Stop foisting AI-generated garbage on your colleagues, they could have been far less busy.
AI can write an email in seconds, but the recipient may spend far longer figuring out what it is actually trying to say. Shopify CEO Tobi Lütke recently coined a term for this kind of work output: “slop grenades”.
In April 2025, Tobi Lütke sent a memo to employees titled Using AI Is Now a Basic Requirement. He believed that AI is transforming the way merchants operate: it can not only provide suggestions, but also directly complete part of the work for merchants.
1 The CEO Who Advocates AI the Most Is Tormented by AI
To keep Shopify’s products up with this change, employees must also stay ahead of the curve. Lütke wrote that “using AI instinctively” has become a basic requirement for Shopify for its employees. He said he often uses AI as a thinking partner, research assistant, critic, mentor and pair programmer, and has also used AI agents to prepare speeches, hoping to encourage everyone in the company to try it out.
Lütke explained in the memo that proficient use of AI is a skill that can only be acquired through repeated practice. Previously, he only invited employees to explore AI, but now he believes that staying at the “suggested to try” level is no longer enough. He calls AI a “multiplier” of capabilities, saying that some employees use it to start handling tasks that they would not have chosen to take on in the past, and even complete workloads that were unimaginable in the past.
In his view, continuous learning and adapting to changes are not empty slogans, but basic requirements for joining the Shopify team.
To what extent has Lütke elevated the importance of AI?
Even before the team applies for additional manpower or resources, they need to first explain why the existing work cannot be completed with the help of AI. This memo pushes AI from an optional tool to the forefront of daily work and resource decision-making. It also inspires employees to devote themselves to AI “crazily”, leading to a flood of a large number of low-quality outputs generated by AI.
After this initiative was implemented for a year, Lütke himself said he was already overwhelmed by the negative consequences.
More than a year later, a few days ago Lütke talked about the other side of the story on the podcast The Knowledge Project: some people let AI generate emails or codes, and hand over the finished products to colleagues without careful inspection. The recipients have to re-read, check and modify them before they can continue working. Shopify internally calls such deliveries “slop grenades”.
During the podcast, the host asked Lütke: Has AI made some things inside the company worse? Lütke’s answer was “yes”. He then talked about a new type of work error: in the past, laziness might be manifested as no output; now, with the help of AI, laziness may instead be manifested as too much output. Shopify internally named this kind of unchecked AI content directly handed over to colleagues for processing as “slop grenades”.
This term comes from a company that has been deeply using AI. Lütke said that in Shopify, few engineers write code completely manually, and many people use multiple agents to assist development at the same time. The company’s internal AI assistant River can participate in discussions in public work groups, sort out requirements, research materials, make prototypes, and propose code modifications based on conversations. According to Lütke’s estimate in the interview, about half of the code change requests in Shopify now come from conversations in the company’s shared chat. But this proportion is only his oral estimate, not the statistical data disclosed by the company.
The design of River also undertakes an organizational task: to let employees see how colleagues use AI in public channels. Lütke believes that proficient use of AI is a capability that employees need to cultivate, and the public collaboration process helps everyone learn from each other. When employees discuss a feature, they can ask River to summarize opinions, create task lists, or make a prototype at any time. Some work that used to require dedicated personnel to take over can now be promoted in the same discussion.
But problems also followed. Lütke gave an example: someone wanted to modify the system, so he asked River to generate a code change request, and submitted it without checking it carefully. When his colleague reviewed the code, he found something wrong with it. The initiator saved time, but the inspection work fell on others. “You just let AI do the work for you,” Lütke said.
The same is true for emails. An email that could have clearly explained the problem was expanded by AI into a long content. The recipient finished reading and found that the information is not that complicated, and even has to use a large language model to compress it back. Lütke is dissatisfied with this cycle of first expanding and then compressing: since you are already using AI, why not let it help you clarify your point of view, but generate a large paragraph of text that takes others time to process?
This is also the reason why he repeatedly emphasized “responsibility” in the interview. Lütke himself also uses AI to assist in handling complex decisions: he lets different agents collect materials and put forward opinions from the perspectives of data, business and engineering, and then he makes judgments. But he believes that machines cannot take responsibility for the results. For Shopify, the problem exposed by “slop grenades” is therefore not whether employees are using AI, but who checks the final delivered content and who is responsible for the work handed over to colleagues. Lütke is still actively promoting AI; what he is alert to now is that AI can easily make a piece of work that has not been completed carefully look like “already completed”.
For this Shopify executive who once flocked to AI, this may be ironic, but he is far from the only one.
2 Managers and Employees Have Different Perceptions of AI’s “Efficiency Improvement”
An interview by The Guardian found that some companies require employees to use AI more, and even hope that the remaining employees can complete more work with the help of AI after layoffs. But for those who receive these work outputs, the faster generation speed does not necessarily mean less workload: the seemingly finished emails, copies or codes may still need to be verified, modified, or even completely redone.
A copywriter under the pseudonym Ken told The Guardian that the Miami cybersecurity company he works for laid off several colleagues and required the remaining employees to use AI chat tools. The first draft is indeed easier to write, but the time he and his colleagues spent on rewriting and error correction afterwards is even more than when they did not use AI.
Sometimes, they also have to deal with the contradictory content given by the chat tools they use respectively. Ken said that the quality of the articles has decreased, the production time has become longer, and team morale has also been affected.
Similar rework also appears in scenarios with higher accuracy requirements. A student in the joint training program of Doctor of Medicine and Doctor of Philosophy at the University of Michigan found during a survey at a primary care clinic that after medical staff use AI to draft emails for patients, they still need to spend time editing and correcting errors, and worry that wrong replies will be sent to patients. For these jobs, “writing an email” is only part of the process; whether the content is accurate and can be sent to patients still needs to be judged by humans.
This type of content is called “workslop” by researchers: it looks complete and fluently worded, but lacks reliable information or the judgment required to complete the task, and eventually leaves the work to the colleagues who take over.
In April this year, a survey of American office workers conducted by researchers from BetterUp and Stanford University showed that about 40% of respondents had received such content within one month.
The Guardian cited relevant research stating that respondents spend an average of about 3.4 hours per month dealing with it; the estimated productivity loss for an enterprise with 10,000 employees is 8.1 million US dollars.
Managers and front-line employees may also have different perceptions of “efficiency improvement”. The Guardian cited another survey targeting American white-collar workers, stating that 40% of respondents in non-management positions said that AI did not save them working time; at the same time, 92% of senior managers believed that AI improved their own efficiency. The two sets of answers measure the subjective experience of different positions and cannot be simply subtracted, but they suggest that enterprises may miss the time for receiving, verifying and reworking when calculating benefits.
There is also noteworthy evidence in the software development field. The research institution METR once asked senior open, and randomly decided which tasks could use AI
As a result, in this experiment targeting tools, specific developers and tasks in early 2025, tasks using AI took an average of 19% more time. The researchers clearly reminded that this cannot be generalized as “programmers will all slow down when using AI”. It shows that after the tool generates the code, the time required to understand the project background, check the results and complete the delivery also needs to be included in the efficiency calculation.
When Lütke complained about “slop grenades”, he pointed out exactly this easily overlooked link: one person uses AI to hand over materials faster, but another person has to complete the inspection that should have been done before delivery for him. If an enterprise only looks at how much content is generated and how much drafting time employees save, it will be difficult to see clearly whether the whole work has really become faster.
3 When “Using AI First” Becomes a Requirement, Who Will Measure the Follow-up Work?
Shopify’s experience has brought to the surface an easily overlooked problem when enterprises promote AI: the usage rate can be counted, and the generation volume is also easy to display, but verification and rework are often scattered in other people’s schedules.
Lütke’s 2025 memo was intended to solve another real problem: employees may follow the old process and put forward demands for additional resources without seriously trying new tools. Requiring the team to first explore the capabilities of AI has a clear management logic. However, employees will also respond to the requirements of the organization. When “how much AI is used” is easier to see than “how much extra work others need to do after delivery”, a longer and faster output material is not necessarily a more useful material.
The real question worth asking is therefore not whether employees should use AI, but what counts as completing the work.
After an email is sent, the task does not end automatically just because the text has been generated; after a piece of code is submitted, it does not automatically have maintenance value just because it can run. If the delivery standard only stays at “what is produced”, verification will easily be pushed to the next link. If the standard includes “whether the person who takes over can continue to act based on it”, the person using AI needs to complete the necessary checks before sending.
For the person writing the email, this means deleting the unnecessary elaboration that does not help decision-making, verifying links and facts, and making clear the questions that need to be answered by colleagues. For the person writing the code, this means explaining the purpose of the change, checking the critical path, and confirming that the recipient can understand its impact. There is no mysterious rule exclusive to AI here: whoever submits the work is responsible for the quality of the work first.
AI just makes it easier to skip this step, and also allows problems to enter team collaboration in larger volumes.
Lütke named “slop grenades” and also pointed out the way it harms efficiency. Next, Shopify and other companies that require employees to use AI widely will face the same more difficult problem: when evaluating a piece of work, can they see the time borne by others after generation. Otherwise, the personal efficiency figures of employees may rise, but the team is still busy defusing bombs.
https://futurism.com/artificial-intelligence/shopify-ceo-ai-workplace-staff-horrified-backfire
https://finance.yahoo.com/technology/ai/articles/shopify-ceo-made-staff-prove-105644220.html
https://www.shopify.com/news/unserious-exploration
https://futurism.com/artificial-intelligence/ai-workslop-gridlock
This article is from the WeChat official account “InfoQ” (ID: infoqchina) published with authorization from 36Kr
