> ## Content Index
> Fetch the complete content index at: https://nnutshell.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Chargebee wants a telecom-style standard for AI agent costs
- URL: https://nnutshell.com/chargebee-wants-a-telecom-style-standard-for-ai-agent-costs/
- Published: 2026-10-25T23:00:25.000Z
- Updated: 2026-10-25T23:00:25.000Z
- Description: Chargebee's Venet Sashadri argues AI companies can chart revenue but not margin, and that an open, telecom-inspired record standard matters even as model prices fall.
- Author: Admin
- Tags: Venet Sashadri, Chargebee, Snowflake, Twilio, Agentic AI, Enterprise SaaS, Fintech, Open Source, #rec-2026-09-28

*Recorded 28 September 2026 · 14 min*

Venet Sashadri, senior director of product management for monetization and PLG solutions at Chargebee, argues that AI companies cannot see their margins because the cost of an agent run is scattered across applications, harnesses and model routers. His proposed fix is a shared data standard, not a shared billing system. Chargebee has just launched one, the Agent Usage Detail Record (AUDR), modeled on the telecom call detail record. He spoke with host Sophia.

The tension is between a falling cost and a rising one. Model prices are dropping, which suggests the problem might solve itself. Sashadri says the harness and eval layers, where companies differentiate, are getting more expensive and harder to measure. The conversation is also a product launch: the spec is open, adapters are only partly shipped, and Chargebee is asking people to try to break it.

1. **Telecom's data-standard precedent** — [▶ 1:25](https://www.youtube.com/watch?v=a95PcDWm2%5F4&t=85s&ref=nnutshell.com)  
Chargebee's AI-era customers, including CodeRabbit, DeepL and Lambda Labs, increasingly ask what a customer's agent runs cost them this month and what margin remains. Sashadri says the data is not hidden, only distributed and in need of joining. He points to telecom 40 years ago: a call crossed two carriers and three switches, each knowing its own leg's cost, none the total. The industry solved it with a shared call detail record, not one billing system. AUDR copies this, joining the application (user), the harness (tool calls) and routers (models, tokens). The transfer of the analogy is his argument, not demonstrated.
2. **Revenue chart, no margin chart** — [▶ 4:12](https://www.youtube.com/watch?v=a95PcDWm2%5F4&t=252s&ref=nnutshell.com)  
Every AI company can now chart revenue but not margin, Sashadri says, because usage varies sharply month to month and day to day. Model choice ranges from expensive to cheap, and tasks differ in compute and tool-call intensity, so one agent run can cost 10x another for the same user. The host's own team sees this with Cursor, where one person uses $100 of credits and another $1,000\. Variable costs are not new, as Snowflake and Twilio show, but their magnitude has grown and their predictability has fallen. Granular instrumentation, he argues, is a precondition for margin-healthy business models.
3. **Harness costs are rising** — [▶ 6:50](https://www.youtube.com/watch?v=a95PcDWm2%5F4&t=410s&ref=nnutshell.com)  
Model cost, the figure most companies track, is falling relatively fast. The harness and eval layer is moving the other way, Sashadri says, and it is also where each company builds differentiated value atop frontier-lab models. That layer is compute-heavy and tool-call-heavy, and can involve third-party APIs. A month-end bill can be spread across customers, but that is a blunt average; understanding margin requires capturing cost at runtime, at the tool-call and agent-run level. The inference is that a company's most differentiating costs may be its least visible ones. Sashadri gave no figures on the size or growth of these costs.
4. **Why launch now** — [▶ 8:12](https://www.youtube.com/watch?v=a95PcDWm2%5F4&t=492s&ref=nnutshell.com)  
Chargebee is launching AUDR adapters for LiteLLM and Nvidia's NeMo relay, with more coming for OpenRouter, Mastra and Merge, plus one further name garbled in the transcript. Merge and Mastra are publicly backing the launch, and recent customer meetings kept raising the topic. Sashadri's stated goal is feedback: the most valuable contribution would be someone spending an hour on the spec trying to break it and finding where it fails to scale. The request acknowledges the standard is unproven. Its value depends on enough systems across the stack emitting compatible records.
5. **Cheaper inference won't suffice** — [▶ 9:47](https://www.youtube.com/watch?v=a95PcDWm2%5F4&t=587s&ref=nnutshell.com)  
Asked whether falling inference prices will make the cost problem disappear, Sashadri answers that it might, but might not. Frontier models keep arriving and always cost more than other models. Total cost is also not just model cost, because harness and eval spending sits on top. He welcomes smarter routers, saying "you don't want a Ferrari to run the grocery run," but argues routing gains still need measurement at the run level so costs can be aggregated into a profitable business model. The rebuttal is reasoned but unquantified.
6. **Finance agents on proprietary data** — [▶ 10:52](https://www.youtube.com/watch?v=a95PcDWm2%5F4&t=652s&ref=nnutshell.com)  
Chargebee's next-year roadmap centers on agents for the finance office. Receivables agents are coming online, along with a growth agent for growth managers. These will draw on Chargebee's billing, payments and growth data to produce insights and recommendations on growing revenue and margins. Sashadri gave no timelines or performance evidence, and did not say how the agents relate to AUDR. It is an editorial inference, not his claim, that standardized per-run cost records could eventually enrich the margin side of such recommendations.

The AI industry's margin problem is, in Venet Sashadri's account, a data-joining problem rather than a data-scarcity one. Every AI company, he says, can now chart revenue but cannot chart margin. The inputs exist. They are simply scattered: the application knows who the user is, the harness knows which tools an agent called, and model routers know which models ran and how many tokens were consumed. No single system sees a whole agent run. Answering what a customer's agents cost this month therefore means stitching together records that were never designed to be combined. Sashadri says Chargebee's AI-era customers, including CodeRabbit, DeepL and Lambda Labs, increasingly ask exactly this question.

His proposed solution borrows from telecommunications. Forty years ago, a phone call could cross two carriers and three switches. Each switch knew what the call cost on its own leg, yet no carrier could say what it cost overall. The industry did not fix this by adopting one billing system. It agreed on a data standard, the call detail record, which captured who the caller was and how much was consumed. Each carrier emitted the record, and Sashadri credits it with enabling roaming, interoperability across carriers and an itemized bill showing start time, end time and rate for every call. The Agent Usage Detail Record applies the same logic: an open standard recording who initiated an agent run and what it consumed across models, tools and infrastructure.

The urgency comes from how unlike seat-based software this has become. The host noted that on her own team, a Cursor subscription yields one user consuming $100 of credits and another $1,000, depending on workflow. Sashadri adds that usage swings month to month and day to day, that models range from expensive to cheap, and that tasks differ in how compute-intensive or tool-call-intensive they are. The cost of a single agent run can differ 10x for the same user from one day to the next. Variable costs are not new, he notes, pointing to Snowflake and Twilio. What has changed is their magnitude, which has grown, and their predictability, which has fallen. Account-level averages become a poor guide to profitability.

Where the cost sits matters as much as how large it is. Model cost, the figure most companies track, is falling quickly. The harness and eval layer is moving the other way, according to Sashadri, and it is also where each company builds differentiated value on top of what frontier labs provide. That layer is compute-heavy and tool-call-heavy, and may involve hitting third-party APIs. A month-end bill can be spread across the customer base, but that is a blunt average. What is needed is instrumentation at the tool-call and agent-run level, captured at runtime. An editorial inference follows: the costs most tied to a company's competitive advantage may be the ones least visible in its accounting, while the more visible model line shrinks.

The obvious objection is that cheaper inference will make the issue moot. Sashadri's answer is hedged: maybe, maybe not. Frontier models keep arriving and always cost more than the alternatives, and model price is only part of the bill. He welcomes better routers and smarter model choice, saying "you don't want a Ferrari to run the grocery run," but argues that routing improvements still need run-level measurement so costs can be aggregated into a profitable business model. The rebuttal rests largely on assertion. He offered no figures on how large harness and eval costs are relative to model costs, or how fast they are rising, so the central premise remains plausible but unquantified.

A standard is only as useful as the number of systems that emit it, and the work here is early. Chargebee is launching adapters for LiteLLM and Nvidia's NeMo relay, with others, including OpenRouter, Mastra and Merge, described as coming. Merge and Mastra are publicly backing the launch. Sashadri asked for something unusual: an hour of someone's time trying to break the spec and showing where it fails to scale. That request concedes the design is unproven. It also sits alongside Chargebee's commercial position. The company sells billing infrastructure and says it is building receivables and growth agents that draw on its billing, payments and growth data. Sashadri did not connect these efforts, but the overlap is worth watching.

The unresolved issue is whether measurement leads to better economics. Sashadri's own framing is that instrumentation is a prerequisite for sustainable, margin-healthy business models, a necessary condition rather than a sufficient one. Per-run records would show which customers and workflows are unprofitable, but they do not by themselves change pricing, routing or product design, and the discussion offered no customer results. Adoption is the other dependency. The standard works only if applications, harnesses and routers all emit compatible records, as telecom carriers did. Whether competing vendors in those layers will do so, rather than keeping their own formats, is a question the launch cannot yet answer.

**The open question.** Applications, harnesses and routers sit with different vendors, and each holds only part of the picture. Will they agree to emit a common record, and will run-level cost data actually change how AI companies price and route? Sashadri offered no adoption data or customer results. Only some adapters have launched, with the rest described as coming. The premise that harness and eval costs are rising enough to matter is asserted rather than quantified.

Watch the full video at [https://www.youtube.com/watch?v=a95PcDWm2\_4](https://www.youtube.com/watch?v=a95PcDWm2%5F4&ref=nnutshell.com).