The productivity case for AI tools is built on a straightforward premise: if AI can generate a draft, analyze data, or produce a report in seconds, the time saved represents a direct efficiency gain. Research from Zapier introduces a complication that the straightforward version of this premise does not account for. Only 2 percent of AI outputs are ready to use without revision or correction, which means that 98 percent of the time, someone is spending additional hours editing, fact-checking, rewriting, or discarding what the AI produced. The efficiency gain from faster generation is real. The cost of supervision, correction, and cleanup is also real, and most organizations are measuring the former while ignoring the latter. The net result is teams that feel busy but ineffective, budgets absorbing costs that do not appear in AI tool line items, and the particular organizational damage that comes from publishing or sharing AI output that was not adequately reviewed before it reached its audience.
The term that has emerged for low-quality AI output that looks complete but lacks substance, accuracy, or genuine value is AI workslop, and the business costs it generates extend beyond the hours spent fixing it.
What the 98 Percent Revision Rate Actually Means for How You Are Accounting for AI
The Zapier finding is worth examining in terms of what it reveals about how AI productivity is being measured rather than just what it reveals about output quality.
When an organization evaluates whether AI tools are delivering productivity gains, the measurement that tends to be captured is the time between task initiation and initial output generation. An AI tool that produces a draft report in two minutes rather than the two hours a human would require represents a substantial apparent efficiency gain. What is rarely measured with equivalent rigor is the time required to bring that draft to the standard at which it can actually be used: the fact-checking that catches errors the AI introduced confidently, the structural revision that addresses the coherence problems that AI drafts characteristically produce, the contextual adjustment that replaces generic content with the specific knowledge that the situation requires, and the judgment calls about what the AI included that should not be there and what it omitted that should be.
When those supervision and correction costs are measured against the generation time savings, the net efficiency calculation looks different from the gross calculation that most organizations are using to justify AI tool investment. For some tasks and some users, the net calculation is still strongly positive. For others, particularly tasks that require significant contextual accuracy, nuanced judgment, or output that will be attributed to the organization in contexts where errors carry reputational or legal consequences, the net calculation may be negative. Producing AI output and then spending the time required to correct it to an acceptable standard can take longer than producing the work from scratch with human expertise, while also introducing the cognitive overhead of switching between generation and critical evaluation modes.
The business costs that Zapier’s research associates with this dynamic range from hundreds of dollars per employee per month for smaller teams to millions annually for larger organizations. These are costs that are not appearing in AI tool subscription budgets. They are appearing in employee time that is being consumed by cleanup rather than by the higher-value work that AI was supposed to free up capacity for.
The Trust Damage That Circulates With Uncorrected AI Output
The efficiency cost of AI workslop is the cost that is most directly measurable. The trust damage is the cost that accumulates more slowly and produces more durable organizational harm.
When AI output that has not been adequately reviewed is shared with colleagues, clients, or external audiences, the errors, generic framing, and factual inaccuracies it contains are attributed to the person who sent it, not to the AI tool that generated it. A colleague who receives a report containing factual errors does not conclude that the AI tool failed. They conclude that the person who sent the report did not apply sufficient care or expertise to what they were sharing. The reputational consequence attaches to the human sender, and it attaches in ways that affect how that person’s future work is received, how their judgment is evaluated, and whether they are included in work streams where quality and reliability matter.
This dynamic creates an organizational incentive structure that is the opposite of what AI adoption is supposed to produce. Employees who care about their professional reputation become more cautious about using AI tools because they have experienced or observed the reputational cost of sharing output that did not meet the standard their audience expected. The employees who are least attentive to output quality continue to share AI-generated content without adequate review, accumulating reputational costs they may not immediately perceive. The net effect is that AI tool adoption increases organizational output volume while potentially degrading the quality signal that audiences use to evaluate the reliability of what they receive.
The collaboration damage compounds over time. Teams that have experienced receiving low-quality AI-generated output from colleagues develop skepticism about AI-assisted work that extends beyond the specific instances that warranted it, creating friction around AI adoption that would not exist if the output had been appropriately reviewed before sharing.
Why the Junior Team Member Framework Produces Better Outcomes
The deployment model that consistently produces better net productivity outcomes frames AI tools not as autonomous output generators but as capable assistants who require guidance, oversight, and review before their work is ready to use.
This framing changes what AI tools are asked to do and how their output is treated after generation. A junior team member who produces a first draft is expected to produce something that requires experienced review and revision before it is finalized. That expectation is not a criticism of the junior team member’s value. It is an accurate description of their role in the production process and a realistic calibration of what their output requires to become usable. Treating AI output with equivalent expectations, as a starting point that experienced judgment needs to refine rather than a finished product that needs only light review, prevents the quality gaps that inadequately supervised AI output characteristically contains.
The task selection that makes AI assistance genuinely valuable follows from this framing. AI tools add the most net value on tasks where the generation speed benefit is high, the accuracy requirements for initial output are relatively forgiving, and the human review required to bring the output to final standard is modest relative to the time saved. Brainstorming and ideation, initial research aggregation, structural outlines for documents that humans will draft from, first-pass data processing, and routine communications that follow established templates represent the category where net productivity gains are most reliably positive.
AI tools add the least net value, and can produce negative net value, on tasks where accuracy requirements are high, where contextual knowledge that the AI does not have is essential to correct output, where the output will be attributed to the organization in contexts where errors carry significant consequences, or where the human review required to bring AI output to standard approaches the time that producing the work from scratch would require. Identifying these tasks and keeping them human-led is not a failure to capture AI’s potential. It is an accurate application of where AI’s potential is actually positive.
Building the Organizational Practices That Make AI Investment Pay Off
The practices that convert AI tool investment from a source of busywork to a genuine productivity driver require explicit design rather than emerging naturally from making tools available.
Defining which tasks receive AI assistance and which remain human-led removes the ambiguity that leads to AI being applied where its limitations produce more cost than value. This is not a permanent categorization. It should be revisited as AI capabilities develop and as the organization develops better data on where AI assistance is producing net positive outcomes and where it is not. But the absence of any categorization, the default of applying AI wherever it seems potentially useful and evaluating the results informally, is what produces the widespread cleanup costs that Zapier’s research documents.
Prompt quality training is the investment with the highest return relative to its cost for organizations that are serious about AI productivity. The quality of AI output is substantially determined by the quality of the instructions it receives, and most employees using AI tools have not been taught how to construct instructions that produce outputs requiring minimal correction. Training that teaches employees how to provide the context, constraints, and specificity that AI tools need to produce useful outputs reduces the revision burden that 98 percent revision rates represent, without requiring any improvement in the underlying AI technology.
Measuring actual net productivity, the efficiency gain from AI generation minus the time cost of supervision and correction, rather than measuring AI adoption and output volume as proxies for productivity, gives organizations the data they need to evaluate whether their AI deployment is working. The organizations that make this measurement honestly are the ones that will identify where AI is genuinely adding value and where it is generating the appearance of productivity while consuming the time that real productivity would require.
