A shopper grabs five tomatoes and walks

out. Who knows how much they paid when

the cameras guess wrong? A pizza bakes

in a moving truck. Who stops the cheese

from sliding? A remastered Gone with the

Wind hits streaming. Who approves it

when the faces look off? Automation

works well when there is an army of

people behind it. Today's story is about

five failed AI pivots where product bets

were hit with physics, market reality,

and margins. We will study five big

cases where the unit economics snapped

and the product failed and figure out

what we can learn from them. Let's dive

in.

Amazon's just walk out was supposed to

be a checkout free project at grocery

and convenience stores. The idea was to

let customers enter a store, pick up the

items, and leave without scanning

anything or stopping to pay at the till.

The project launched in 2018, and later

on they rolled it out in Amazon Go and

Amazon Fresh stores. The idea was to

create a new age frictionless shopping

experience. The Just Walk Out project

was the first example of a massive

investment in AI retail automation, but

nevertheless, the company retired the

system from its stores in 2024.

Let's understand why. Amazon reportedly

spend around $1 million annually on Just

Walkout between 2019 and 2020. This

investment covered all R&D and capital

expenses. Now, how much is it? a billion

dollars. Is it a lot or is it a little?

The implementation of this system in

large supermarkets required from 10 to

15 million per store, which is far above

traditional checkout systems. Behind the

automation curtain, Amazon hired more

than 1,000 people in India labeling

videos in real time. That human layer

kept things accurate, but it also

destroyed the margins. The cost of

processing data and fixing mistakes

outpaced a normal retail operation,

especially in big box stores with

complex product mixes. It is a lot

easier to implement a system like this

at a convenience store like Amazon Go or

Amazon Fresh. And the reason for that is

because the product mix and the possible

combination of products at those stores

are quite limited. Now, if you think

about a store like Walmart, for example,

you can get a few tomatoes that need to

be weighted. You can get a box of

tomatoes that needs to be scanned and

perhaps a pillow that is 25% off if you

get a toaster by the end of the week.

When you work with a product mix like

that, and that's just a tiny example,

Amazon's just walk out starts to produce

a lot of errors, which is why they

struggled with the adoption among the

large stores. It was estimated that up

to 70% of transactions required human

intervention. The labor cost savings

expected from removing cashiers were

wiped out by the substantial off-site

human operation. And if you think about

it, the need to have the humans behind

it really defeats the purpose of

automation in the first place. The

project also required lots of hardware,

thousands of cameras, shelf sensors, a

backend infrastructure for computer

vision and AI, all of which resulted in

high setup and maintenance costs. In

this business model, stores would have

needed to dramatically increase sales to

offset capital and operational expenses.

And that was the threshold that Amazon

never met. The end result, Amazon

continues using this project and smaller

Amazon Go stores, but they have

abandoned the grocery store application.

You know how every meeting ends in the

same way. 10 people nod, three people

take notes, and only one remembers what

got decided. If you're lucky, you might

say, "But costs get recorded." Now,

sure, if you're paying for those

recordings, and even if you do, what

happens after? You've got a pile of text

that you still need to do something

with, creating action items, updating

tickets, drafting follow-ups, or

cleaning up the chaos. And this is why

we partnered with Radiant, because in a

world where you're paying for every

single app, [music] the smartest move is

to pay for one that replaces five. And

Radiant is actually totally free while

it's [music] in beta, so you don't have

to pay for anything at all. I personally

love tools that work across multiple

platforms. It's a lot cleaner. It gets

you the best value for your money, and

you don't have to remember which app

handles what. Radiant is an AI meeting

assistant that captures your calls,

converts them into notes, and then turns

those notes into completed actions.

Slack messages, project updates, backlog

tickets, drafted documents, and even

prototypes. And it works across Zoom,

Teams, Google Meet, Notion, Cursor,

Lovable, Asana, ClickUp, and so many

more. So your follow-ups appear exactly

where you already work. Radiant is not a

bot that joins your calls. It runs

silently in the background and stores

your transcripts locally for privacy so

that they never leave your device.

Radiant takes care of the admin [music]

work that drains your time. And instead

of dumping a pile of text on you after

the meeting, it immediately takes

action. It's available for M1 and newer

Macs. And it already feels like a

glimpse of what future of workplace AI

should feel like. Multi-functional,

crossplatform, invisible, and secure.

So, if you want to know what it feels

like to have your meetings end and your

work is already [music] done, check out

Radiant using the link below. No bots,

no setups, just results. Huge thanks to

Radiant for sponsoring this part of the

video.

Moving on to IBM's Watson Health. IBM's

Watson Health is perhaps the most

expensive AI pivot failure in corporate

history. The one that has burned through

four to five billion dollars in cash

over a decade. And in the end, it ended

up being sold for a fraction of the

money that was invested. IBM is an

example of how AI marketing can outpace

technological reality. IBM's failure is

rooted in extremely high operational

costs, deeply flawed unit economics, and

glaring gaps between AI promise and

practical medical results. IBM

positioned its product Watson as a

revolutionary cancer diagnosis

technology. They also partnered with MD

Anderson Cancer Center. Now, if you

aren't familiar with the world of cancer

care, MD Anderson is a pretty big deal.

The promise was ambitious. AI that could

analyze vast amounts of medical

literature and patient data to recommend

personalized and tailored cancer care.

Onto the unit economics and costs. IBM

spent an estimated $5 billion on

healthcare acquisitions. And that was

done with a goal to feed Watson's AI

engine with data. At its peak, Watson

Health employed more than 7,000 people

of various medical and scientific

specialties, which meant a massive fixed

cost spent in salaries, technical teams,

and support staff. Watson's cost per one

oncology diagnosis or support case was

reportedly much higher than usual. And

on the consumer side, consumer being a

hospital, the product ended up being

very expensive to purchase and maintain

with no clear ROI and most importantly

no proven clinical benefit. The MD

Anderson Cancer Care Partnership alone

consumed $62 million without producing a

deployable product. Watson Health's

implementation at hospitals was often in

the millions per contract but required

additional budget for ongoing support

customization and integration with

electronic health records and electronic

health records is the backbone of any

hospital. Most hospital deployments

required custom configuration because

Watson struggled with unstructured

patient data. System costs grew faster

than revenues. IBM never scaled Watson

Health to profitability. Their peak

financial performance was when they

broke even by slashing jobs and

eventually selling the unit for about $1

billion, which is the loss of about 80%

of investment. So why did it fail?

What's in health is an example of a

fundamental misalignment between AI

expectations and the reality of a

medical practice. IBM's marketing

promised revolutionary breakthroughs the

technology couldn't deliver. They did

not account for realworld messy patient

records and contextrich doctor's notes.

Watson could only analyze well

ststructured data. A minority at most

hospital databases. Doctors, the

ultimate users of Watson, found it hard

to use and often preferred their own

expertise. All of this resulted in many

abandoned deployments after months of

frustration with complicated interfaces

and most importantly unreliable

recommendation system. And on top of

that, privacy concerns and multiaceted

healthcare laws were not addressed by

the product. In the meantime, Oracle and

Microsoft took advantage of the fact

that IBM was distracted by Watson's

complexity and issues and decided to

make large and focused acquisitions that

yielded clear enterprise value. End

result, by 2021, Watson Health became

unprofitable, which prompted IBM to sell

it. All in all, what's in health

enormous losses were a result of

unrealistic expectations from AI,

unchecked spending, hight touch

implementations, and AI that never

matched its marketing narrative, and the

needs of clinical medicine.

Moving on to Netflix, AI content

upscaling. In early 2025, Netflix pushed

to use AI for content enhancement and

upscaling, which draw a lot of attention

and multiple controversies, especially

around the mangled restoration of

classic shows. Customer backlash has

been intense and on top of this, their

use of AI triggered broader questions

raised about the role of artificial

intelligence in media and specifically

in media preservation, content quality,

and authenticity. Netflix is a prime

example of PR crisis around AI. Let's

look at what happened. Netflix used AI

upscaling to convert older, lower

resolution television shows to high

definition, notably A different world.

What they did is that they bypassed

traditional manual remastering that

requires original film negatives and

detailed human restoration work. The

goal behind this move was rapid,

cost-effective restoration of old film.

But as it turned out, AI really

struggled with grainy source materials.

It produced lots of distorted visuals,

blurred faces, overly smooth textures,

and extremely distorted and unrealistic

backgrounds. Netflix also faced massive

backlash when fans discovered AI

generated promotional posters for Arcane

season 2. Those posters included

telltale signs of artificial generation

included distorted anatomy and unnatural

details. The controversy reached 6

million viewers and forced Netflix to

remove the content. In Netflix's case,

unlike Amazon Just Walk Out or IBM

Watson, it's less about lost revenue or

burned cash and a lot more about damaged

customer trust and brand perception.

Let's go through the unit economics of

this move. If we put ourselves in the

shoes of Netflix and look at it from the

business perspective, it makes sense to

try AI and use it for content upscaling

because manual remastering is very

expensive and slow. And if there is a

part of it that can be restored by a

model so you can delegate the most

delicate work to an experienced

professional, why not do it? AI

upscaling can process large libraries at

scale with minimum labor. Netflix

invested heavily in machine learning for

video restoration and the key goal was

to reduce the per title cost. But in my

opinion, it's one of those things that

need to be done in a waterfall sequence.

It only works when it's done well

because when it fails, the brand damage

and customer dissatisfaction offset

those savings 10fold. So why did it

fail? The models that Netflix deployed

couldn't faithfully recreate the

artistic intent, colors, or subtle

details of the original footage. Instead

of enhancing nostalgia, that is actually

very much a trend, especially among Gen

Z, with non-conforming anti-perfect

aesthetics, it alienated longtime fans.

Another curious thing is that the

critical flaws with the upscaled content

like garble text, cut off scenes,

blurred faces, all of them pass

Netflix's quality checks. And that in

turn signals extreme over reliance on

algorithmic solutions and the absolute

need for human oversight. AI was used

not just in upscaling but also to

manipulate actors mouth movements and

dubbed films which in turn raised

broader ethical concerns over artistic

integrity and performer rights. This was

not the first time when Netflix faced

backlash for AI generated visuals in

original documentaries and marketing

which they did not adequately address.

All in all, to me, this is a PR failure,

even more so than the technical.

Technical, too, for sure, but mostly PR.

This entire situation has triggered

industry-wide debate on the limits of

automation in creative media. So, this

is a lot more than just Netflix, meaning

this whole situation had an

industry-wide impact.

Moving on to Zoom Pizza. Zoom Pizza was

an ambitious softbankbacked

startup whose ambition was to

revolutionize pizza delivery. Their

vision was to cook pizzas inside trucks

and deliver them to customers homes much

faster than any other traditional pizza

chain. Trucks were filled with dozens of

internet connected ovens. Orders were

placed through the app. Ingredients and

toppings were robotically assembled,

although some human workers did some

prep. And pizzas were cooked just before

delivery arrival. The goal was maximum

freshness and efficiency. But there was

one painfully simple problem that sort

of killed the business. Cheese kept

sliding off pizzas while they were

cooked and moving trucks. If you're

thinking that it doesn't sound like a

big problem, why close the entire

business because of the cheese? I'd like

to argue and say that on the surface it

actually does sound like a good idea. I

understand why they got funding, but I

would like to invite you to play product

management 101 with me and really

understand the problem. But before we go

there, unit economics and cost

breakdown. Zoom's entire automation bet

was based on staggering capital and

operational costs in an industry known

for very tight margins. They raised $445

million from Soft Bank, but they did not

do a due diligence calculating

operational costs. Zoom had one

fundamental flaw. Despite hundreds of

millions in funding and teams of

engineers, Zoom forgot about the physics

of cooking pizza and moving vehicles.

And when the business began operating,

the cheese was simply sliding off. They

had teams of engineers trying to solve

the cheese problem after years of R&D.

But the entire operation got essentially

defeated by simple physics and realworld

conditions that robots could not

overcome. Zoom's trucks cost millions to

design and build. with each truck

fitting 56 mini ovens and was supplied

with highly advanced GPS-driven

scheduling. Now, I haven't found a lot

of data on the whole thermo isolation

piece of it. But can you imagine having

a truck full of gas with 56 working

ovens inside of it? Ingredient costs

were around $6 per pizza and the selling

price was set at $18 or more per pizza.

Now, the price, in my opinion, is

actually quite competitive. Yeah, it's

not cheap, but it is a typical price of

a restaurant quality pizza. Soft Bank

invested $375 million in 2018 and valued

Zoom at $2.25 billion based on visions

of becoming the Amazon of pizzas. They

saw Zoom as part of their strategic

portfolio alongside food delivery

companies like Uber. Now, Zoom had other

investors as well, but the largest sum

came from SoftBank. For Soft Bank, the

investment became a complete loss

because Zoom shut down in 2023. They

tried to make a pivot by 2020. They

abandoned the pizza delivery service as

a whole and pivoted to sustainable

packaging manufacturing using the very

robots that they already paid for.

Investor pressure, especially from Soft

Bank, led them to scale prematurely and

make an aggressive pivot into unrelated

packaging and logistics industry, which

kept burning cash without any sight of

product market fit. End result, Zoom's

failure demonstrates how fundamental

physics can defeat all kinds of

sophisticated AI and robots. Now, let's

come back to Product Management 101 for

a second and realize that they created a

business around a problem that didn't

really exist. When doing research for

this video, I found this awesome, very

niche YouTube channel that talks about

how to start and operate pizza

businesses and they explain this very

problem as people who specialize in

pizza business. The thing is, pizza

delivery service as we know it works

fine. It's not ideal, but it's good

enough. And there isn't enough evidence

for a service gap that Zoom set out to

fix. Zoom's focus on technology over

basic product quality created an

unscalable business model. They went

into aggressive expansion without

validating their business model and

product in initial markets. I'm frankly

extremely surprised how this problem did

not come up in initial testing. But if

the initial testing never happened, then

that would answer my question.

And lastly, Quibby, the 1.75 billion

short form streaming disaster. Quibby

was a short form video streaming

platform in early 2020s. They raised

approximately $1.75 billion from

investors before launching in April 2020

and shutting down in December 2020. The

company was alive for 6 months. Quibby

is probably the fastest and the most

expensive content platform failure. And

the core theme here is that they failed

to understand how people consume mobile

content. This is definitely the fastest

failure given the funding of all the

companies that we covered today. Let's

understand why. Quibby raised funds

through multiple rounds and multiple

investors. They were backed by top tier

Hollywood studios. NBC, Sony, Warner

Brothers, Lionsgate, MGM, tech firms

like Alibaba, and investors like Goldman

Sachs and JP Morgan. The company spent

lavishly on content production. Budgets

reached up to $100,000 per minute for

original scripted shows. They produced a

wide range of original short form

content with a heavy focus on high value

production and big Hollywood talent.

Given that they were on the market for

six months, they've got quite a few

shows under their belt. Quibby invested

heavily in content, announcing plans to

spend over $1.1 billion on original

programming during its first year of

operation. prestigious filmmakers and

stars signed on, and many shows had

massive production budgets, very much

comparable to high-end cable TV and even

Netflix originals. Advertising spending

reached $63 million in just 6 months of

operations. But despite all of that,

Quibby's revenue from subscriptions and

ads was $7 million, and advertiser

payments started being deferred due to

low viewership. So, let's talk about

what went wrong. The biggest theme with

Quibby is that they really misunderstood

the market. Quibby's executives believe

that people wanted high quality short

form content for commuting, but they

launched in 2020 when commuting

disappeared. The service charged from $5

to 7 for content that users could get

free on Tik Tok or YouTube. Now, another

problem is the production costs. Despite

spending over $1 billion on content from

Hollywood talent, Quibby failed to

produce any breakout hits. Now, as

someone who got into content creation

fairly recently, I genuinely think that

there is no point in making large

investments into production until you

start feeling the content market fit, so

to say. And to be clear, I'm not

comparing myself to a media company with

massive Hollywood investors. They're in

a polar opposite scale. But

nevertheless, this is a content business

and I treat my own channel from a

perspective of a product manager. This

is my product. The point where we knew

that it makes sense for us to start

investing in better visuals was when we

hit several consecutive breakout videos.

When our channel had 500 subscribers

back in April, our video started

reaching 50 to 80,000 views, which was a

massive number for our channel size at

the time and given our subscriber count.

and we hit those views when videos were

filmed on my iPhone 13 with one $80

softbox. I deliberately kept everything

extremely low budget because I wanted to

see if my content could get breakout

through the value of the research and

the content that I produce. So,

investing billions in production when

you have absolutely no idea whether your

content is going to be watched makes no

sense. They also had a lot of technical

limitations. The app's inability to

share content on social media prevented

organic growth and word of mouth

marketing. And that is the bread and

butter of modern content platforms.

Users could not take screenshots. They

could not create clips. They could not

share anything with friends. End result,

within 6 months since inception, Quibby

announced it shut down. They admitted to

have failed to achieve sustainable

subscriber numbers for profitability.

The company returned some remaining

investor funds as goodwill and laid off

nearly 250 employees. Roku acquired

Quibby's content library in early 2021

for hund00 million and rebranded it as

Roku Originals.

Building a scalable AI business is very

hard. It's just as hard for larger corps

as it is for startups. The difference is

in the scale, but it's equally as hard.

Don't overestimate the model and

underestimate basic product quality.

Price the invisible labor. Account for

basic physics. Test before you invest

and benchmark against the boring product

management metrics that always work.

Ship the smallest thing that pays for

itself and let AI amplify an already

solid product. If the unit economics

don't beat the baseline, it's not a

pivot and it's not a business. You're

building a demo and demo doesn't pay the

bills. If this video was helpful, let us

know what you think in the comments.

Till next time. Bye.