Workslop is a context problem, not an AI problem

By Kaia Colban

95% of organizations pouring money into AI report no measurable return on it. Not smaller than hoped. None. That's from a 2025 MIT study, as reported in Harvard Business Review, in the same stretch that AI use at work nearly doubled.

So much activity. So little return. Where are the gains going?

They're getting spent on cleanup, at two ends of every AI interaction. Researchers have already named both, and neither is an AI-quality problem. Both are context problems. And context problems don't get solved by a better model. They get solved by a better place to keep context.

Context rot: the tax you pay before the AI helps

Before an AI produces anything useful, you have to hand it everything it doesn't already know. The codebase conventions. The decision from three weeks ago. The reason we don't use that library anymore. The customer who churned over exactly this.

Here's the cruel part. Loading more context doesn't reliably help. Chroma tested 18 leading models, including GPT-4.1, Claude 4, and Gemini 2.5, and found they don't read their context uniformly. Performance gets less reliable as the input grows, even on simple tasks. They call it context rot. The more you stuff in, the more the answer can degrade.

So you do it by hand instead, one careful prompt at a time. Last month I opened a fresh session to change how we handle a retry, and spent the first fifteen minutes re-explaining to the assistant why the retry existed at all. Context I had personally typed into a different session, in a different tool, two weeks earlier. Gone. Re-typed. Didn't I already do this?

Every new tool is a fresh amnesiac. Every session starts from zero. That reconstruction is invisible on any dashboard, but it's where a real chunk of the workweek now goes.

Workslop: the tax your teammates pay after

When feeding context is tedious and nobody measures it, people stop doing it. They ship the polished-looking draft anyway. BetterUp Labs and Stanford Social Media Lab gave that a name too: workslop, AI-generated work that looks finished but lacks the substance to actually move the task forward. It doesn't save effort. It shifts effort downstream, onto whoever opens the doc next.

The numbers are not small:

The workslop tax	Finding
Workers who received workslop in the last month	40%
Time to resolve each instance	1 hr 56 min
Invisible cost per employee	$186/mo
Annual cost for a 10,000-person org	~$9M
Who saw the sender as less trustworthy afterward	42%

Read that trust number twice. Workslop doesn't just cost hours. Roughly half the people who receive it rate the sender as less capable, and 42% as less trustworthy, than they did before. The corner you cut to save ten minutes quietly taxes your reputation with your own team.

Why smart people ship it anyway

This isn't about lazy people. It's about two well-documented human defaults meeting a tool that rewards them. We cognitively offload: hand judgment to the machine and stop checking its work. And we satisfice, a term from Herbert Simon, taking the first good-enough answer instead of the right one. The AI produces a confident, fluent draft in seconds. Of course the good-enough answer wins when the deadline is 4 p.m.

None of that is new. What's new is the scale, and the fact that the cost lands on someone else.

The gains stay personal because the context stays private

Put the two taxes together and you can see the whole mechanism. The productivity is real. One person, in one session, genuinely moves faster. But the context that made it work, and the reasoning that would let a teammate trust the output, lives and dies inside that private session. So the speed stays individual. The cleanup becomes everyone's.

That's the gap Stanford researchers have long called the knowing-doing gap: organizations that know things without being able to act on them. AI didn't create it. AI just poured fuel on it, by generating more individual knowledge than ever while leaving it stranded in ten thousand disconnected chat windows.

The fix isn't a better model. It's a shared substrate.

Here's the part I find genuinely hopeful, and it runs against the usual doom framing. The reasoning isn't disappearing. It's the opposite. For the first time, the thinking is written down as it happens. When you work through a problem with an AI, the whole trail exists: what you tried, what you rejected, what you decided, what's still open. That used to live only in your head.

The problem was never that the context vanished. The problem is that it's trapped in a session your teammates can't see. We finally have more context than ever. We just haven't been using it.

That's the bet behind Lore. Your AI sessions get captured, structured, and shared, so the team runs on shared context instead of everyone privately reconstructing it. The substrate turns raw sessions into the nouns a team actually needs: decisions (what we chose and why), dead ends (what we already tried, so nobody retries it), open threads (what's still unresolved), and handoffs (so the next person starts where the last one stopped, not from zero).

When the retry decision I made two weeks ago is findable, the next engineer doesn't re-explain it to a fresh model. When the reasoning behind a change is attached to the change, the review is fast and the work is defensible. That's the difference between AI speed that stays with one person and AI speed the whole team compounds.

What this means if you lead a team in 2026

The instinct, when AI gains stall at the org line, is to buy a better model or a fifth tool. The data says that's the wrong move. Adoption already doubled and the return still isn't showing up. Another amnesiac in the stack doesn't help.

The move is to make context a shared asset instead of a private tax. Capture the sessions your team already produces. Make the decisions and dead ends inside them searchable. And measure both sides of the ledger, throughput and rework, so the workslop tax stops hiding. The teams pulling ahead aren't the ones using the most AI. They're the ones who built somewhere for the context to live.

So here's mine for you: how many hours did your team spend last week re-explaining something one of you already told an AI?