The 80% refund rate is the story

Vaudit, a four-month-old startup that audits enterprise invoices from AI providers, said on 30 June that it had found roughly $1.7 million in disputed charges across $34 million of AI spend at about 60 companies, including Panasonic, HP and Honda. Anthropic, OpenAI, AWS, Google Cloud and Azure credited back around 80% of the contested amount. None of them acknowledged systemic error.

That last sentence is the interesting one. The headline number is small and the sample is small, but the refund behaviour is a structural signal about how the inference-billing stack actually works, and I think a verification category is now forming around it.

What was actually announced. Vaudit's Business Wire release says it has "identified approximately $1.7 million in disputed charges" through its TokenAudit product, which reconciles enterprise AI invoices against contracted pricing and usage logs.¹ The flagged issues cluster around Claude Code (Anthropic's agentic coding product, which runs long multi-step sessions at $3 per million input tokens and $15 per million output tokens on Sonnet), failed requests billed as successful, retry storms, and model-tier pricing discrepancies where a customer was charged for a more expensive model than the one that ran.¹ Vaudit has been operating since March. There is no disclosed funding, revenue, or pricing for TokenAudit itself.¹

The dispute rate. $1.7M out of $34M is roughly 5%. That is a small sample from a single auditor with a commercial incentive to find errors, and it should be read as such. But 5% of enterprise AI spend is not a trivial line item when the customers named are the size of HP and Honda, and the workload types most exposed, agentic coding, multi-step tool use, are the ones enterprises are scaling into hardest in 2026. If 5% is the steady-state error rate as agentic traffic grows, it becomes material fast. If it is only the ceiling of what one auditor can find with limited instrumentation, it is still a category signal.

The refund rate is the tell. Anthropic and OpenAI told reporters they dispute that overbilling is widespread.² They are also, per Vaudit's account, crediting back roughly four out of every five dollars contested.¹ These two positions are not obviously consistent. If the billing is accurate, the refund rate should be lower, or the refunds should come with technical explanations of why a specific charge was reversed. Neither has surfaced publicly.

There are two clean readings. One: the disputes are largely meritorious, the providers know it, and they would rather issue credits quietly than argue about instrumentation gaps in an agentic-workload billing stack they are still building. Two: the refunds are relationship management — enterprise-account goodwill, cheaper than litigation, and especially cheap ahead of an IPO cycle where a "frontier lab in billing dispute with Fortune 500 customer" headline is worth avoiding at almost any credit-memo price. Both readings support the same conclusion, which is that vendors have decided contesting these charges is not worth it. That is the kind of decision that creates a market for the people generating the disputes.

The frame this fits. Inference economics — the shift from training cost to per-token running cost as the binding commercial constraint — predicts that the economic relationship between enterprises and AI vendors starts to look like cloud infrastructure billing rather than software licensing. Cloud billing produced a durable third-party category: Apptio, CloudHealth, Cloudability, all built on the premise that hyperscaler invoices are too complex and too opaque for customers to verify without help, and all eventually acquired. The conditions that made that category work — vendor-controlled metering, unilateral pricing changes, consumption volatility, line items customers cannot independently reproduce — are all present in per-token AI billing, and arguably sharper.

Tokenisation is defined by the vendor. Model routing is opaque. Retry logic in agentic products like Claude Code compounds cost in ways the customer did not authorise per se, because they authorised the agent, not each retry. A failed request that still consumed input tokens is billable under most current terms. None of this is fraud. It is what happens when a metering system designed for API calls gets stretched over multi-step autonomous workloads that the pricing page was not written for.

Where the frame bends. Cloud billing audit worked because AWS, Azure and GCP invoices are structurally reconcilable — the SKUs are stable, the metering is documented, and disputes turn on customer misconfiguration more than vendor error. AI billing is less reconcilable today because model versions, token counts, and routing decisions change under the customer, and the vendor's own logs are the source of truth for what happened. That makes third-party audit harder to do rigorously, and easier to do aggressively. Vaudit's fee structure is undisclosed, but any percentage-of-recovery model creates an incentive to classify ambiguous charges as disputes. The 5% number should be read with that in mind.

It also matters that some of what Vaudit is flagging is arguably a product problem, not a billing problem. Claude Code retry storms are a UX failure, an agent looping on a task the customer would have killed if they had seen it, that surfaces as a billing complaint because the meter kept running. Enterprises will not fix that with an audit. They will fix it with contract terms.

What procurement does next. HP, Honda and Panasonic hiring an outside auditor is the leading edge. The lagging edge is the 2026 renewal cycle, where I would expect enterprise procurement teams to push for billing SLAs, audit-right clauses, shadow-metering provisions, and caps on agentic-session token consumption — the AI-billing analogue of the Enterprise Discount Program terms that cloud customers extracted from hyperscalers in exchange for volume commitment. Frontier labs approaching IPO have a strong reason to concede these terms rather than fight them.

The specific dollar figure Vaudit disclosed is not the thing to track. The thing to track is whether the refund rate stays near 80% as more auditors enter the category and sample sizes grow, and whether any provider changes its terms to publish reconcilable per-request billing detail. If refund rates stay high and terms do not change, the verification layer becomes a permanent tax on inference revenue. If terms change, the category never grows beyond a compliance line item. Either way, the interesting move already happened: five vendors decided arguing about the invoice was more expensive than paying it.

Footnotes

¶