Meta CTO Andrew Bosworth told staff to convert AI productivity gains into more work rather than shorter days, and the reaction was faster and angrier than I expected. By the next morning the HN thread had converged on a phrase that’s now everywhere: productivity pollution. One person generates ten times the volume, and the cost of that volume gets paid by whoever sits downstream reading it, checking it, and putting it where it actually belongs.

Meanwhile the GDP figures keep coming in flat. Trillions of dollars of capex, no measurable productivity boom, and a growing pile of documents nobody asked for.

I think the phrase is diagnosing a symptom and blaming the wrong organ.

The cleanup is the tell

Go read the actual complaints instead of the summaries of the complaints. They all share a shape. Someone’s AI produced a thing, and then a human had to move that thing by hand into the place where work is really tracked. A meeting summary that gets re-typed into Jira. A draft reply that gets pasted into Zendesk and then edited because the tone is off. A competitor analysis that a PM has to chop into six fields in a Notion database because the AI wrote paragraphs and the database wants values.

The pollution isn’t the text. It’s the manual transfer step nobody budgeted for.

The model never saw the system of record

Here’s the part that gets skipped. When you ask ChatGPT to draft a customer response, it is writing into a blank chat window with no knowledge of the ticket ID, the account’s plan tier, the two previous escalations, the refund policy your team quietly changed in April, or the six required fields the form will reject on submit. So it produces something plausible-shaped, and plausible-shaped is exactly the thing that costs a colleague twenty minutes to fix, because verifying a confident wrong answer takes longer than writing a hesitant right one from scratch.

The agent produced output for a context it was never shown. That’s not a model quality problem and better models will not fix it, which is why the last two years of benchmark jumps have not moved the numbers economists care about. GPT-5.2 writes beautifully into the void. So does Opus. So does Gemini.

We wrote about this before as the context tax, and the framing still holds: the expensive part of using AI at work is assembling the situation the AI needs, then disassembling its answer back into the systems that hold the truth. Both ends of that sandwich are human labor. Both are invisible on any productivity dashboard. And both scale linearly with how much the AI produces, which is precisely why “just generate more” makes the aggregate worse rather than better.

That’s the pollution. Not slop. Displacement.

The tab is already open and already authenticated

You are logged into Salesforce right now. And Gmail, and Linear, and the internal admin panel with the ugly table. Your browser holds session state that no API integration will hand you without a two-week procurement conversation.

What changes when the agent works in the real form

An agent running in your browser side panel reads the ticket that’s on screen, with the account history and the fields as they exist, then fills the actual form. Not a doc. The form. Dassi lives in that side panel and works against whatever tab you’re in, using the logins you already have.

Nobody re-types anything. Nobody receives a summary and wonders which of the nine numbers in it were hallucinated, because the numbers came off the page they’re looking at.

There’s a second effect that I like more than the first one. Because the work happens where you can watch it, the review step collapses into the doing step. You see the field get filled. You correct it in place. The whole “AI produces artifact, human audits artifact later” loop, which is where basically all the resentment lives, never forms.

Bosworth is not wrong, exactly

If AI genuinely made a task cheaper, converting that into more output is a rational thing for a company to want. The problem with the memo is that it assumes the gain is real and captured, and mostly it’s neither. It’s been shifted onto someone in another department who now spends their Thursday reconciling AI-generated crap against the CRM.

We’ve argued that browser agents are the productivity stack everyone forgets, and this is the sharpest version of the argument. An agent with no view of the system of record can only produce candidates. A candidate is a promise of future work.

I don’t know whether any of this shows up in the productivity statistics. Honestly I doubt one Chrome extension moves a national accounts figure, and I’d distrust anyone who claimed otherwise. But the local version is measurable in about a week: count how many times you copy something out of an AI and paste it into a form. That number is the tax. Watch it go to zero and decide for yourself whether the economists were measuring the right thing.