Amazon’s AI Projects Blew 860% Over Budget. Here’s Exactly Why.

Late July 2026, several senior Amazon developers notified colleagues that multiple internal AI projects had completely lost cost control. The worst offender: an internal tool built on Anthropic's Claude Sonnet model, designed to match author data against product listings, which spent roughly $1.8 million over five months, pushing its budget 860% over target. Management did not notice the overrun for five months because the tool was never released publicly. Two additional projects in logistics and financial auditing brought the total surprise bill to around $2.5 million.

What happened

Detail	Figure
Worst single project overspend	~$1.8M (approx. €1.57M)
Budget overrun percentage	860%
Time before management noticed	5 months
Total surprise spend across 3 projects	~$2.5M (approx. €2.18M)
AI model involved in worst case	Anthropic Claude Sonnet
AWS China outage caused by internal AI agent	13 hours (late 2025)

The tool at the centre of the story was an internal automation built on Claude Sonnet. Its job was to compare author records against product listings, a routine data-reconciliation task. Because it was never launched publicly, there was no external signal to alert anyone that costs were mounting. Management only found out five months later when a developer raised the alarm internally.

Two further projects, one in logistics and one in financial auditing, also overran badly. Combined with the Claude Sonnet tool, the three projects account for roughly $2.5 million in unplanned spending, according to the source report.

This is not Amazon’s first internal AI mishap. In late 2025, the company’s internal AI coding assistant Kiro caused a 13-hour outage in AWS China after the agent deleted its own working environment.

Why token-based billing turns small mistakes into large invoices

People familiar with the situation pointed to a billing model shift as the core problem. Many AI vendors, including Anthropic, have moved customers away from flat monthly subscriptions toward per-token pricing. A token is roughly equivalent to a word fragment; the more an AI processes or generates, the more tokens are consumed and the higher the bill.

When an AI agent runs in a loop, produces incorrect code, and then tries to correct itself, token consumption spikes rapidly. Tasks that once cost almost nothing under flat pricing can generate thousands of dollars in a single day under per-token billing. Without hard spending caps or automated alerts, those costs accumulate invisibly.

If your business is evaluating AI integration for internal workflows, the billing model you negotiate upfront is not a detail. It is the single biggest factor in whether a project stays affordable.

How KiroRank made the problem worse

Amazon ran an internal leaderboard called KiroRank, meant to encourage developers to use the company’s AI tools regularly. Instead, it created a perverse incentive: developers assigned unnecessary tasks to AI agents specifically to generate token activity and climb the rankings. This “token maximisation” behaviour drove costs higher without producing any business value.

Amazon has since abolished KiroRank. The lesson here applies to any organisation rolling out internal AI tools: gamified usage metrics without cost guardrails will almost certainly inflate your bill.

The pattern is similar to what we have seen in recent agentic AI rollouts at OpenAI, Anthropic, and Microsoft, where autonomous agents running without tight constraints create unpredictable cost and reliability risks.

Why it matters

Amazon is not a naive adopter. It built AWS, runs some of the world’s largest cloud infrastructure, and has direct relationships with leading AI labs. If Amazon’s own teams could rack up $2.5 million in surprise AI costs without detection for months, the same thing can happen in organisations with far fewer controls.

The specific risks are:

• Per-token billing with no spend caps. Agents that loop or fail silently can consume budget in hours.
• No production monitoring. The Claude Sonnet tool ran for five months unnoticed because it had no public-facing output to trigger an alert.
• Gamified incentives. Internal leaderboards tied to AI usage, not AI value, push employees to generate wasteful activity.
• Agent access to critical systems. Kiro’s 2025 deletion of its own environment shows that autonomous agents with write access to infrastructure are a reliability risk.

Our take

Amazon’s characterisation of these incidents as “isolated cases in a learning phase” is plausible but not reassuring. Five months without noticing an 860% budget overrun is a governance failure, not just a learning experience. The KiroRank story is particularly telling: it shows how quickly internal AI adoption programmes can be distorted by the wrong incentives.

For smaller businesses and agencies, the good news is that the guardrails needed here are not complicated. Spend caps, token-consumption alerts, and a clear sign-off process before an agent touches production data would have caught this early. The technology is maturing fast, but the operational discipline around it is lagging. That gap is where real money gets lost.

What to do about it

1. Set hard monthly spend caps on every AI API key before you run a single agent in production.
2. Configure automated alerts when token usage crosses 50% and 80% of your cap, not just at the limit.
3. Require human sign-off before any AI agent gets write access to databases, infrastructure, or financial systems.
4. Audit your internal AI incentive programmes. If you reward usage volume rather than business outcomes, you are paying for noise.
5. Run a post-mortem on any AI project monthly, even if no one has flagged a problem. Silence is not a sign that everything is fine.

If you want help designing cost-controlled AI workflows for your business, the AI integration services page outlines how we approach this, or get in touch directly.

Source: Bing News · Claude AI