10 Most Expensive AI Mistakes Companies Ever Made

The most expensive AI mistakes often come from shipping a working model into a process nobody governs.

There's no validation checkpoint, human oversight, bias testing, or cap on how much money a model can move before human intervention.

Below are 10 documented AI fails, each covering what happened, what it cost, what went wrong, and one thing you can check in your organization.

80%+ of AI Projects Fail. These Were the Most Expensive

AI failures can cost far more than the initial loss, and they are happening as companies scale adoption.

RAND found more than 80% of AI projects fail, while a 2025 MIT report found 95% of generative AI pilots failed to deliver measurable returns. Meanwhile, 42% of CIOs ranked AI and ML as their top technology priority.

The true cost can include operational disruption, legal liability, brand damage, and years of wasted investment, especially when a system scales before it has been properly validated.

The examples below are ranked by documented financial impact.

1. Google Bard – $100 Billion Lost After an Unchecked AI Demo

Financial impact: Roughly $100 billion off Alphabet's market cap on 8 February 2023, with shares down about 7.7%.

Google announced Bard on February 6, 2023, with a promotional post describing it as a launchpad for curiosity. In the demo ad, the bot claimed that the James Webb Space Telescope captured the first images of an exoplanet.

It didn't. The European Southern Observatory's Very Large Telescope did that in 2004, a fact NASA publishes on its own website.

Reuters flagged the mistake two days later, hours before Google executives took the stage in Paris to pitch Bard as the future of the company.

Alphabet shares plunged that day, wiping roughly $100 billion off the company's market value.

Language models get facts wrong constantly, but the problem was where the error occurred: in an ad Google chose and published as its best available example.

The error came as investors were already questioning Google's ability to compete with Microsoft, which had unveiled that it was folding ChatGPT into its Bing search engine.

That an easily verifiable error survived Google’s internal review and reached its own showcase material, which it chose as its best available example, did little to reassure them.

Social media users pointed out the obvious thing, which is that the claim could have been checked by Googling it. That is the actual failure.

AI errors like this are common, but Google had no review process for AI output inside advertising.

The Aftermath

• Alphabet's share price did recover within weeks, and the company went on to a far higher valuation.
• Google restricted Bard to a trusted-testers program before public release, and a spokesperson pointed to the importance of a rigorous testing process.
• Bard was folded into Gemini in early 2024, and the name was retired.

What to do differently: Any AI-generated line in customer-facing material should name a credible person who verified it and can be asked why they signed off.

2. Volkswagen CARIAD – $7.5 Billion Spent Building One OS for 12 Brands

Financial impact: CARIAD accumulated more than $7.5 billion in operating losses from 2022-2024, while software delays pushed back the Porsche Macan Electric and Audi Q6 e-tron by more than a year.

Volkswagen ultimately shifted toward external software partnerships and confirmed 1,600 job cuts at CARIAD in 2026.

Volkswagen founded CARIAD in 2020, initially as a Car.Software Organisation, to build one OS and centralized electronics architecture for all 12 group brands.

It includes AI-powered features such as automated driving and intelligent vehicle functions.

The problem was scope. Volkswagen tried to replace legacy systems and build custom AI capability at once.

Each is a five-year program with its own risk of failure. Sequenced, you learn before scaling, but run in parallel, and you risk three failures at once.

Software problems delayed the Porsche Macan and Audi Q6 e-tron by more than a year each, and CARIAD grew to around 6,000 employees before having to lay off 1,600 jobs in March 2026.

The Aftermath

• Volkswagen shifted away from an entirely in-house strategy. Its Western software architecture is being developed with Rivian, while CARIAD China is co-developing its China architecture with XPeng.
• CARIAD now focuses on automated driving with Bosch and Qualcomm, AI-based voice assistants, and cloud services.

What to do differently: Start with one narrow deployment and require measurable results within two quarters before expanding the scope.

Volkswagen's eventual fix was to turn to external partners after years of development and billions in losses.

3. IBM Watson Health – $4 Billion Spent on a Cancer AI Trained on Hypothetical Patients

Financial impact: Over $4 billion invested across development and acquisitions; Watson Health assets sold to Francisco Partners in 2022 for a reported figure around $1 billion; the MD Anderson partnership was cancelled after roughly $62 million

IBM marketed Watson as a system that would help cure cancer, backed by acquisitions including Truven Health Analytics, Merge Healthcare, Phytel and Explorys.

Then internal documents reported by STAT showed that Watson for Oncology had produced multiple examples of unsafe and incorrect treatment recommendations.

It was largely because of its training data. Watson for Oncology learned from a small set of hypothetical cases and expert opinions written by physicians at one institution, rather than a broad set of real patient records and outcomes.

Watson learned how doctors thought treatment should go, not what actually happened to patients. It was strong at parsing text, but cancer treatment is about judgment under uncertainty.

Watson for Oncology is still cited as one of the clearest bad AI examples in healthcare. IBM sold the promise of AI-powered cancer care before it had validated the product for real-world clinical use, where tolerance for a wrong answer is zero.

The Aftermath

• The assets were sold and renamed. Francisco Partners rebranded them as Merative. IBM kept the Watson name and stepped back from clinical medicine.
• IBM's AI business survived the failure. The company redirected toward enterprise AI tooling, which is where its actual strength always was.
• The lesson about training data stuck around. Healthcare AI buyers got a lot more skeptical of models built on synthetic or narrow datasets.

Check this yourself: Know what real-world outcomes a model learned from, and how they differ from your target population.

4. Zillow Offers – $407.9 Million Lost After an AI Model Misjudged Home Prices

Financial impact: Zillow wrote down $407.9 million in homes it had bought through Zillow Offers after its models overestimated their future selling prices.

Roughly 2,000 people, about a quarter of its workforce, lost their jobs as the business shut down.

Zillow Offers used the company's home-valuation modeling to buy houses directly, renovate them, and resell. In the third quarter of 2021, it bought 9,680 homes and sold 3,032.

Zillow's 8-K disclosed a $304.4 million write-down for the quarter and warned of another $240 to $265 million on homes already under contract, because it had been buying above its own revised estimates of what the houses would fetch.

The board voted to shut the business down on 2 November 2021, saying the unpredictability of home prices had far exceeded what it anticipated.

The Aftermath

• The model survived. The Zestimate still runs. Zillow's FY2025 10-K reports a median error rate of 1.8% for listed homes and 7.2% for off-market homes.
• Zillow went back to being a marketplace. Revenue reached roughly $2.2 billion in 2024, and the company reported a return to GAAP profitability in 2025.
• Competitors learned it too late. Opendoor absorbed months of heavy losses, continuing the model Zillow had exited. It’s a clear example of AI gone wrong at scale.

What to do differently: For every model that triggers a financial commitment, find the number that caps how much it can commit before a human reviews realized error against predicted error.

5. Arup Deepfake Heist – $25.6 Million Sent After a Video Call With a Fake CFO

Financial impact: $25.6 million, or HK$200 million, across 15 transfers to five Hong Kong bank accounts. The money was not recovered.

An employee in the Hong Kong office of engineering firm Arup received a message from the apparent CFO in the UK about a confidential transaction.

The employee was suspicious and asked for a video call to confirm.

On the call was the CFO. So were several recognizable senior colleagues. All of them were AI-generated, built from ordinary public video and audio of Arup executives.

Reassured, the employee made 15 transfers totaling $25.6 million, and only discovered the fraud after contacting head office.

Arup confirmed the incident to CNN, saying fake voices and images were used and that no internal systems were compromised.

The Aftermath

• Arup went public deliberately. CIO Rob Greig has spoken openly about the attack, describing it as technology-enhanced social engineering rather than a cyberattack. That
• It became the reference case for deepfake fraud. Video was treated as proof of identity. And now, it isn’t. The incident now appears in bank fraud training, insurer risk models, and the OECD-tracked AI incident record.

What to do differently: Verify through a channel the attacker can't touch. Video and voice are no longer enough proof of identity in the age of AI.

6. Earnest – $2.5 Million Bias Settlement for a Credit Variable That Stood In for Race

Financial impact: $2.5 million to the Commonwealth of Massachusetts, announced July 10, 2025, plus a mandated governance overhaul and ongoing reporting to the Attorney General's office.

Student loan lender Earnest used AI underwriting models that considered applicants’ college Cohort Default Rate when refinancing student loans.

However, cohort Default Rate can correlate with the racial makeup of a student body, making it a potential statistical proxy for race.

The Massachusetts Attorney General alleged the practice risked disparate harm to Black, Hispanic, and non-citizen applicants. Earnest denied the allegations and settled without admission.

The Aftermath

• The variables are banned. Earnest agreed to stop using Cohort Default Rate and the immigration-status knockout rule in its lending models.
• The remedy was governance. Earnest agreed to written AI policies, model inventories, risk assessments, fair lending tests, and an oversight team.
• The case is now cited as a template for state AI enforcement, with several attorneys general treating disparate impact as live regardless of federal posture.

What to do differently: Run correlations between every model input and protected class before launch, then test disparate impact on approvals, pricing, and terms. Do it before a regulator asks.

7. iTutor Group – $365,000 for Age-Filtered AI Hiring Discrimination

Financial impact: $365,000 to settle the EEOC's first AI hiring discrimination lawsuit in 2023, plus mandated policies, training, monitoring, and reporting under a five-year injunction.

According to the EEOC, iTutorGroup programmed its tutor application software to automatically reject women aged 55 or older and men aged 60 or older.

More than 200 qualified US-based applicants were screened out.

An applicant, Wendy Pincus, submitted two identical applications except for date of birth. The one with the younger birthdate got an interview. She filed a complaint. The company denied wrongdoing and settled quickly.

The Aftermath

• The company had to change its controls. iTutorGroup agreed to anti-discrimination policies, staff training, monitoring, reporting, and reconsideration of rejected applications.
• The injunction lasts five years. The settlement put the company under federal monitoring rather than simply requiring a one-time payment.

Check this yourself: Export your screening rules and check whether any field lets someone set a rule like this unblocked, then sample rejections weekly.

8. Deloitte Australia – A Partial Refund on a AU$440,000 Report Citing Sources That Did Not Exist

Financial impact: About AU$440,000 paid by the Australian government, with Deloitte later repaying the final AU$97,587 instalment after errors were found in the report.

Deloitte Australia delivered a report to a federal government department citing academic work that did not exist and quoting a Federal Court judgment that was never made.

The errors were consistent with unverified generative AI output. After the problems surfaced, Deloitte agreed to refund part of the contract.

This is damaging to Deloitte’s reputation as a firm that sells assurance. Clients pay for the firm's name on a conclusion, with the expectation that qualified people checked the work.

A report with invented citations undermines the product being sold.

The Aftermath

• The report was corrected and reissued with amended references and disclosure of the AI tooling involved.
• The case exposed a basic control gap. AI can draft the work, but someone must verify every factual source before it ships.

What to do differently: Make source-checking mandatory before anything AI-assisted ships and disclose AI use to the client upfront.

9. UnitedHealth & Humana – An AI System Got 90% of Appealed Care Denials Wrong

Financial impact: Class actions, a Senate investigation, securities and derivative suits, and reputational damage in an industry where trust is a licensing condition.

UnitedHealth and Humana used nH Predict, an algorithm from naviHealth, which UnitedHealth's Optum acquired in 2020. The tool estimates how many days of post-acute care a Medicare Advantage patient should need.

Lawsuit alleged the system was optimized to maximize cost savings rather than medical accuracy

In Estate of Lokken v. UnitedHealth Group, plaintiffs alleged the tool overrode physician judgment and cut off medically necessary care, and that about 90% of appealed denials were overturned.

Employees were reportedly told to keep patient stays within 1% of the algorithm's prediction.

A Senate Permanent Subcommittee on Investigations report in October 2024 found UnitedHealth's post-acute denial rate rose from 10.9% in 2020 to 22.7% in 2022.

The Aftermath

• CMS issued guidance in February 2024 clarifying that algorithms cannot be the sole basis for Medicare Advantage coverage decisions.
• Congress investigated insurers' use of technology in post-acute coverage decisions, while litigation over nH Predict continued.
• In 2026, the HHS Office of Inspector General found that 95% of appealed skilled nursing facility denials were overturned, including 97% of denials issued by naviHealth. OIG said the results raised concerns about initial reviews and recommended stronger oversight of denial patterns.

What to do differently: Pair every denial with a real justification and give clinicians a real-time override instead of relying on appeals, which move too slowly to help.

10. Taco Bell – 500 Drive-Throughs Rolled Back After Voice AI Failed on Real Orders

Financial impact: Deployment strategy reversed from broad rollout to selective use, plus brand damage

From 2023, Taco Bell put voice AI into more than 500 US drive-throughs to cut order errors and speed up service.

Instead, clips went viral of the system looping the same question, repeatedly upselling drinks, and being crashed by an order for 18,000 cups of water, placed deliberately to bypass the AI and reach a human being.

The rollback turned Taco Bell’s rollout into one of the most viral AI fails of 2023.

Taco Bell’s Chief Digital and Technology Officer Dane Mathews told the Wall Street Journal in August 2025 that the company was “learning a lot” and that the system sometimes let him down.

The Aftermath

• Taco Bell narrowed the rollout. Restaurants got more control over where and when to use the system, with humans stepping in when needed.
• Selective deployment became the safer model. Better to discover that at 50 locations than 500.

What to do differently: Before scaling any customer-facing AI, spend have people try and break it, then measure total staff minutes per transaction with the system versus without to know if the deployment helps.

All 10 Cases Side by Side: What It Cost, What Broke, What to Fix

Here’s a quick comparison of the ten cases, their costs, and the failures behind them.

Why the Most Expensive AI Mistakes Happen: 7 Recurring Patterns

Ten different companies all experienced roughly the same seven AI fails. Each pattern below is paired with the controls that could have prevented it.

• The big bang trap: Transforming everything at once
• The speed trap: Launching before validation is complete
• The scale trap: Scaling AI before proof
• The wrong metric trap: Optimizing the wrong outcome
• The data trap: Training AI on the wrong data
• The missing review trap: Skipping the human review step
• The accountability trap: Giving AI decisions no clear owner

1. The Big Bang Trap: Transforming Everything at Once

This is the most expensive pattern here. Big programs hide their failures, and by the time one surfaces, the budget is already committed. It’s a textbook case of corporate AI implementation failure driven by scope creep.

CARIAD is the clearest case. Volkswagen tried to replace legacy systems, build custom AI capability and deliver one architecture for 12 brands at the same time.

Volkswagen's eventual answer was to split the work between Rivian and XPeng. That option existed in 2020, for a fraction of the price.

How to avoid it:

• Pick one narrow use case with a checkable answer, and require a measurable business result within two quarters before adding scope.
• Sequence programs that depend on each other instead of running them in parallel, so each one can fail cheaply and separately.
• Budget for upfront costs. AI often costs money before it produces returns, and unrealistic payback forecasts create pressure to cut validation.

2. The Speed Trap: Launching Before Validation Is Complete

Competitive pressure sets the launch date before validation is finished, so validation gets cut because it produces no visible progress.

Google published the Bard ad one day after Microsoft announced ChatGPT was coming to Bing. The error inside it took seconds to check.

Zillow more than doubled its home purchase rate every quarter while the housing market turned, turning a pricing error that might have been survivable at 2,000 homes into a catastrophe at 9,680.

In both cases, the company could have caught the problem with enough allotted time.

How to avoid it:

• Make validation a named deliverable with an owner and deadline, so cutting it requires an explicit decision.
• Treat every AI-generated claim in customer-facing material as unverified until a named person confirms it.
• Tie growth in deployment volume to measured accuracy. A calendar date or a competitor's announcement is not a reason to scale.

3. The Scale Trap: Scaling AI Before Proof

AI lets an organization make the same mistake thousands of times before anyone notices. A wrong model 2% of the time is harmless across ten decisions but dangerous across ten thousand.

iTutorGroup's age filter rejected more than 200 qualified applicants. The nH Predict lawsuits involve denials issued to thousands of patients.

How to avoid it:

• Set a ceiling on how many decisions or dollars a model can commit before human review, and tighten it automatically when errors rise.
• Sample decisions nobody looks at. Reading 20 automated rejections a week is one of the cheapest controls available and would have caught iTutorGroup in the first month.
• Roll out in stages with a real stopping point. Finding a problem at 50 locations costs far less than finding it at 500.

4. The Wrong Metric Trap: Optimizing the Wrong Outcome

These AI failures above optimized the number that was easy to measure instead of the one that mattered. Then, the model did exactly what it was told, at scale.

The nH Predict lawsuits allege the algorithm was tuned to cut authorized care rather than predict what patients needed.

A model built that way will cut authorized care. It can't tell you whether the care was necessary, because nobody asked it to know.

Zillow governed how many houses it bought by growth targets, when it should have used forecast error.

How to avoid it

• Write down what the model optimizes for. Then, write down what that produces at scale if it's slightly wrong. If the answers conflict, the metric is wrong.
• Measure the outcome you care about, including the human cleanup. For a drive-through, that means completed orders and total staff minutes per order.
• Never make cost reduction the sole objective for a system making decisions about people.

5. The Data Trap: Training AI on the Wrong Data

A model learns what's in the data, not what you intended. Three of these ten cases failed here, in three different ways:

• Watson for Oncology learned from a small set of hypothetical cases. That teaches a model how a few experts thought treatment should go, not what happened to real patients.
• Earnest's models used Cohort Default Rate. It looks like an ordinary credit signal, but it correlates with the racial makeup of a student body, which made it a statistical stand-in for race. Race was never a variable.
• Zillow's model was trained on rising prices. It stayed accurate right up to the moment the market turned.

How to avoid it:

• Before approving a model, ask what real-world outcomes it learned from and how that population differs from the one you're about to use it on.
• Correlate every input against race, age, gender, disability and national origin. School attended, ZIP code, name, commute distance, employment gaps and device type all carry demographic information.
• Test disparate impact on approvals, pricing, rejections and terms before launch, and keep the results.

6. The Missing Review Trap: Skipping the Human Review Step

In several of these cases, the entire failure was a missing human step.

Nobody Googled the claim in Google's own Bard ad. Nobody at Deloitte Australia opened the sources in an AI-assisted government report, so invented academic citations and a made-up court quotation went to a federal department.

How to avoid it:

• Make source verification mandatory for anything AI-assisted, with a named person responsible. No citation ships unless a human has read it.
• Add a mandatory delay above a set transfer threshold. Deepfake fraud runs on urgency, and a one-day hold costs almost nothing.

7. The Accountability Trap: Giving AI Decisions No Clear Owner

This pattern sits underneath the other six. In nearly every case, no single person owned the question of whether the system should keep running.

The best evidence is what regulators demanded after the incidents.

• Massachusetts required Earnest to produce written AI policies, model inventories, risk assessments, fair lending tests, and a standing oversight team.
• The EEOC went further with iTutorGroup and imposed five years of federal monitoring on top of the payment.

Regulators are writing these controls into settlements because their absence is what caused the cases.

How to avoid it:

• Name one person who owns the decision to shut the system off. They need to see model performance, business results, and legal exposure at once, and to act without calling a committee.
• Write down the specific result that triggers a halt before launch. Afterward, every number becomes negotiable.
• Track your reversal rate. If humans overturn your automated decisions, that's your error rate, and nobody will surface it unless it belongs to someone.

AI Gone Wrong: Final Words

In all ten cases, the technology worked, but the process around it failed. AI risk comes down to three questions: what can the model do when it's wrong, how quickly will anyone find out, and who can stop it?

Volkswagen employed capable engineers, Deloitte employed qualified consultants, and Arup's employee followed training exactly.

In each case, the missing piece was a governance decision that cost almost nothing to make early and a great deal to skip.

The companies that avoid these failures set clear limits before they deploy AI and enforce them when something goes wrong.

Our team ranks agencies worldwide to help you find a qualified partner. Visit our Agency Directory for the Top Business Consulting Firms as well as:

1. Top Startup Consulting Firms
2. Top Business Operations Consulting Firms
3. Top Small Business Consulting Firms
4. Top Consulting Firms in Columbus

Our design experts also recognize the most innovative design projects across the globe. Visit our Awards section for the best & latest.