The Queue That AI Helped Create

Code review is where we see this show up first and most visibly. A pull request written by a senior engineer typically came with an implicit signal: someone competent thought this through, so a reviewer could focus on judgment calls rather than basic correctness. A pull request written by an AI assistant carries no such signal. It compiles, the tests often pass, and the logic can still be subtly wrong in ways that take a careful human longer to find than it would have taken that human to generate the code in the first place. Multiply that by volume, and the review queue doesn’t just grow — it changes. Reviewers are no longer skimming for whether the approach was right; they're checking for plausible-looking mistakes buried in otherwise clean-looking work. The data bear this out: GitClear’s January 2026 analysis of 623 million code changes found code duplication up 81% since 2023 and refactoring collapsing to under 4% of changed lines⁽¹⁾, while Stack Overflow's February 2026 developer survey found that only 33% of developers trust the accuracy of AI output, even as nearly four in five developers use AI tools in their own development work⁽²⁾.

We’ve seen product specs follow a similar arc. A spec used to represent hours of a product manager's thinking, which meant a stakeholder reading it could reasonably assume the hard tradeoffs had already been made. Now, a spec can be generated from a prompt in minutes, and it can read just as polished as one that took a week: full sentences, clean structure, plausible rationale. The polish is no longer a signal of the thinking behind it, so the reviewer has to do the thinking the document was supposed to have already done. This pattern now has a name: “workslop” — AI-generated work that masquerades as good work but lacks the substance to meaningfully advance a task. In research published by Harvard Business Review, 40% of U.S. full-time employees surveyed had received workslop in the prior month, and each incident cost its recipient nearly two hours of rework — an invisible tax the researchers put at $186 per employee per month.⁽³⁾

The Fix That's Compounding the Problem

Faced with review queues they can no longer keep up with, we’ve seen a growing number of companies have reached for an obvious next step: use AI to do the reviewing too. Let a model triage pull requests, flag likely issues in a spec and summarize what changed and whether it looks sound.

This is where the story gets worse instead of better. While today's review agents are improving quickly, capability is not the issue; accountability is. The best review agents already ingest a codebase’s history, style rules and even product context, and they will keep getting better. This makes the reviewer model well-suited to catching the kinds of errors a generator would itself make, but, in our view, poorly suited to catching the kinds that come from nobody having asked whether the work was worth doing at all. The result is two AI systems in a closed loop, one generating, one nodding along, with human judgment pushed further from the process precisely when the volume of unvalidated output is at its highest.

The Other End of the Ledger

While this plays out inside engineering and product organizations, another problem has been unfolding: the AI bill. Inference spend, per-seat licenses, review tooling and the human hours spent cleaning up after AI-generated work rarely land on a single line item, so leadership often sees rising costs scattered across a dozen categories rather than one clear number. In some organizations, AI spend may outpace any value it can credibly be said to have produced, which reflects the difficulty in measuring the ratio, rather than any failure of the tools themselves.

The pattern connects directly to what we see happening in engineering and product. Every hour a senior engineer spends checking a subtly wrong AI-generated PR, or a product leader spends catching a spec that skipped the hard tradeoff, is a cost the AI bill may not show but the P&L eventually does. Workslop isn't just an engineering-quality problem; it's an unbudgeted cost.

For growth-stage companies, where the people writing the roadmap are often the same people reviewing the output, the exposure can be particularly acute. And boards are now asking management teams a pointed question: what is your AI spend actually buying? Among the growth-stage companies we work with, the leading organizations treat review capacity and cost attribution as first-order constraints on AI adoption rather than afterthoughts.

Breaking the Loop

While the instinct is to slow AI adoption down, we believe that that's the wrong lever. The tools aren't the problem; the missing judgment layer is. We recommend CEOs and their leadership teams take four steps to avoid the pitfalls of workslop — none of which requires slowing adoption.

Route review by risk, not by source

Set explicit tiers. A configuration tweak, a UI copy change and a payments-logic change don't belong in the same review lane, regardless of who or what wrote them. Define what routes to a senior reviewer, what routes to a peer and what can merge on tests alone, so scrutiny scales with blast radius instead of headcount.

Never let AI be the only reviewer of AI

Put a named human on every review chain who owns the sign-off, not just the tool that generates a summary or a risk score. If an AI reviewer clears something, treat that as one input to check, not as the final call.

Require a decision before you generate, and a threshold before you ship

Before a spec, pull request or prototype gets built, write down the decision it's meant to inform and the bar it has to clear, such as a metric, a stakeholder sign-off or a specific tradeoff resolved. If none exists, it's not ready to build yet, no matter how fast it would be to produce. This is the same discipline I outlined in “Disciplined Decision-Making in the Age of AI” — define the decision before you commission the work — applied here to the build pipeline.⁽⁴⁾

^‍Track AI cost against outcomes, not activity

‍Start by separating line items. AI spend is often scattered across model-provider invoices, per-seat tool subscriptions and the general compute bills of cloud and data platforms, which can make the total cost difficult to see without a deliberate review. Tag AI spend at the source, attribute it to initiatives and pair the cost line with the metric it's supposed to move. Attribution is fast becoming table stakes: in the FinOps Foundation’s 2026 State of FinOps survey, 98% of organizations reported managing AI spend as part of their cost practice.⁽⁵⁾