Is My Slop Slowing Us Down?

What motivates this essay is personal experience, but it is a well-documented, more universal problem and opportunity. Over the past three years or so, in my own work and in the organizations I work with, I have watched the same thing happen again and again. Nearly everyone is personally more productive with AI. They draft faster, research faster, prototype faster, get faster results. Ask them individually and they will tell you, with real conviction, that these tools have changed how they work. Then ask their colleagues down the hall — the ones who receive that work — and there can be a different story. The organization as a whole slows down for a while. Review cycles are longer. More meetings are spent untangling what a document or workflow really means for the organization. Somewhere between the individual and the institution, the productivity can stall out.

There is documentation. Researchers at Stanford and BetterUp Labs have given the phenomenon a name — “workslop” — AI-generated content that looks like good work but lacks the substance to move a task forward. Their study found that four in ten workers had received something in the past month that cost its recipient nearly two hours of cleanup. Ethan Mollick has made the related observation that AI use which boosts individual performance does not naturally translate into organizational performance — and I would add that it does not naturally uphold an organization’s quality standards either.

All of that matches what I have seen up close. When I use AI for a team project and am blown away with what is often called the “AI augmentation” of my effort, I also immediately worry: What does this mean for my organization? Who does it affect? How? Have I checked it all? Does it scale? Is this just MY too-quick slop? This essay is my attempt to describe the mechanics: where the individual gains go, and what the path through can look like.

The Old Constraint Was Doing Us a Favor

For most of my career, the effort required to produce work was a natural filter. Writing a strategy memo took a while, so you thought hard before starting one. Building a financial model took a week, so you asked colleagues whether the assumptions made sense before you built it. Producing a campaign took a team, so the team argued about the audience and the message first. The cost of production forced a certain amount of forethought, collaboration, scoping, and process — because nobody wanted to waste that much effort on the wrong thing.

AI removed the cost of production without replacing the filter. A plausible-looking sixty-page report now takes an afternoon. A batch of two hundred prospect emails takes an hour. The thinking that used to happen before the work — the conversations about context, scope, purpose, and goals — can now be skipped entirely, and often is. Not out of laziness, but out of ambition. The tool simply makes it possible to start producing before anyone has decided what should be produced, and starting feels like progress.

Everyone knows the phrase “garbage in, garbage out.” That problem is real and everywhere in AI work: vague prompts, missing context, unexamined assumptions fed into a system that will confidently build on all of them. But GIGO alone does not explain what I am seeing. The second problem is volume. Poor inputs degrade the quality of each artifact, and the ease of generation multiplies the number of artifacts. What lands on the rest of the organization is a large quantity of work that is some blend of magic, genius, and bullshit — or is simply, though perhaps just slightly, unmoored from what the business needs.

And here is the part that shows up in no individual’s productivity metrics: someone else has to deal with it. Every unvetted work product becomes an unfunded mandate for a colleague. It has to be read, edited, fact-checked, tested, and frequently rebuilt. The effort that the author saved did not disappear. It moved downstream — usually to the most experienced people in the organization, the ones whose judgment is needed to separate the sound from the merely fluent. We have created a machine that converts the time of junior enthusiasm into the time of senior review, and the exchange rate is terrible.

Where It Shows Up

Marketing

Content generation was the first and most visible case. A marketer can now produce a month of blog posts, social copy, and landing pages in a day, and many do. The result reads fine sentence by sentence. But who sat down on the ninth blog post this week to decide what this company believes, who exactly it is talking to, or what a reader should do differently after reading? The content team’s output charts go up and to the right while the brand voice dissolves into randomness. Then a senior editor — or worse the CMO — ends up rewriting the pieces that matter, which takes longer than writing them properly would have, because now there are more off-brand versions to argue with.

Sales

Automated SDR tools are hypothetically a marvel when set up with care, which is rare. Set up in an afternoon by one motivated rep, they send thousands of messages that open with a personalization token and proceed to say nothing specific to the prospect, the industry, or the moment. The messages do not convert. Worse than not converting, they burn the territory: prospects who receive obviously automated outreach remember — and ignore — the company that sent it. The pipeline dashboard shows enormous activity. The revenue line shows nothing. And sales leadership then spends a quarter repairing sender reputation, rewriting sequences, and rebuilding the targeting logic that should have been the first conversation, not an afterthought.

Finance

This is the one that worries me most, because the failures are quieter. Financial AI agents — whether reconciling accounts, building forecasts, or drafting board materials — will produce confident numbers with or without proper vetting. An agent configured quickly, without thorough testing against known-good data, without clear boundaries on what it should and should not touch, produces outputs that look exactly like correct outputs. The errors surface later: in an audit, in a board meeting, in a decision made on a valuation that was subtly wrong. The controller who must now verify every agent-produced number by hand is not being paranoid. She is doing the setup work that was skipped, at the most expensive possible point in the process.

In all three cases the pattern is the same. The individual using the tool experienced a productivity gain. The organization absorbed a productivity loss. The gap between those two is the editing, verifying, testing, and repairing done by others.

What Organizations Move Through

I want to be careful here, because none of this is an argument against using AI. I use these tools daily and I would not go back. I have seen and been part of massive efficiencies and new value creation. What I am describing is a stage that organizations pass through — and my conviction is that the stage is avoidable, or at least can be shortened, if you can anticipate it and recognize it while you are in it.

The phases are not inevitable in their full painful length. But skipping ahead requires leadership to slow down the most visibly productive people in the building and ask them to spend more time on inputs, collaboration, and verification — at exactly the moment the tools are indicating that none of that is necessary anymore.

Relocate the Effort, Don’t Reduce the Use

The core move is to relocate the effort, not reduce the use. Put the hours back where the old constraint used to put them: before generation, in the shared work of deciding what is worth making and what good looks like; and after generation, in testing and review that is planned and budgeted rather than absorbed invisibly by whoever happens to be downstream.

Measure the whole cycle — from first prompt to accepted work product — rather than the raw output of individuals, because raw output is now nearly frictionless and therefore nearly meaningless as a metric. And treat every AI agent, whether it writes emails or reconciles ledgers, as a new hire: it gets onboarding, it gets supervision, and it earns autonomy through demonstrated reliability, not through the enthusiasm of the person who installed it.

I will close with an obvious irony that makes me smile. AI can help with the organizing too. It is a fine assistant for drafting a project charter, keeping a decision log honest, surfacing the questions a plan has not answered, or naming the steps for an AI project. Used that way, the tools become part of the cure for the very mess they helped create.

But the parts that matter most remain stubbornly, wonderfully human: the planning, the collaboration, the arguments that end in actual agreement, the allocation of scarce people and money, the setting of goals worth pursuing. No model can execute those for us. And when we do them well, everything the machines produce gets better. That work has always paid off. It pays off more now than it ever has.

The tools are ready, yet continuously new. The question is whether we are organized enough to adapt and organize ourselves anew.