AI Slop Is Costing Businesses Real Money: The Documented Cases, and Why Most AI Projects Fail

Something strange happened over the last three years. AI became the easiest expertise in the world to claim and the hardest to verify. Anyone can wire a language model into a chat widget in an afternoon, call it an AI transformation, and invoice for it. The gap between that and a system a business can actually stake its reputation on is enormous, and the businesses paying the price for that gap are rarely the vendors. They are the clients.

Gartner has a name for one flavor of this: agent washing, vendors rebranding ordinary chatbots and automation scripts as AI agents. Of the thousands of vendors calling themselves agentic AI companies, Gartner estimates only around 130 are building the real thing. The rest are selling packaging. And the results are now well documented enough that nobody gets to claim they were not warned. What follows is not speculation or fear of new technology. Every case below is on the public record, with sources linked at the end of this post.

The chatbot stories are embarrassing. The operational ones are expensive. Zillow bet its balance sheet on a pricing model, buying thousands of homes based on algorithmic estimates. When the market shifted in 2021, the model kept overpaying, and Zillow shut the entire division down, took write downs exceeding $500 million, and laid off around a quarter of its workforce. McDonald's spent three years piloting AI drive through ordering with IBM across more than a hundred locations, then ended the partnership in 2024 after viral videos of the system adding 260 Chicken McNuggets to an order and putting bacon on ice cream.

Klarna is the most instructive case because the company said the quiet part out loud, twice. In early 2024 it announced its AI assistant was doing the work of 700 customer service agents and froze hiring. By May 2025 the CEO was publicly reversing course and recruiting humans again, admitting that organizing around cost as the dominant factor produced lower quality service. And iTutorGroup paid $365,000 to settle the first EEOC lawsuit over AI hiring discrimination, after its recruiting software automatically rejected older applicants, more than 200 of them, based on age alone.

None of these companies lacked budget, talent, or access to the best models available. What they lacked was scope discipline: each one handed AI an entire judgment heavy job instead of the bounded, verifiable piece of it the technology could actually carry.

The same research that documents the failure rate also documents the escape route. McKinsey tested 25 attributes and found workflow redesign has the single biggest effect on whether generative AI produces real earnings impact: the winners redesign the process around the tool instead of bolting the tool onto the old process. CEO level oversight of AI governance is the factor most correlated with bottom line impact at large companies, yet only 28% of organizations have it. And the MIT study found external partnerships with specialists succeed around 67% of the time, roughly double the success rate of internal builds, mostly because specialists have already made these mistakes on someone else's pilot.

RAND's prescription list is almost boring, which is exactly the point: pick the problem before the technology, invest in data quality up front, make sure leadership understands what AI cannot do, and commit past the pilot phase. Nothing in that list requires a frontier model. All of it requires discipline that the overnight AI expert selling a chat widget has no incentive to bring.

We build AI systems for a living, so this post is not an argument against the technology. It is an argument against shipping it carelessly, and the discipline shows up concretely in our own work. The loan approval system we built holds 85% model accuracy and cut processing time by 40%, but the design decision we are most confident in is that no application is ever decided by the model alone. The AI recommends with a stated reason, a human verifies every single case, and the verified decisions feed back into the model. That is the exact opposite of the unreviewed output pattern running through every disaster above.

The same restraint applies to scope. The trading automation we built cut manual effort by 80% and lifted revenue 18%, and it still does not choose the strategy. The client supervises strategy, the system executes it flawlessly at a speed no human can match. When a client asks us for a customer facing bot, the first deliverable is not the bot, it is the boundary: what it is allowed to say, what it must refuse, what gets escalated to a human, and what happens when someone deliberately tries to break it. If a vendor cannot describe that boundary for the system they are proposing, they are selling you a future incident report.

The projects in this post did not fail because their budgets were too small. Most of them had enormous budgets. They failed because the work was not engineered for the messy reality of production, and that discipline is available at every price point, including from smaller studios that charge a fraction of what the failed projects cost. What guardrails and proper scoping cost for your use case is a question with a specific answer, once you share the use case.

We build AI systems with the boring parts included: fallbacks, human review where it matters, and testing against the inputs real users actually produce. Tell us what you want AI to do in your business, and we will quote the version that survives contact with production, priced against the return it has to generate.