FT Reports Amazon AI Projects Ran Over Budget
Reports published July 30 said several internal Amazon AI projects exceeded their budgets, led by a failed Claude Sonnet deployment that cost about $1.8 million, ran 860% over budget, and went undetected for five months. Amazon said the cases were isolated examples from teams learning to use the technology and that it was developing automated spending guardrails.
The Financial Times reported that Amazon employees had identified large cost overruns in several internal AI projects. Tom's Hardware and Cybernews, both citing the FT report, said the largest disclosed case was a failed Claude Sonnet deployment intended to match author information with product listings on Amazon's ecommerce platform. They reported that it cost about $1.8 million, was 860% over budget, and was not detected for five months.
What the reports say
The same accounts described two other projects with unexpected costs: about $541,000 for work on a financial-auditing tool and about $134,000 for an AI task connected to delivery-speed improvements. Cybernews reported that engineers discussed the incidents at an internal staff meeting and that Amazon was developing automated guardrails to reduce the risk of similar overruns.
Amazon pushed back on treating the cases as representative of its broader AI program. In a statement reported by the FT and reproduced by both outlets, the company said teams were experimenting and improving cost efficiency, and that isolated examples should not be portrayed as normal practice across Amazon.
What remains unverified
The available public reporting does not include Amazon's internal budgets, invoices, token logs, request volumes, model configuration, or project postmortems. The dollar figures, failure description, and detection periods therefore remain attributed to the FT's reporting rather than independently verifiable from public records. No retrieved exact-event Amazon announcement or other first-party public document confirms the cases.
The reports also do not establish how much of each overrun came from model tokens, automated retries, engineering errors, data processing, or other infrastructure. That distinction matters because the phrase "AI cost" can combine several different expenses.
Why it matters for data and ML teams
LDS interpretation: the reported five-month gap is a cost-observability warning, not proof that every agentic workload will overspend. Teams evaluating similar systems need workload-level attribution—project owner, model, usage, retries, budget, and business outcome—to distinguish productive experimentation from an unattended failure. In this case, the absence of those details limits any stronger conclusion about Claude, agent architectures, or Amazon's wider AI strategy.
Key Points
- 1The Financial Times reported that a failed Amazon Claude Sonnet project cost about $1.8 million, ran 860% over budget, and went undetected for five months.
- 2Amazon said the disclosed cases were isolated learning examples and that it was developing automated guardrails to control future spending.
- 3No retrieved public record exposes the underlying budgets, usage logs, or cost breakdown, so the figures remain report-attributed rather than independently verified.
Scoring Rationale
The reported overruns are a useful enterprise AI cost-governance case for data, ML platform, and FinOps teams. The impact is moderate because the evidence comes from internal cases reported by the Financial Times without public budgets, usage logs, or technical postmortems.
Sources
Public references used for this report.
Practice with real Retail & eCommerce data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Retail & eCommerce problems
