Working Next to the Machine: What AI Assistants Actually Changed About a Workday

Ask someone whose job changed because of AI what actually happened, and you rarely get the story the headlines promised. Nobody says "the machine took half my work." What they say is stranger and more specific: "I don't write first drafts any more." "I stopped dreading the inbox." "The part I was good at is now the whole job, and I'm not sure I like it."

That last sentence is the real story of assistants at work. They did not remove jobs at the rate the forecasts suggested. They removed tasks — and tasks are not evenly distributed inside a role. Pull out the routine forty per cent of a job and what remains is not a smaller job. It is a denser one.

The tasks that went first

The work that assistants absorbed most completely shares a profile: it has a clear input, a conventional output format, a tolerance for revision, and a cost of error that is embarrassment rather than catastrophe. Meeting summaries. Boilerplate correspondence. First-pass code for a well-understood problem. Translating a spreadsheet into a paragraph and a paragraph into a spreadsheet. Reformatting. Renaming. The long tail of getting something into the shape where the real thinking can start.

None of that work was ever the point of anyone's job. It was the tax paid to reach the point. But it was also, quietly, where new people learned. A junior analyst who never assembles the raw deck by hand may never develop the intuition for when the numbers are wrong. This is the most serious open question in the workplace right now, and it does not show up in productivity statistics for another decade.

Automation removed the rungs of the ladder that nobody enjoyed climbing — and then we noticed they were rungs.

The verification tax

The recovered hours are real, but they are gross rather than net. Every assistant output arrives with an obligation attached: someone has to decide whether it is right. And checking is a different cognitive mode from producing — less creative, more vigilant, and far harder to sustain for long stretches.

Teams that handled this well made the verification explicit. They decided, per task type, what level of checking was required: none for a meeting summary nobody will act on, line-by-line for anything touching money, law, health or production systems. Teams that handled it badly left it to individual judgement under deadline pressure, which is a reliable recipe for the one unchecked output that becomes an incident report.

The pattern that works is unglamorous: treat machine output as a competent draft from a fast, tireless colleague with no accountability and no memory of the consequences. You would not publish that colleague's work unread. You also would not refuse their help out of principle.

What it replaced

Three things, mostly. The blank page — which sounds trivial and is not, because a substantial fraction of professional procrastination is the friction of starting. The specialist bottleneck, where a whole team waited on the one person who could write SQL or read a contract. And a category of outsourced micro-work: transcription, basic translation, formatting, simple research summaries, much of which used to be somebody's paid job.

That last displacement is the one with a face attached, and it deserves to be named rather than folded into an aggregate. The freelance transcriptionist and the entry-level copywriter did not experience a productivity gain. They experienced a market disappearing.

The hours that did not change

It is worth being precise about the limits, because the limits are stable and predictable. Assistants have barely touched:

• Work whose difficulty is political rather than informational. Getting two departments to agree on a number is not a text-generation problem. No summary has ever resolved a turf dispute.
• Work that requires accountability. Someone has to sign. Signature is not a task; it is the acceptance of consequence, and it cannot be delegated to a system that cannot be fired.
• Work with physical presence at its core. The trades, care work, most of logistics, most of hospitality. The economy's largest employment categories were the least affected, which is the opposite of what early commentary predicted.
• Work whose value is that a person did it. A condolence note. A performance conversation. A pitch where the point is that you cared enough to prepare.

What it cost

Beyond displacement, three costs are showing up consistently in the organisations that adopted assistants earliest.

Volume inflation. When producing a document becomes cheap, more documents get produced. Several companies report that the reading load rose faster than the writing load fell. A six-page memo generated in four minutes still takes twenty minutes to read, and it is now competing with eleven others.

Skill hollowing. The capabilities you stop exercising, you stop having. This is not new — few people can navigate by paper map any more — but it matters more when the skill in question is the ability to notice that an answer is wrong.

Ambient monitoring. Assistant tooling sits inside the same systems that log activity. The step from "summarise this meeting" to "score participation in this meeting" is short, technically trivial, and being taken. Employees notice this faster than employers expect, and it poisons adoption of tools that would otherwise be welcome.

The honest summary

The workday did change. It changed less in total hours than in texture: fewer stretches of low-effort assembly, more stretches of judgement, decision and review, with the boring parts that used to serve as mental rest removed. For people who liked the judgement parts, this is a good trade. For people whose job was mostly the assembly, it was not a trade at all.