The receipts

AI that paid off — and AI that didn’t.

Forget the hype reels. We pulled the documented projects — the Harvard cases, the SEC filings, the post-mortems — and lined the wins up next to the wrecks. One thing separates them, and it isn’t the technology.

The odds, in three numbers

The backdrop these eight cases sit against — from RAND, MIT and BCG.
AI projects that fail to deliver business value80%
Generative-AI pilots with zero measurable return95%
Companies reaching real value at scale5%
Sources: RAND (2025), MIT (2025), BCG (2025).
What worked

AI that paid off

Different industries, different tools, same move: each one picked a job worth doing, wired the AI into real work, and counted the results. The model was the easy part.

Moderna
Biotech · enterprise AI
750+
custom AI assistants deployed across the company in ~2 months
Instead of one shiny tool, Moderna handed every team the keys — legal, research, manufacturing, the lot — and backed it with real training. Some teams hit 100% adoption; the legal team built its own contract reviewer.
Why it worked: they ran it like an operating-model change, not an IT project. The tech was there for everyone — the effort went into getting people to actually use it.
Sources: Harvard Business School, OpenAI
Klarna
Fintech · customer service
$40M
projected profit improvement; resolution time cut from 11 min to under 2
Klarna’s assistant took two-thirds of all service chats in month one and dropped resolution time from 11 minutes to under two. It later dialed humans back up — proof that even a win needs constant tuning.
Why it worked: one high-volume job, two numbers that actually mattered — speed and repeat contacts — tracked from day one.
Sources: OpenAI, Fast Company
JPMorgan Chase
Banking · legal operations
360,000
annual lawyer-hours of loan-contract review, done in seconds
Lawyers and loan officers used to burn 360,000 hours a year on the same loan clauses. COIN now reads them in seconds — 12,000 contracts, 150 data points each — and made roughly 80% of the costly compliance slips disappear.
Why it worked: a boring, repetitive job with a hard number on it. You can’t prove a payoff you never measured.
Source: ABA Journal
Google DeepMind
Infrastructure · energy
−40%
cooling-energy use in Google data centers
Fed years of sensor data and one clear goal — use less power to keep the servers cool — the model cut cooling energy by up to 40%. No reorg, no drama; just a well-measured problem.
Why it worked: a tightly wired system, mountains of real data, and a target you could read straight off a meter.
Source: Google DeepMind
What failed

AI that fell over

Not one of these flopped because the model was dumb. They died in the gap between a slick demo and the messy real world — missing data, missing oversight, missing the people who’d actually use it.

Zillow
Real estate · pricing algorithm
$500M+
losses; iBuying shut down and ~25% of staff cut (2021)
Zillow let an algorithm buy ~7,000 houses at prices it set itself. When the market cooled, nobody could hit the brakes — a $304M write-down ended the whole business and took a quarter of the staff with it.
Why it failed: a confident model, a volatile market, and no human hand on the wheel when reality moved.
Sources: GeekWire, SEC filing
IBM Watson & MD Anderson
Healthcare · clinical AI
$62M
spent before the cancer-treatment project was shelved
After $62M, Watson for Oncology never treated a single patient. Internal reviews flagged “unsafe and incorrect” advice, and it was never plugged into the hospital’s real medical records.
Why it failed: it lived in a demo, not the clinic — cut off from the data and tools doctors actually use.
Sources: IEEE Spectrum, JNCI
Amazon
Tech · AI recruiting
Scrapped
after the hiring model learned to penalize women
Trained on ten years of mostly-male résumés, Amazon’s recruiting AI taught itself that men were the answer — downgrading any résumé that so much as mentioned “women’s.” Engineers couldn’t fully un-teach the bias, so it was quietly killed.
Why it failed: point AI at a biased process with no guardrails and you don’t fix the bias — you scale it.
Source: MIT Technology Review
McDonald’s
QSR · voice ordering
100+
restaurants pulled the AI drive-thru after a 2-year test (2024)
Two years and 100-plus restaurants in, the voice-ordering AI still choked on noise, accents and crosstalk — once piling bacon onto a customer’s ice cream. McDonald’s pulled the plug in mid-2024.
Why it failed: dropped into the loud, chaotic real world before it was ready for it — and customers felt every miss.
Source: CNBC
The pattern

It was never about the model.

Eight projects, eight capable models. What split the wins from the write-offs was the unglamorous last mile — whether the AI actually reached how the business runs, with the people, data, conditions and oversight to make it stick. That last mile is the whole job. It’s also the part almost everyone skips.

The discipline
Winners
Failures
Picked a clear, narrow target
Set a baseline and measured the result
Built into real workflows and live data
Drove genuine adoption with the team
Kept humans in the loop

Which list do you want to be on?

That gap — between a model that works and a business that’s better — is exactly what we close. Let’s find where AI actually pays off for you.

Book a free intro call

Sources

Every figure on this page is drawn from public reporting and published research. Facts and figures are summarized in our own words; follow the links for the originals.

Moderna — Harvard Business School: “Moderna: Democratizing Artificial Intelligence”; OpenAI
Klarna — OpenAI; Fast Company
JPMorgan COIN — ABA Journal
Google DeepMind — Google DeepMind
Zillow Offers — GeekWire; SEC 8-K
IBM Watson & MD Anderson — IEEE Spectrum; JNCI
Amazon recruiting AI — MIT Technology Review
McDonald’s drive-thru — CNBC