Observed Signal · Jul 31, 2026 · Investigative Report · Source: t3n · Impact: 3/5 · Sentiment: Neutral

Amazon wastes $2.5M on faulty AI projects

Executive Signal Summary

A Financial Times report says Amazon incurred about $2.5 million in unexpected costs from several faulty internal AI projects. The largest overrun — roughly $1.8 million — came from an initiative that used an Anthropic model called Claude Sonnet; other overruns included $541,000 for finance-audit tooling and $134,000 for a delivery-speed improvement tool. Causes cited include programming errors, token-based billing, expensive models and AI agents driving extra requests. Amazon says the incidents affected only a handful of teams and downplays broader significance; the company plans to add automated guardrails and reduce dependence on Claude while exploring alternatives such as OpenAI. The report notes some overruns went undetected for weeks or months and that AWS cheaper options could have limited the costs.

Polaris7 AgentPolaris7 Strategic Assessment
High Confidence

Illustrates operational and cost-control risks when enterprises deploy LLMs at scale; relevant to cloud cost management, model choice and vendor dependence for a major platform (Amazon), but the monetary impact is small relative to Amazon's financials.

SIGNAL RADAR

Track Amazon Signals & Market Shifts in Real-Time

Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your Polaris7 workspace to monitor these entities.

Start Monitoring
Verified primary sourcesAutomated briefingsInstant watchlist setup

Key Takeaways & Evidence Grounding

  • Amazon incurred approximately $2.5 million in unexpected costs from internal AI project errors.
  • The largest single overrun (~$1.8 million) was linked to a project using Anthropics' AI model Claude Sonnet; the developed program was never deployed.
  • Additional overruns included $541,000 for AI-based finance-audit tooling and $134,000 for a delivery-speed improvement tool.
  • One cost overrun remained undetected for five months; other overruns were discovered after more than two weeks.
  • Amazon plans to implement automated guardrails for future AI projects and to reduce reliance on Claude while exploring alternatives such as OpenAI.

Connected Companies & Entities

“Amazon recently cut thousands of jobs and is simultaneously investing tens of billions of US dollars in expanding its AI infrastructure.”

“A report by the Financial Times now highlights another uncomfortable chapter for Amazon related to the company's AI orientation.”

“The largest of the unexpected expenses was caused by an AI project based on Anthropics' AI model Claude Sonnet.”

“Amazon recently cut thousands of jobs and is simultaneously investing tens of billions of US dollars in expanding its AI infrastructure.”

“You will find external content from TargetVideo GmbH that complements our editorial offering on t3n.de.”

“According to The Next Web, AWS Bedrock Batch Inference is available at half price along with other cost-saving Bedrock options.”

“Options being explored to reduce dependence on Claude include OpenAI, according to The Next Web.”

Ontology Mapping & Concepts

Large Language Models (LLM) & AIRetailer & Marketplace#token-based billing#AI agents#Claude Sonnet
Primary Source Grounding & Direct Attribution
Direct Origin Attribution
Primary Reporting: t3n•Published: Jul 31, 2026
Original Coverage Title: “Katastrophal teuer: Amazon verbrennt 2,5 Millionen Dollar mit fehlerhaften KI-Projekten”

Related Market Signals & Shifts

Recent verified developments and strategic activity across this market segment.

InfrastructureMar 10, 2026

Amazon's Deep Dive: Tackling Recent Outages and AI Concerns

Amazon held an internal "This Week in Stores Tech" (TWiST) deep-dive meeting to investigate a series of recent outages and high-severity incidents affecting its online store. Dave Treadwell, senior vice president of eCommerce Foundation, told employees the meeting would focus on multiple Sev 1 incidents that degraded site availability. Internal documents initially flagged "GenAI-assisted changes" and the use of GenAI tools as a factor in a trend of incidents, though Amazon later clarified a single incident was related to AI and said none involved AI-written code. A separate outage prevented some users from checking out or accessing account details for roughly six hours and was attributed to a software code deployment. Amazon said it will reinforce safeguards for generative-AI usage, introduce temporary safety practices requiring additional review of GenAI-assisted production changes, and invest in more durable deterministic and agentic safeguards.

Read assessment
FinancialsJul 30, 2026

One in Four AI Dollars Wasted

A Harness report finds roughly one in every four dollars spent on AI is wasted as enterprises scale AI adoption faster than governance and cost controls. The survey of 700 engineering leaders and practitioners shows more than half of organizations lack a dedicated owner for AI costs and only one in five can trace unexpected AI cost spikes within hours. AI expenses span infrastructure, foundation models, SaaS subscriptions and managed services, and productivity tools like AI copilots are emerging as major cost drivers. Major cloud providers (including Oracle and AWS) and industry groups (Linux Foundation) are rolling out features and initiatives for AI cost visibility, while vendors and finance teams push for centralized ownership and better forecasting.

Read assessment
Large Language Models & AIMay 17, 2026

Amazon Employees 'Tokenmaxxing' Wastes AI Resources

Amazon has set internal targets encouraging wider AI use — aiming for more than 80% of developers to use AI weekly — and recently began recording AI token consumption in internal leaderboards via an in‑house tool called Meshclaw. Reporting by t3n (citing the Financial Times) describes a trend called "tokenmaxxing," where employees deliberately inflate token usage to look productive. Amazon says token stats are not used for formal performance reviews, but employees report managers still check the data. The article places the practice in a broader context of corporate AI metrics (Meta plans to measure employee "AI‑impact" from 2026), lower effective code‑acceptance after AI drafts (Waydev CEO Alex Circei), and environmental and electricity‑price concerns tied to large AI workloads, which has prompted local resistance to new AI data centers.

Read assessment
Executive Briefing & Watchlist

Receive all future signals for Amazon & competitors in your briefing

Never miss material M&A deals, leadership changes, account wins, or product rollouts. Polaris7 monitors 20,000+ primary sources 24/7, delivering verified telemetry directly into your executive briefing.

Daily primary source scansCurated executive briefingsUnsubscribe anytime