Why AI Feels Useless at Work
My father worked at a local branch of a railway company in his entire professional career. For a bureaucratic institution, it adopted new technology surprisingly fast. Or, more precisely, it was fast at buying the machines.
I remember those unofficial “take your child to work” days in the 1990s. Windows 3.1 was already out, so I could open Paintbrush, draw the nine (now eight) planets of our solar system, and print them on a dot-matrix printer. That was my first memory of using a computer. Later, once I could read and write, I typed out classical poetry in Microsoft Word and printed that too.
To the employees, though, computers were peripheral. The typists used them seriously. Everyone else typed up the occasional document. A general-purpose computing platform, and the killer use case was “slightly better typewriter”.
Was this because computers were useless? Obviously not. It was because the daily workflows had not been redesigned around computers. Work still moved through paper forms, rubber stamps, and inter-office envelopes. The computers sat on office desks like expensive furniture with keyboards.
New tools, old habits
This pattern is similar to many technology transitions in the past 15 years. Adopting a new tool is the easy part. Getting people to stop using it exactly like the old one is the real work. Teams get version control and use it as a backup system. People move to a new programming language and write code that looks exactly like the old one with different syntax. The technology works fine. The workflow doesn’t change.
The same thing is happening with AI now. I notice it most clearly in my own behavior.
I have been using agentic AI tools for coding (both open source and proprietary) for about 18 months. For this use case, it has been genuinely transformative. The gap between intent and working code has narrowed in ways I would not have believed three years ago.
But for everything else? My AI use amounts to: drafting emails, polishing documents, sanity checking ideas. Useful, certainly. But useful in the way that typing classical poetry in Microsoft Word was useful: a real improvement in execution, while the work itself goes unquestioned.
I don’t think I’m unusual. Ask people at most organizations how they use AI at work and you will hear the same short list: summarizing meetings, drafting messages, generating first drafts that need heavy editing anyway. The tools are capable, and people are trying. Yet the prevailing feeling feels more like convenience than transformation.
Not the tool’s fault
The instinct is to blame the technology. AI hallucinates, doesn’t understand our domain, can’t be trusted with anything important. These are real limitations. But I don’t think they explain the gap between what AI can do and what it actually does for most of us. AI is already capable enough to write working software, pass professional licensing exams, and reason through problems that would challenge a specialist. The capability is there. We lack the same thing my father’s colleagues lacked when they used their PCs as typewriters: workflows designed around what the technology makes possible, rather than what the previous technology required.
Think about what happened when workflows were eventually redesigned around computers. We didn’t just type faster. Spreadsheets didn’t speed up accountants but they let one person do what previously required a team. Databases didn’t make filing faster but they made filing obsolete. The structure of work changed, not just one step in the existing process.
For AI, most workplaces are still waiting for that structural change. AI-assisted email is still email. AI-summarized meetings are still meetings. The workflow is untouched; one step got slightly faster.
The hard questions
Having been through technology transitions before, I know what makes the next phase hard: our willingness to question the workflow itself. In my own field, when the industry shifted to open source technology foundations, the real argument was never about programming languages. It was about whether a statistical analysis should be a static document or a reproducible computation. That was a workflow question disguised as a tooling question, and it took years to settle.
I suspect the equivalent questions for AI are harder, because we don’t know what the redesigned workflows look like. With previous transitions, we at least knew the destination, such as reproducible research and version-controlled collaboration, had been articulated as goals years before the infrastructure caught up. With AI, the destination is unclear. We know the current workflows are underusing the technology. We just don’t know what the better ones are, yet.
That uncertainty is uncomfortable. But if the history of technology adoption teaches anything, it is that the organizations that eventually benefit most rarely turn out to be the fastest adopters. They tend to be the ones willing to let the tools reshape how they work, even when that means admitting the old way was more habit than necessity.