Four separate pieces trended this week with essentially the same thesis. “AI Productivity Paradox: Multiplier Under 2x.” “The AI Productivity Trap.” “The AI Productivity Lie.” And my personal favorite, “More Done and Still Zero Free Time.” The commentary split predictably between skeptics dunking and enthusiasts defending, everyone arguing about whether the gains are real.

They’re real. And the critics’ numbers are probably right. But the diagnosis is completely wrong.

What forty percent of your AI time actually looks like

I timed myself last Tuesday. Straightforward task: take a messy email thread from a client, extract three action items, draft responses. I opened ChatGPT and spent the next several minutes doing what I always do — copying the email chain, pasting it in, explaining which parts mattered, describing the client relationship, noting the tone I wanted. By the time I had finished feeding context into the chat window, I could feel the productivity gain evaporating in real time. Not once during those minutes did the AI do anything wrong. It was waiting. I was the bottleneck.

So I started tracking it. Over three days I measured how I was actually spending time during AI-assisted work, and the split was ugly. Somewhere around 40% of every interaction was context delivery. Selecting text, copying, switching tabs, pasting, typing descriptions of what was on my screen or what I was trying to accomplish. The model nailed the task once it understood the situation. But getting it to understand the situation was its own damn job.

And this is not unique to me. A friend who manages a sales team described the same pattern, where his reps use AI to draft emails and spend half their time explaining deal context that sits right there in Salesforce, two tabs over, while they type summaries into a chat window by hand.

Nobody times the copy-paste

Because it doesn’t feel like overhead. It just feels like using the tool.

The multiplier math stops working at the keyboard

Those trending articles cite studies showing AI gives knowledge workers roughly a 1.5-1.8x productivity multiplier. And honestly that sounds about right once you factor in the context tax. Take a task that should be 10 minutes of AI-assisted work. If you spend 4 of those minutes just getting the AI up to speed on what you’re looking at, your effective multiplier on a 10-minute manual task drops from a potential 3-4x down to about 1.7x. The context tax is not some minor friction. It is the gap between what AI should deliver and what people actually experience.

But nobody writing these “AI productivity is disappointing” pieces seems to notice. The bottleneck is not intelligence. GPT-5.2 can reason through just about anything you throw at it. The bottleneck is the copy-paste-explain ritual that precedes every useful interaction, a ritual so ingrained that most users do not even register it as overhead anymore.

Your screen is already the context

A browser agent that can see your screen doesn’t need you to describe what you’re looking at. Because it is already looking at it.

Dassi runs in your browser’s side panel and reads the page you have open. So when I need to draft a reply to that messy email thread, the agent already has the thread. When the sales reps want to write a follow-up, the deal context in Salesforce is visible in the tab. Zero copying, zero pasting, zero explaining. And suddenly the productivity multiplier looks a lot closer to what the models are actually capable of delivering.

I wrote about how setup friction eats AI gains a few weeks ago. The context tax is the same species of invisible overhead that accumulates silently until the tool barely outperforms doing things manually, except this one recurs on every single interaction rather than hitting you once during onboarding and then going away. Setup friction you pay once, and Dassi solved that by inheriting your browser sessions. But context tax you pay forever, and the only fix is an agent that can actually see what you see.

The critics measured a workflow, not a technology

The AI productivity critics are looking at real data and drawing the wrong conclusion. They see a mediocre multiplier and blame the models or blame the hype. But the data is measuring a workflow that forces humans to be the bridge between their screen and the model, and that bridge crossing eats half the time on every single task. Or remove that bottleneck entirely and you are measuring something completely different.

A 1.7x multiplier is exactly what you’d predict for incredible technology trapped behind a copy-paste workflow designed in 2023. The models got better. The plumbing did not.