Major insurance companies are actively lobbying state regulators to carve out exemptions from covering AI-related workflows. That single fact tells you more about where enterprise AI actually stands than any benchmark or product launch.
The detail comes from Dorian Smiley and Connor Deeks, co-founders of AI advisory firm Codestrap and former PwC consultants, speaking to The Register. Their argument is blunt: most businesses using AI are measuring the wrong things, the quality problems are about to surface, and the financial fallout will be severe.
The Metrics Lie
Smiley's core claim is that companies track vanity metrics like lines of code and pull requests, which look great when AI is generating output. But the metrics that actually matter - deployment frequency, change failure rate, mean time to restore, incident severity - show no improvement. The AI-generated work passes unit tests but breaks in production.
The most concrete example: an AI-generated rewrite of SQLite in Rust produced 3.7 times more code while performing 2,000 times worse than the original. Not 2,000 percent worse. Two thousand times worse. That's not a rough draft that needs polish. That's a non-viable product.
Deloitte had to issue refunds to the Australian government after delivering a 440,000-word report riddled with AI-generated errors. When a Big Four firm can't catch the mistakes, smaller companies with less review infrastructure don't stand a chance.
The Insurance Signal
Insurers don't make emotional decisions. When they start excluding AI from liability policies, they're doing it because their actuarial models say the risk is unpriced. Smiley reports that underwriters are "seriously trying now to remove coverage in policies where AI is applied and there's no clear chain of responsibility."
This creates a feedback loop. Companies that can't get insurance coverage for AI-assisted work will either need to prove human oversight at every step (expensive, defeating the cost savings) or accept uninsured liability (reckless for anything consequential).
The Consulting Problem
There's a structural incentive problem in how AI gets deployed. Consulting firms make higher margins when AI does the work and fewer humans review it. The same firms selling AI transformation services have zero financial incentive to slow down and check quality. As Smiley puts it, clients are already pushing back: KPMG reportedly pressured another accounting firm to lower prices after discovering they were using AI, treating it as a reason to pay less rather than a premium feature.
Smiley estimates quality problems will surface within eight to nine months for heavy AI users, bringing lawsuits from bad advice and accelerated system failures.
None of this means AI is useless in business. It means the current approach of deploying fast, measuring output volume, and skipping rigorous quality checks is heading for a correction. The insurance industry just got there first.