I am about as pro-AI as a person can be. I build the systems behind Atlas, I use models all day, and I am genuinely excited about where this goes. So when I say that a lot of the AI productivity claims do not hold up, understand that I am not a skeptic looking for reasons to doubt. I am a practitioner who got tired of demos and started watching what actually happens when these tools meet real work. The pattern, once you look past the marketing, is more interesting and more honest than anything I read in a press release.
The short version: AI made some work dramatically faster, made some work feel faster while actually being slower, and made a third category of work worse in ways that only showed up two weeks later. The trick to getting value from AI at work is knowing which bucket a given task falls into, and almost nobody selling you AI tools wants to have that conversation. Here is what we found.
The myth that AI productivity is uniform
The headline claim is always a single number. AI makes developers forty percent faster, support agents twice as productive, writers three times as fast. Whenever I see one figure stretched across an entire job, I know the measurement was done on the easiest slice of the work. Productivity from AI is wildly uneven, and the average hides the only thing that matters, which is the variance.
In practice, the gains cluster hard around tasks with three properties: the output is easy to verify, the cost of being wrong is low, and the work is something a competent person could do but finds tedious. Writing a first draft of boilerplate code, generating test fixtures, summarizing a long thread to catch up, drafting a routine email, translating a rough note into a clean ticket. On that kind of work the speedups are real and large, often two to four times faster, and they hold up when you check the output.
The moment any of those three properties breaks, the gains evaporate or reverse. Hard to verify, expensive to get wrong, or genuinely requiring judgment, and AI goes from accelerator to liability. The aggregate "forty percent" is just these two populations averaged together, and the average tells you nothing about your specific task.
The myth that faster always means more productive
This is the one that fooled me at first, and it is the most expensive myth in the whole category. Speed and productivity are not the same thing, and AI is exceptionally good at making slow work feel fast. A developer using an AI assistant generates code faster. That feels like productivity. But if that code takes longer to review, hides a subtle bug that surfaces in production, or quietly accumulates patterns nobody on the team chose, the time you saved writing it gets paid back with interest somewhere downstream where it is harder to see.
This shows up whenever someone measures it carefully. A team using AI on a set of medium-complexity tasks finishes writing meaningfully faster, often around a third faster. But track the same work through code review, rework, and the first two weeks of production, and the net time advantage shrinks toward nothing, and on the trickiest tasks it goes negative. The work was not more productive. It was front-loaded. The fast part happens first and feels great, and the slow part happens later and gets blamed on something else.
Example: Picture an engineer who ships a data migration with AI assistance in an afternoon that would normally have taken two days. Everyone is impressed. Three weeks later the team spends four days untangling an edge case the generated code handled wrong in a way plausible enough to pass review. Net, time lost. The lesson is not "do not use AI for migrations." It is "the verification cost is the real cost, and AI does not reduce it, it often raises it."
The deeper issue is that a verification bottleneck is harder to manage than a production bottleneck, because it is invisible until it fails. When a person is slow at writing code, you can see the queue backing up and you staff for it. When a person is fast at producing code they cannot fully check, nothing looks wrong, the dashboards stay green, and the cost lands later as defects, rework, and the slow erosion of confidence in the codebase. The teams that handle this well consciously rebuild their review practices around the assumption that more code will arrive, written faster, by people who understand it less deeply than if they had typed every line. That is a real organizational change, and treating AI as a simple speed dial means you never make it.
The myth that AI replaces the thinking
The most seductive claim is that AI handles the hard cognitive work so you can focus on higher things. In practice it is closer to the opposite. AI is brilliant at the mechanical middle of a task and weak at both ends, the part where you decide what is worth doing and the part where you judge whether the result is actually good. Those two ends are the thinking. They are also the parts that are hardest to measure, so they are the parts the productivity claims quietly ignore.
What we found is that AI shifts where your effort goes rather than removing it. Less time typing, more time specifying and reviewing. For people who are strong at specifying clearly and reviewing critically, this is a genuine win, because it moves their hours toward their strengths. For people who are weak there, AI is dangerous, because it lets them produce a large volume of plausible work they cannot actually evaluate. The tool amplifies whatever judgment you already have, including the absence of it.
- If you cannot tell good output from bad in a domain, AI will make you faster at being wrong, not faster at being right.
- The bottleneck moves from production to verification, and most teams have not staffed or sized for that shift.
- The biggest gains go to experienced people on tedious work, not to junior people on hard work, which is the opposite of how it is usually sold.
What actually held up
I do not want this to read as a takedown, because the real gains are real and we have built our company around them. When we point an AI assistant at the right kind of task, the results are not marginal. Inside Atlas, the assistant we call Ask Atlas earns its place precisely because it lives where the work and the context already are, so it is drafting against real data, summarizing a real inbox, or pulling the right record into view, rather than guessing in a blank box. The wins are largest when the AI is grounded in your actual material and the human stays firmly in the loop on judgment. You can see how we think about that in Atlas, where the assistant is deliberately built to assist rather than to pretend it can decide.
The honest summary is this. AI is a real productivity gain on verifiable, low-stakes, tedious work, applied by people who can judge the output, and a net loss almost everywhere else. The companies getting outsized value are not the ones using AI for more things. They are the ones who figured out which tasks belong in the first bucket and ruthlessly stopped using it for the rest. Future-of-work conversations rarely get this specific, because specificity is less exciting than a big round number. But the specificity is the entire game. Measure your own work, sort it into the buckets, and ignore anyone who hands you a single percentage for everything.