The Slopification Crash: How Rushed Deployment and Usage-Based Pricing Broke the Numerator and the Denominator

Two numbers frame the post-bubble picture. MIT's review of more than three hundred publicly disclosed generative AI initiatives found that ninety-five percent of them, against roughly thirty to forty billion dollars of enterprise investment, produced no measurable return. A Deloitte survey of nearly two thousand executives, taken as agentic AI adoption accelerated, found that only ten percent of organizations were realizing significant returns on that spend. The crash the market is now digesting was not a technology failure. It was the arithmetic catching up.

Two failures, one crash

The collapse has two independent causes that compounded. The numerator — the value AI was supposed to produce — was eroded by what the trade press has begun calling slopification: rushed deployment, absent governance, and output quality low enough that the rework cost of using the system exceeded the labor cost of not using it. The denominator — the cost of running the systems — was inflated by the industry's rapid shift from fixed-price subscription software to consumption-based pricing, which turned every model call into a variable expense with no ceiling.

The numerator: slopification

Reporting by TechTarget's Carrie Pallardy documents the pattern in detail. Enterprises rushed to be labeled AI-native. Employees were given tools and told to experiment. Individual experimentation quietly became departmental practice, then production dependency, without ever passing through a governance review. What was produced was, in the executive summary's phrase, average at best and confabulated at worst — and the cost of cleaning it up compounded from there.

• Rework cost — entire projects scrapped and rebuilt to add the governance, security, and data controls that were skipped in the first pass.
• Brand and customer damage — low-quality outputs shipped to customers, discovered late, and remedied at the slowest and most expensive point in the feedback loop.
• Productivity and talent loss — internal teams spending their days fixing AI output produced by peers, with the strongest performers leaving first.
• Disjointed shadow use — parallel AI tooling adopted department by department, with no consolidation, no visibility, and no attributable spend.
• Wrong-vendor lock-in — enterprise agreements signed for the discount, well before the enterprise had enough evidence to know whether the vendor was the right long-term choice.

“If you care about quality, AI slop has a huge rework price tag that may make you run in the red on AI. A lot of times people think about the gross savings — they don't think about the net savings after the rework has been accomplished.”— Tim Sanders, Chief Innovation Officer, G2

The denominator: consumption pricing broke the budget

Global Finance's June 2026 reporting, drawing on McKinsey, traces the second half of the problem. Between 2015 and 2024, the number of consumption-based software companies more than doubled. Agentic AI accelerated the shift, because every autonomous action a system takes generates billable usage. McKinsey's Nicolai von Bismarck describes the resulting cost uncertainty as structural and accelerating, and cites it as one of the top operational barriers to scaling AI. In practical terms, the budgets that were approved as software line items behaved, in production, as variable cost of goods sold.

Why the banks noticed first

The credit market registered the mismatch before the operating market did. JPMorgan Chase, Morgan Stanley, and Sumitomo Mitsui Banking Corp began working to offload data-center loan exposure to private funds and insurers. A thirty-eight billion dollar debt package tied to Oracle was in syndication for six months, ultimately offered at a discount. When the lenders financing the infrastructure begin discounting the paper, the recoverable value question stops being an opinion and starts being a market price.

The dot-com parallel

Sanders draws the comparison plainly. The dot-com crash was filled with companies that did it fast and were wrong. Webvan and eToys are the names that recur because their operating models were built for a growth rate that never arrived, and their records were built for a fundraising cycle rather than for a workout. The pattern this decade is not identical, but it rhymes. Speed was rewarded until the moment the market stopped paying for speed and started asking for evidence.

“There's just no good proof behind 'move fast' with AI. Doing it right was better than doing it fast.”— Tim Sanders, Chief Innovation Officer, G2

What's left to recover

The distinctive feature of this crash, compared with the dot-com bust, is that the artifacts are harder to see. A failed e-commerce company left a warehouse, a fleet, and a customer list. A failed AI initiative leaves a spread of cloud invoices, seat licenses, model API calls, prompts nobody kept, outputs no one owns, and a vendor contract that has been silently amended by model deprecation. The workout question is not whether the capital was lost. It is whether enough of the record survives to make any defensible decision about what is left. The provenance problem explains why so little of it is reconstructable without deliberate effort; the six-disposition framework is what the reconstructed record is fed into.

1. 01Reconstruct the invoice trail — every AI-adjacent expenditure, embedded seat, cloud allocation, and rework hour, mapped to a system or ruled out with a documented reason.
2. 02Reconstruct the value trail — every claimed benefit tested against the operational data that would confirm or contradict it.
3. 03Classify each asset on evidence recoverability, not on original intent — retain, restructure, impair, terminate, or sell.
4. 04Take the workout decision with the uncertainty explicit rather than concealed, so it survives audit, litigation, and diligence.