Why most AI projects fail: it’s infrastructure and people (The New Stack)

An article by Meredith Shubel, The New Stack, 2026-07-06, on why AI projects stall between a working demonstration and a production system.

What is actually recorded here

The body has never been read. The page is client-rendered and returns navigation and signup forms to every extraction route tried — WebFetch twice (including the canonical URL after a redirect), Firecrawl, and a direct HTML fetch and parse. What this wiki holds is the headline, the byline, the date, and the thesis the headline states: the causes are infrastructure and people, not model quality.

That framing is worth recording even unread, because it names the two candidate explanations for the demo-to-production-gap and rules out a third. Everything below the headline — the evidence, the interviewees, any statistics — is unavailable.

Tier

T4. Not a judgment on The New Stack, which is a competent trade outlet and would ordinarily be T3 for a reported piece. The tier reflects what reached this wiki: a title. It establishes that a working journalist framed the problem this way, and nothing about whether the framing is supported.

Named for a later pass: re-fetch by another route, or replace with a source that measures the failure rate rather than asserting it. Until then this page must not be cited for any claim beyond its own headline.

Why it’s here anyway

It is one of the two founding sources for demo-to-production-gap, and it is the counterweight to ml-system-design-case-studies: that catalog collects systems that shipped, this headline points at the ones that didn’t. Keeping an honest stub is the hub’s rule — a source is never dropped for being hard to fetch, and the weakness gets recorded instead of hidden.

Related

demo-to-production-gap · ml-system-design · ml-system-design-case-studies · the-new-stack · synthesis