I’ve been working with a lot of finance teams and consulting clients over the past year who are deep into AI tool rollouts. And almost universally, the story is the same: people feel more productive, output is up, and leadership is happy with adoption numbers. That’s the story getting told in the board deck.
But there’s another story in the data. And a 2026 report from Faros — a company that tracks software development telemetry across thousands of teams — puts numbers to it in a way I haven’t seen before. Their dataset: 22,000 developers across 4,000 teams. Two years of data covering the exact period when AI tools went mainstream.
The headline finding: under high AI adoption, the probability of a production failure tripled per code change merged. Not doubled. Tripled. And organizations with strong, mature processes saw the exact same deterioration as everyone else. Engineering maturity offered zero protection.
They called it the Acceleration Whiplash. I think it’s one of the most important concepts for CPAs to understand right now — not because of software, but because the same dynamic is playing out in finance functions everywhere.
How AI Became the Author
When AI coding tools first launched, the framing was clear: human in charge, AI suggests, human decides. That was the pitch. For a while, that’s how people used it.
But the acceptance rate of AI-generated code — meaning how often a developer takes what the AI produced and puts it directly into the codebase — went from 20% to 60% in two years. Part of that is the models getting better. Part of it is people getting more comfortable just hitting accept. I’ve seen acceptance rates in the high 90s at companies I’ve worked with. Every time the tool asks “can I do this?” — people say yes.
Think about it like a factory line where someone doubled the conveyor belt speed without telling the quality inspector. Output volume doubles. The inspector is now buried. Defects get waved through — not because the inspector stopped caring, but because the math no longer works. You can’t inspect twice as many units in the same amount of time.
Most organizations didn’t slow the belt down. They kept driving adoption.
Two Very Different Pictures
The Faros data tells two stories depending on which metrics you look at.
The adoption metrics look great. Task completion per developer is up 34%. Epics completed per developer up 66%. Code-specific work up 210%. If you’re measuring what people are getting done, AI is delivering.
The downstream metrics tell a different story. Bugs per developer up 54% — and that’s up from just 9% in their prior year’s study, so it’s accelerating, not stabilizing. Incidents per code change merged up 243%. And 31% more changes merged into production with zero human review. Not less review. No review.
Individual productivity up. System quality down. The gap between those two things is what the Acceleration Whiplash describes.
One analyst applied Little’s Law to the lead time data — the time from a completed change reaching the end user — and estimated that despite all the throughput gains, finished output actually reaching customers may be down 70–80%. More output into the system. Less getting through it cleanly.
What This Means for Finance Teams
The software framing is useful but the parallel to finance is direct. Imagine an analyst who used to take two hours on a variance analysis. Now it takes 20 minutes with AI. That’s real. But their manager now has three times as many analyses sitting in their inbox, each of which needs the same level of scrutiny — maybe more, because AI-assisted work can look polished on the surface while missing something important underneath.
The output volume doubled. The review capacity didn’t. That’s the same problem.
Three Questions to Ask Any Client
When a client tells you AI adoption is going well, here’s where I’d push:
- What’s your review policy for AI-assisted output — and how is it enforced? Not “do you have one.” Everyone has one. How is it enforced at the process level? If the answer is “we trust people to review before finalizing,” that’s the 31% problem waiting to happen. Enforcement has to be structural, not cultural. Culture breaks down under volume pressure.
- Has your error or rework rate changed since AI adoption scaled? In finance: reconciliation exceptions, returned deliverables, revised analyses, reopened items. These signals rise on a lag — by the time they’re obvious, the problem is already compounded. Ask now, before it’s a crisis.
- Are you measuring system throughput separately from individual throughput? Individual productivity is easy to measure and easy to report upward. Quality-controlled output actually reaching the end user is harder to track — and usually isn’t. That’s where value destruction hides. If a client can tell you their team is saving 30% of their time but can’t tell you whether deliverable quality has held, they have a visibility gap worth flagging.
The Fix Is Upstream, Not Downstream
The instinct when quality drops is to add more reviewers, tighten gates, extend QA. The Faros researchers are direct about this: that’s the wrong response. More reviewers treat the symptom. The problem has to be addressed at the point where the output is generated — before it ever reaches a human to review.
Governance at the output layer will always be playing catch-up with AI volume. Governance at the authoring layer — how AI is prompted, constrained, and instructed — is where the leverage is. That’s a useful frame for any client conversation about where to invest in AI governance infrastructure.
Key Takeaways
- AI adoption metrics and AI quality metrics measure different things. Track both separately — they will often point in opposite directions.
- No client is protected by process maturity alone. The Faros data found zero protection from strong foundations. The whiplash hits regardless of how good the team is.
- Ask about downstream quality signals — error rates, rework, unreviewed output — not just adoption rates. The gap between those two is where the risk lives.
- The fix isn’t more reviewers downstream. It’s better controls upstream, at the point of authoring.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply