What is work slop?
Work slop is a recent coinage, generally defined as low-effort, AI-generated content that looks polished but lacks the substance to meaningfully advance a task. But we can push the definition further. Work slop is output that looks like work while transferring the real cognitive effort from the person who produced it to the person who receives it. It shifts the burden from producer to recipient.
Work slop is output without presence. The sender gets the appearance of diligence; the receiver gets the actual cognitive labor.
- An email drafted by AI that the sender never read but the recipient must parse.
- A report generated in seconds that takes an hour to verify.
- A Slack message that covers the sender's accountability while creating a sorting problem for everyone else.
- A training session full of buzzwords with no connecting logic or principles from which to form real behavior patterns.
The “AI-generated meeting summary” is a useful case study. Imagine five people working together on an initiative. In one particular working session, four of them meet, work through a hard problem, and reach a decision. The fifth person, who missed the meeting, receives an AI-generated meeting summary afterward. This high-level document notes that a decision was made. But It does not say why, or what exactly was decided in enough detail to act on.
This absent colleague is left with two bad options:
- Act on the summary without the rationale and risk getting it wrong, or
- Go back and consume the full recording, which defeats the purpose of the summary.
The artifact meant to save effort has added effort for the one person it was supposed to help!
Why does it matter?
Work slop works against real work
We tend to treat the meeting as the work and the follow-up communication as logistics. But rearticulating what was discussed, in a new form for a new audience, is refinement work. It filters the conversation through a second layer of judgment that operates differently from our free-talking, free-ideating mode. This is why a meeting full of argument can produce a write-up that contains none of the arguing: in its ideal form, the write-up presents a single synthesized way forward that resolves the tensions felt in the room. Much of the time, that restatement is the cognitive effort.
When we hand that restatement to an unsupervised AI, we do not just save time. We skip the judgment layer where positions get reconciled and decisions get sharpened. The lesson is not to simply avoid AI or return to pen-and-paper processing. Rather, it is to recognize where the work actually happens and make sure it still happens, even as we automate away the drudgery surrounding it.
Work slop dilutes information quality
In my current consulting engagements, I maintain GitHub repositories of meeting transcripts that Claude Code analyzes on demand. I might prompt Claude Code like this: “Pull everything Stakeholder A said about DevOps transformation this month, compare it against our technical discovery findings, and return a maturity gap analysis.” However, Google Meet delivers each transcript with a Gemini-generated summary attached on top, and I deliberately strip that summary before filing anything so that only raw transcripts go into these repos.
But let’s say I upload the entire Google Meet package into GitHub without stripping out the AI-generated summary. Now each transcript contains not only what people actually said, but also several paragraphs of what Gemini thinks they said. Here’s the problem: Gemini only summarizes the content of each individual call. Stakeholder A mentions “the document we sent you yesterday,” and the AI-generated summary interprets this into a bullet point: “Stakeholder A sent you the document yesterday.” However, in yesterday’s working session, right after Stakeholder A dropped the call early, Stakeholder B noted that a permissions issue was preventing them from sending the document. Gemini’s summary couldn’t possibly know this, so it slightly mis-documented the truth. The document, of course, has fallen through the cracks because we are relying on Gemini summaries to capture complex action items. But even though we have a complex Claude-based knowledge management approach in GitHub, that assumption has been smuggled into our system via pre-pended Gemini summaries and is being tracked as if it was true. And two weeks later, when Stakeholder A insists that we should have known about an upcoming deadline, we face hours of decision archaeology to address the escalation and understand the gap.
The appropriate solution, of course, is a tried and true one: enforce pull requests and implement automated code reviews, instructing a smart frontier model to strip out any Gemini artifacts before they enter the repo. Curate your sources of truth to support interpretability and provenance, and generate summaries that are fit for purpose each time one is needed.
Work slop promotes "productivity" over alignment
Real alignment between people is more than sharing notes. It is individuals engaging their own minds with concepts, with other positions, and with the perspectives of other people to arrive at a way forward that draws on everyone's strengths. When someone produces work slop, they and their actual opinion are absent from that process. They have substituted their presence and instincts with a probability-based approximation of text drafted somewhere else.
In communication-theory terms, work slop is without intention. It carries meaning only because words themselves carry meaning, but the point of communication is to externalize something that lies within, an opinion or instinct, in a form another person can interrogate from their own experience.
This does two kinds of damage.
- First, the producer waives themselves from their own work. While some of us may legitimately desire to disentangle ourselves from our professional responsibilities, it’s worth asking whether it serves the people we work with or disrespects them. It’s almost always the latter.
- Second, it degrades everyone else's contribution. Colleagues bringing their best selves are now forced to engage with a selfless blob of text. The entire reason we work together rather than alone is to accomplish more than any one of us could individually. When one of us checks ourselves at the door, everyone else is reduced in their capacity to create and realize value.
Where does it come from?
Work slop comes from bad context engineering, which Anthropic defines as the process of curating and maintaining the optimal set of tokens during LLM inference, including everything that lands in the context window beyond the prompt itself. Along this trajectory, we can trace the sources of work slop from the AI model outward:
Improperly scoped and parameterized models.
The closest source is the model itself: temperature, max token limits, and model selection that are not fit to task. Run a complex transcript through a cheap, low-reasoning model with the temperature too high and the output capped too small, and the summary cannot match the information variety of the underlying conversation. This is a question of requisite variety in the AI configuration itself.
Bad architecture and unoptimized context.
The next layer out is everything surrounding the model. The AI may have no ability to reach systems of record beyond itself, or those connections may be poorly built: weak integrations, non-descriptive error messages that fail to help the model navigate its information space. And even where the architecture is sound, the context itself may be uncurated, so the information and tools provided at inference time do not fit the task at hand.
Uninformed prompt engineering.
People often prompt in suboptimal ways: asking for a vague outcome rather than walking the model through a process, or failing to define good versus bad results. Within this category, Gemini's meeting summaries are a prime example. The model selection, routing, and prompting are automated by the platform, built for broad usage rather than custom fit to your task, and often invisible to you entirely.
Bad processes.
This outermost layer has nothing to do with AI necessarily. Organizations inject AI into broken processes without realizing how much humans with intuition, taste, and smarts have been quietly compensating in between receiving tasks and delivering on them. In those cases, if you strip out the humans, the cognitive compensation goes with them.
The common thread is a failure of externalized cognition: identifying the thinking that goes into a high-quality artifact, then writing it down or architecting it into a system that performs it reliably. The less of that cognition is externalized, the more the output trends toward slop. I call this capability enterprise metacognition: not just an individual's ability to think about their thinking, but the business's ability to think about its thinking.
Work slop, distilled to its essence, is a lack of metacognition.
How do we avoid it?
The layers above are the roadmap:
(1) select and parameterize AI models correctly, (2) engineer your context through careful architecture, (3) improve your prompting, and (4) fix your processes. Treated as a deployment checklist, this is how you design AI implementations that avoid work slop from the start.
But notice what the checklist implies. Metacognition across IT systems, individual enablement, and company-wide processes is transformation: operating-model-level change in governance. That is far bigger than what most companies mean when they ask (as several have asked me directly), "How do I ‘switch on’ AI for my organization?" Simply ‘switching on’ AI - usually through a platform ecosystem selling vendor-locked-in agents - often means switching on work slop across the business.
The alternative is an agentic operating model: designing AI around your business, and your business around AI, so that it amplifies the people and opinions contributing to the company's success rather than diluting them. I believe this is a new domain of consulting and advisory services, and one we are well positioned to meet if we can pull it off ourselves. For more, see Harnessing the Chain Reaction, which I co-wrote with Marcus Corpening at Foray Consulting.
How do we deal with it?
Suppose we’re already downstream of work slop. It is circulating in the environment and getting in the way of current work, and re-architecting toward an agentic operating model takes time. Just as social feeds are overflowing with AI-generated videos and articles, many work environments are becoming overgrown with AI summaries, status updates, and slide decks that carry no substance, no opinion, and no human perspective, just an amplification of vague intent. Four practices help right now:
1. Choose quality over speed. It is better to forgo AI on a task and produce high-quality output than to use AI to speed up the task and reduce quality. It’s worth noting that most pushback on this principle is really just a disagreement about the definition of quality. If my manager wants a task done in five minutes, I tell him doing it “right” takes twenty, and he still wants it in five, he is not ultimately challenging the timeline. He is challenging the definition of right.
2. Get focused on requirements. Understand what is actually being asked of you. In knowledge work, the negotiation of trade-offs usually happens beneath the level of organizational documentation. It occurs in the minds of individual workers and is audited informally (often instinctually) by whoever consumes the output. But AI requires you to externalize cognition, and therefore to decide explicitly what the quality requirement is for any given task.
3. Track and document the downstream impact. When an AI-generated meeting summary sends a team down two weeks of work on a course of action that was never actually agreed upon, do not cover it up. Surface it and document it. This builds the evidence base for pushing back, and it shows precisely where the complexity of your requirements exceeds the capacity of the AI systems participating in your management systems.
4. Redo work with better systems. Return to the Gemini summaries example. Gemini's entire world, at least as deployed, is the space of a single meeting; it cannot contextualize its summary against other decision records. So I ignore it, filter it out of the signal, and replace that layer with your own: raw transcripts in your own repository, processed by Claude Code with specific, contextual instructions. Read all of the raw transcripts; create a list of commitments, decisions, and action items organized in an Eisenhower matrix; cite the person, date, and time where each was stated; flag any conflicts; store the result in a decision-record subdirectory.
This is not replacing AI with manual effort. It is more labor intensive than trusting a headless AI, but still less than working without AI at all. Implementing with intentionality co-opts the entire chain that creates work slop. It starts by curating the information correctly and works back up the stack: a more capable model, better prompts, better-architected context. Within a business, this approach showcases the value of real AI-augmented knowledge work over and against work slop, which is simply AI-augmented bad work.
If your organization is trying to figure out where real AI-augmented work ends and work slop begins, we'd love to talk. Praecipio helps teams design AI implementations that amplify the people behind the work, not replace them.
