Slop Grenades: What Shopify's CEO Gets Right About AI Workslop, and Why CX Teams Should Listen

A slop grenade is a piece of AI-generated work that someone sends to a colleague without reviewing it first. A pull request nobody read. A three-page email inflated from two bullet points. A "summary" that sounds right and can't be checked. Shopify CEO Tobi Lütke gave it that name in September 2026, and it's the sharpest description yet of AI workslop: output that looks finished but pushes the real work onto whoever receives it. For CX teams, the risk is bigger than for most. Customer feedback analysis is the perfect slop factory: lots of text, nobody reads it all, and the summary lands directly in front of leadership.

What is a slop grenade?

Lütke used the term on The Knowledge Project podcast: "So we call those slop grenades that people toss at each other." He credited Harry Brundage with coining it.

His two examples are painfully familiar:

An engineer asks an AI agent to change some code, approves the pull request without really reading it, and leaves teammates to find out what it actually does.

Someone expands a short point into a long email with one model. The recipient compresses it back down with another model. Two people, two AIs, zero information added.

Then the line that matters: "The failure case now of lazy work is not lack of output."

There's an irony here. This is the same CEO who wrote in April 2025 that reflexive AI use was a baseline expectation at Shopify, and that teams asking for more headcount first had to show why AI couldn't do the job. Lütke still wants AI everywhere. What he's naming is what happens when output becomes free and judgment doesn't scale with it.

Slop grenades are workslop with a pin pulled

The broader phenomenon already had a name. In September 2025, researchers from Stanford's Social Media Lab and BetterUp Labs published a study in Harvard Business Review defining workslop as AI-generated work content that looks polished but lacks the substance to move a task forward.

Their numbers, from a survey of 1,150 US desk workers:

40% of them had received workslop in the previous month.

Each instance took an average of 1 hour 56 minutes to deal with.

Estimated cost: $186 per employee per month.

For a large organization, that adds up to more than $9 million a year.

The grenade metaphor adds something the research only implied: blast radius. Workslop doesn't cost the sender anything. It costs the receiver. The person who generated it saved ten minutes. The three people downstream lose two hours each figuring out what's true.

Why CX teams are the most exposed to AI generated insights

Look at how a lot of teams handle customer feedback right now. Thousands of verbatims from surveys, reviews, tickets and calls get pasted into a general purpose LLM. Out comes something like: "Customers are frustrated with delivery times and want better communication." It reads well. It's probably not wrong. It goes into the monthly CX review deck.

That's a slop grenade, and it lands on three people:

1. The product or ops lead who has to act on it. Which deliveries? Which regions? Since when? They go back to the raw data and redo the analysis.

2. The executive who reads it. They can't tell a real pattern from a plausible paragraph, so they either ignore it or overreact to it.

3. The CX team itself, whose credibility takes the hit the first time someone checks and finds the insight doesn't hold.

This is why AI generated insights about customers are a special case. In most workslop, someone eventually notices the email is empty. With customer feedback, nobody downstream has read the 10,000 comments, so nobody can tell the summary is hollow. The grenade goes off quietly, inside decisions.

We covered the technical side of this in our post on the traps of AI text analysis in CX. The short version: a model that summarizes fluently can still measure badly.

How to tell real AI customer feedback analysis from slop

Across 100+ CX projects, our customer success team keeps coming back to the same five questions. Run any AI output on customer feedback through them before it leaves your desk.

1. Can every insight be traced back to the verbatims behind it? If you can't click from "delivery frustration" to the actual comments, it's an opinion, not an insight.

2. Is accuracy measured, theme by theme? "The AI is pretty good" is not a metric. You should know how reliably each topic is detected.

3. Is sentiment read at the aspect level? "Great staff, but the app crashed twice" is not a neutral comment. It's one positive and one negative signal. Averaging them hides both.

4. Is it prioritized by business impact or by volume? The most frequent complaint is rarely the one that costs the most. Volume ranking is how you end up fixing the loudest problem instead of the expensive one.

5. Did a human who knows the business validate the categories? The taxonomy is where your expertise lives. If nobody on your side signed off on it, the AI is guessing what matters to you.

Five yes answers, you have intelligence. Two or three, you have a draft. Zero, you're holding a grenade.

What anti-slop looks like in a customer intelligence platform

This is the problem a customer intelligence platform exists to solve, and it's how we built Feedier. Every theme runs as its own agent with its own rules, outputs are grounded in your ingested feedback rather than the model's general knowledge, and accuracy is scored per theme: 95% out of the box, up to 99% once your team validates the taxonomy.

Sentiment is read per aspect, a Business Impact Score ranks issues by what actually moves your numbers, and weak signal detection separates a one-off complaint from an emerging pattern.

The human stays in the loop, just in a better spot: validating what the categories mean, not rereading 10,000 comments to check a summary. As we argued in AI doesn't think for you, domain expertise is still what drives performance.

The real rule: whoever pulls the pin owns the blast

The fix here is ownership, not another tool. If you send it, you own it, whether you wrote it or a model did. For CX leaders, that translates into one simple standard: no customer insight goes to leadership unless you could defend it with the verbatims in front of you.

Slop grenades are a symptom of AI making output cheap. In customer experience, the fix is AI whose work you can check. If your voice of customer program currently runs on pasted prompts and plausible summaries, that's the gap to close first.

Want to see what traceable, measured feedback analysis looks like on your own data? See how the pplatform works.