwartz is a writer, publisher and founder of GenZ Publishing and Dr. Rissy’s Writing & Marketing.
October 12, 2026
What AI activity counts leave out
by Morissa Schwartz • AI, Comment
That distinction matters to the debate about the AI utilisation gap. In Workplace Insight, Richard Colwell recently explored why people use AI more broadly in their personal lives than at work. His argument raises a further question: what do employers actually want to learn when they measure use? Counting interactions can show whether a tool is being opened. It cannot, by itself, show whether a barrier has been removed or a responsibility has been delegated.
Accommodation makes this particularly clear. A person may use an AI-enabled writing tool to reduce the physical effort involved in getting their own words onto a screen. The assistance may be frequent precisely because the underlying barrier is frequent. Treating that activity as evidence that the person is surrendering more of their work to a machine would confuse access with authorship. Treating it as proof of advanced AI capability would be equally premature.
A useful assessment begins with the task. Transcribing a person’s spoken words, suggesting alternative phrasing, summarising a document and recommending a decision are different interventions. Even within a single application, those functions can be mixed together. The relevant question is what changed between the person’s input and the work that emerged.
Is it accurate?
Consider a dictated message. If the system transcribes it, the review concerns whether the words were captured accurately. If it also rewrites the message, the review must include meaning, tone and any added claims. If it invents a reason for declining a request, something more consequential has happened. The tool’s name and the number of times it was used do not reveal those differences. Nor does a broad label such as ‘AI-assisted’.
Decision authority is the next part of the picture. There is a difference between receiving a suggestion and allowing that suggestion to determine an outcome. A human approval step deserves scrutiny of its own: can the reviewer see the source material, recognise an error and reject the recommendation? Does the workflow leave enough time for that review? An approval button is a weak safeguard when the person pressing it cannot meaningfully assess what they are approving.
This shifts attention from how busy the software appears to the actual division of responsibility. An employee who uses assistance throughout the day may retain control over every consequential choice. Someone who uses it only once may allow a consequential recommendation to pass without examination. Frequency alone cannot distinguish them.
Reversibility adds another dimension. A rough paragraph in a private draft can be corrected before anyone relies on it. A message already sent to a client is harder to repair. A decision affecting another person’s opportunity may cause harm that a later correction cannot fully undo. Even a draft can carry consequences if confidential information has already been entered into an unsuitable service. Reversibility therefore concerns the whole process, including the input, rather than just whether there is an undo button.
These distinctions also suggest a more careful approach to accommodation privacy. Employers need enough information to manage the work and the technology responsibly. A general AI-use register need not become a place where colleagues explain medical histories. The operational record could describe the permitted function, information-handling restrictions and review responsibility, with any necessary personal information handled through an appropriate confidential process.
An employee could then explain that an approved tool supports transcription and that they check the final text, without having to explain a diagnosis to everyone who sees the record. This is a proposed design principle, not a substitute for the organisation’s specific responsibilities. Its value is that it separates the information needed to assess a workflow from personal details that a broad readership does not need.
Is it safe?
None of this makes accommodation use automatically safe. Accuracy, confidentiality and accountability still matter. Equally, those concerns should be examined through the actual task, rather than assumed from the fact that assistance was used. A blanket judgement in either direction leaves too much unexplained.
Employers will still need adoption data for practical reasons, including understanding whether people can access and use the tools provided. But that data needs interpretation. A more informative account would connect use to the task performed, the authority retained by a person and the consequences of error. Evidence of benefit would then come from the work: whether someone could complete it, whether the result was sound and whether mistakes could be caught before they mattered.
The workplace AI debate often asks how much further people can go with the technology. Before asking for more use, leaders should be able to explain what the existing use means. Sometimes the important change is that a person can get their own ideas into the world with less friction. That deserves to be understood on its own terms.
Morissa Schwartz is a writer, publisher and founder of GenZ Publishing and Dr. Rissy’s Writing & Marketing.