Feature Theater: The AI Industry's Real Product Is Mediocrity at Scale

Companies are now able to spin up a new UI element or AI toggle in a single afternoon by feeding a large‑language model a prompt, and they ship it to demos as if it were a breakthrough. The “feature theater” described stems from the fact that the engineering effort to prototype has collapsed from weeks to hours, but the deliberation required to assess customer need, prioritize, or reject an idea has not. Consequently, product managers flood dashboards with half‑baked widgets that look impressive in a slide deck but do not accelerate the core workflow they were meant to improve. The phenomenon is amplified by usage‑based billing models that charge per token regardless of answer quality, encouraging more interactions rather than better outcomes.

This sprint‑fast, AI‑first mindset is reinforced by boardroom pressure to quantify AI progress. Metrics such as “AI features per release,” “engineers using copilot,” or “AI interactions per user” have become the headline numbers investors demand. Goodhart’s law kicks in: as these adoption figures become targets, companies inflate them by default‑on buttons, “generate with AI” links, and rebranded assistants. Gartner’s June 2025 forecast warned that over 40 % of agentic AI projects will be scrapped by 2027, citing spiralling costs and unclear ROI, and identified a wave of “agent washing” where existing bots are merely repackaged as agents—only about 130 of the thousands of vendors are truly delivering novel capabilities. The Guardian’s January 2026 expose of Google’s AI Overviews delivering misleading health data underscores how even top‑tier labs can push large‑scale, poorly vetted features into the most visible user spaces, exposing users to harmful misinformation while the product appears successful on engagement dashboards.

The fallout is three‑fold. First, engineering bandwidth is diluted across a flood of low‑impact features, eroding the depth of innovation. Second, customers confront a barrage of superficial AI add‑ons that do not solve their primary problems, risking churn and brand erosion. Third, the token‑based revenue model incentivizes volume over accuracy, potentially attracting regulatory scrutiny as mis‑information or “good enough” outputs proliferate. Stakeholders should watch for a shift from pure adoption KPIs to outcome‑based metrics—such as task completion time reductions or error‑rate drops—and for board inquiries that demand evidence of real productivity gains rather than feature counts.