Every August produces the same conversation in slightly different words. Somebody points out that artificial intelligence is moving faster than anyone can follow. Somebody else points out that most of what was promised eighteen months ago has not arrived.
Nicole Junkermann thinks both camps are right, which is why the argument never resolves. An entrepreneur and investor, she records short briefings on artificial intelligence and its effect on everyday work, each one taking a single question at a time.
We spoke to her about what has actually changed this year, what has not, and why the gap between the two explains most of the disagreement about where things stand.
“A capability exists the moment somebody demonstrates it. A habit is different.”
Let’s start with the disagreement itself. Why can’t people agree on whether AI is moving fast or slow?
Because they are measuring different things and neither side says which. The people who say it is moving extraordinarily fast are describing capability. The people who say nothing has arrived are describing habit. Both are looking at real evidence.
A capability exists the moment somebody demonstrates it. A habit exists when people reach for something without deciding to. Those are very different thresholds, and almost all of the interesting movement this year has been in the second category, while almost all of the coverage has been about the first.
Is there one change from this year you would call structural rather than incremental?
The shift from answering to doing.
For several years the mental model for these tools was a very well read colleague who would answer any question instantly, occasionally with total confidence about something they had got wrong. You asked, it answered, and the exchange began and ended in one place.
What has replaced that is a system you give an objective to rather than a question. Find the three suppliers who meet these conditions, pull their terms into a table, flag anything unusual. It works out what steps that involves and comes back when it has something.
That sounds like a difference of degree.
It is not, and I would push back on that quite firmly. The unit of work has moved from an answer to a task, and that changes what the person on the other end is actually doing.
Reviewing a paragraph is a familiar activity. Everyone knows how to do it. Reviewing a process that something else followed, when you cannot see most of it, is a completely different skill. Very few organisations have deliberately built it, and most have not noticed that they need to.
“The least discussed change of the year is that the tools stopped being events”
What else looks different now compared with the start of the year?
Three things, and none of them was a headline.
The first is that the tools stopped being events. A year ago, using one of these systems was a small decision. You went somewhere specific, you had a purpose, you came back with something. That is how people use a service.
Now the capability is simply present inside the software people already had open. It is in the document, in the inbox, in the customer record. Nobody navigates to it and nobody makes a decision to use it.
Why does that matter more than a capability improvement?
Because it is the point at which a technology stops being a topic and becomes infrastructure. And notice that nothing underlying had to improve for it to happen. The friction simply came down far enough that people stopped noticing they were using anything.
That is the least discussed and most consequential change of the year.
And the second?
The quiet work overtook the impressive work.
Ask most people what AI automation looks like and they describe something large. An entire function handled end to end. A report that writes itself. Ask the people who have actually recovered time what changed, and the answers are almost embarrassingly small. Something that used to need three tabs now needs one. A file that had to be reformatted every Monday arrives formatted.
Those sound trivial.
The arithmetic says otherwise, and it is consistently underrated. A task that takes four hours and happens twice a year is eight hours annually. A task that takes ninety seconds and happens twelve times a day is roughly seventy five hours.
The second is nine times larger and looks a fraction of the size, because we assess importance by how a task feels rather than by how often it happens.
Organisations that went looking for the dramatic transformation have generally been disappointed. Organisations that went looking for the ninety second task have generally not.
And the third change?
Smaller and closer became respectable.
The assumption for most of the past decade was that capability required centralisation. Useful things happened somewhere else and you sent your material to reach them. That assumption has weakened. A growing share of everyday work now runs on smaller models, sometimes on the device in front of the person, and for ordinary tasks the results are frequently indistinguishable from what the largest systems produce.
Sorting, summarising, tidying, extracting, recognising. None of that needs the most capable system available, and using one is slower and more expensive without being better. Matching the tool to the size of the job is not a technical discipline. It is just being willing to ask what a task actually needs.
“Confidence still carries no information”
What has not changed?
Two things that people keep assuming have moved, and a good deal of poor decision making comes from that assumption.
The first is that confidence still carries no information. These systems remain equally fluent when they are right and when they are wrong. There is no tremor in the voice, no hesitation that reliably tracks reliability.
Is that a problem that gets fixed eventually?
I would not plan on it, because it is structural rather than a defect awaiting a patch. What these systems learned to do is produce text that reads well. Accuracy usually comes along with it, because the material they learned from was mostly accurate. But the fluency is the thing produced directly, and when fluency and accuracy come apart, the fluency does not degrade.
Anyone relying on tone to judge whether an answer is sound is using an instrument that was never connected to anything.
And the second thing that has not changed?
Responsibility did not move.
A system that prepares a decision has not made one. Sorting is not deciding. The decision is the act of treating the output as sufficient grounds and doing something in the world, and a person still performs that act and still answers for it.
This matters more as the systems improve, not less. The better the output looks, the more tempting it becomes to accept it without inspection. A tidy table feels more trustworthy than a paragraph, even when it was assembled from exactly the same uncertain material.
“The adoption gap is not a technology gap”
You have said the interesting differences between organisations are not about which tools they bought. What are they about?
Take two companies that adopt the same system on the same terms in the same month. Six months later one has changed how a process runs and the other still has a pilot serving four people. The difference is almost never the technology, and the reasons repeat.
The pilot ran on a curated set of examples rather than a genuine sample, so it measured performance on the part of the work that was never the problem. Nobody owned the unglamorous half: what happens when it is unavailable, who is told when something looks wrong, who checks it is still working in March. And the tool was never connected to the systems where the information actually lives, so using it meant copying material in and out, which means it got used when people remembered and were not busy.
None of that sounds specific to AI.
It is not, and that is the point. Every one of those failures would have applied equally to a software project in 1998.
Organisations that were good at getting things into production before are good at it now. A genuinely capable new category of tool has not repaired an old organisational weakness, and it was never going to.
“Wait a fortnight”
September brings a dense run of announcements. How should people read them?
With a few habits, because the concentration itself produces a misleading feeling. A dozen significant-sounding releases in six weeks makes it seem as though everything is changing very quickly, and some of that is the conference calendar rather than the substance.
Separate the demonstration from the distribution. Announcements frequently describe something that exists, works, and is not available to anyone yet. A remarkable thing arriving next year belongs in a different mental category from an ordinary thing arriving next week.
Ask what the comparison was. Better than what, measured how, on which tasks? A system described as substantially improved may be improved against the same organisation’s previous version, which tells you about their trajectory rather than about the field.
Notice what goes unmentioned. Latency is often omitted when it has worsened. Cost is deferred. Failure behaviour is almost never described, though how something behaves when it is out of its depth is among the more important things about it.
Any single piece of advice?
Wait a fortnight. Two weeks after a significant release, people who are not selling it have used it on real work and written about what happened. That writing is considerably more informative than the launch material, and there is almost never a cost to waiting.
For someone taking stock in the quiet part of August, what should they be asking?
Four questions, and they are worth more than another round of tool evaluation.
Where is the ninety second task? Not the impressive one. The small, frequent, mechanical step nobody has addressed because there is no business case for two minutes.
What would we notice? For every automated process, what is the actual mechanism by which a problem reaches somebody, and does it survive a colleague going on holiday?
How often does the output need real repair before it can be used? That is the single most informative measure of whether a tool is helping or simply relocating work, and almost nobody tracks it.
And what can we do now that we could not do in the spring? Not what has been announced. What has actually changed about how the work gets done.
So where does that leave things?
The technology is further ahead than most organisations are, and the constraint has shifted decisively from capability to adoption.
I do not find that disappointing. It means the difficult part is now a management problem rather than a research problem, and management problems can be worked on deliberately by people who are not building models.
It also means the advantage will not go to whoever adopts the most. It will go to whoever works out, with some precision, which of their own repeated small frictions is worth removing, and then removes it. That work is unglamorous, it does not demonstrate well, and it compounds.
The demonstrations will continue to be impressive. The returns will continue to come from somewhere much quieter.
About Nicole Junkermann
Nicole Junkermann is an entrepreneur and investor. She writes and records the AI Overview, a series of short briefings on artificial intelligence and its effect on everyday work, publishing one question at a time with a full transcript, at nicolejunkermann.ai.
