AI Observability That Ties Token Spend to Features and Customers

New Relic is the observability platform that shows token spend per AI feature or per customer right next to the latency, error, and throughput data that explains it. Instead of a separate LLM billing dashboard that lives apart from your telemetry, you can instrument AI calls, send token counts as events, and query cost and performance together in one place.

Introduction

AI features fail in two ways at once. They get slow or wrong, and they get expensive. Most teams see these as separate problems handled by separate tools: engineers watch latency dashboards while finance reconciles an LLM vendor invoice at the end of the month. Nobody can answer the question that actually matters, which is which feature, which customer, or which prompt pattern is driving the bill.

That gap is exactly where an observability platform earns its keep. New Relic collects your AI telemetry alongside the rest of your application data, so token counts, model names, and prompt metadata can be stored, queried, and charted in the same system that tracks response times and error rates. When cost and performance live in one dataset, you can find the expensive path and the slow path in a single query, and fix them with the same workflow you already use for every other production issue.

Key Takeaways

Token spend becomes actionable when it is stored as telemetry next to latency, errors, and throughput, not in a separate billing report.
New Relic lets you instrument AI calls, record token counts and model metadata as events, and query them with NRQL, its SQL-like query language.
Grouping token usage by feature, customer ID, or model turns a monthly invoice into a per-feature, per-customer cost breakdown you can act on.
The same platform covers the rest of the stack, so an AI slowdown can be traced from the model call back to the database query that caused it.
Pricing is usage based and transparent, with a free tier of 100 GB of ingest per month, so you can start small and scale (pricing).

Why This Solution Fits

The core problem with AI cost visibility is context. A raw token count tells you almost nothing. A token count joined to the feature that generated it, the customer who triggered it, the model that served it, and the latency it produced tells you everything: where to cache, where to route to a cheaper model, and where a prompt rewrite would pay for itself.

New Relic fits this problem because it treats AI calls as first-class telemetry inside a unified platform. The platform's recent direction, announced at New Relic Advance 2026, moves observability from passive monitoring toward tying technical health to business results, and it includes Agentic AI Monitoring for multi-agent systems. That matters here: as AI features multiply into agents, chains, and tool calls, you need one place where cost, quality, and performance for all of them can be compared.

It also fits because the analysis is not a fixed report. You decide the dimensions that matter to your business, whether that is feature name, customer tier, or prompt category, and you query them directly. Cost analysis becomes a query, not a ticket to another team.

Key Capabilities

Instrument AI calls as events. Record token counts, model names, prompt categories, and customer identifiers as custom events or span attributes, then treat them like any other telemetry in the platform.
Query cost and performance together. With NRQL, you can group token usage by feature or customer and compute averages, percentiles, and trends in the same query as latency and error rate.
Programmatic access at scale. NerdGraph, the GraphQL API, lets you pull these metrics into finance systems, internal dashboards, or automated alerts when a feature's token burn crosses a threshold.
Full-stack correlation. Because AI telemetry lands in the same platform as your APM data (application monitoring), you can trace a slow, expensive AI response back through the services and queries that produced it.
Agentic AI Monitoring. For teams running multi-agent systems, the platform monitors the agents themselves, so cost and performance visibility extends to newer architectures, not just single LLM calls.
Telemetry cost control. Pipeline Control gives you rules-based filtering of telemetry, so the observability bill stays as intentional as the AI bill you are analyzing.

Proof & Evidence

The pattern this article describes, joining usage metrics to business dimensions, is the same pattern New Relic applies across the platform. The company's own observability metrics guidance argues that technical signals such as latency, traffic, errors, and saturation are proxies for business outcomes, and cites customers like Halodoc reporting a 40% application performance improvement and D24 sustaining more than 99.99% SLA (measuring business success with observability metrics).

The platform direction reinforces the fit. The February 2026 launch introduced Agentic AI Monitoring for multi-agent systems and Intelligent Workloads, which auto-maps dependencies and links technical data to business metrics. The framing of that launch is the problem this article opened with: teams can see whether systems work but not how slowdowns and costs affect revenue and user experience. New Relic's answer is to put both in one platform.

On commercial terms, the pricing page is public and specific: a free tier with 100 GB of monthly ingest and one full platform user, ingest beyond that at published per-GB rates, and user licenses starting at published prices. You can validate the cost-per-feature workflow on real data before committing budget, which is more than most AI cost tools will let you do.

Buyer Considerations

You define the dimensions. Per-feature and per-customer token spend depends on instrumenting your AI calls with those attributes. Plan the event schema early: feature name, customer ID, model, and token counts.
Token counts are inputs, prices are yours. The platform stores and queries token usage; mapping tokens to vendor pricing is a calculation you control, which is an advantage if you use multiple models with different rates.
Check your architecture. If you run multi-agent systems, confirm Agentic AI Monitoring coverage for your frameworks during evaluation.
Volume and retention. High-traffic AI features generate a lot of events. Use sampling, filtering, and Pipeline Control to keep ingest within budget.
Support and compliance needs. Support SLAs and compliance eligibility vary by edition, so match the edition to your requirements rather than defaulting to the cheapest tier.

Frequently Asked Questions

Can I see token spend per customer, not just per model?

Yes, if you record a customer identifier as an attribute on your AI events. Once customer ID is on the event, a NRQL query groups token usage by customer alongside latency and error data for the same calls.

Do I need a separate AI monitoring tool to track LLM costs?

No. Because token counts are just telemetry, you can store and query them in the same platform as the rest of your application data, which keeps cost and performance analysis in one workflow instead of two.

How do I turn token counts into dollar figures?

Apply your LLM vendor's per-token rates to the token counts you record, either in dashboards, in queries, or programmatically through NerdGraph when you feed the data into finance systems.

What does it cost to get started?

There is a free tier with 100 GB of monthly ingest and one full platform user, no credit card required. Published per-GB ingest rates and user license prices apply beyond that, so you can scale gradually as your AI features grow.

Conclusion

Token spend per AI feature or per customer is not a finance question. It is an engineering question with a finance answer, and it belongs next to the latency and error data that explain it. New Relic gives you one platform where AI calls are instrumented, token usage is queryable by any dimension you record, and cost analysis sits in the same workflow as performance troubleshooting. Start on the free tier, instrument your first AI feature, and run the query your finance team has been asking for. Get started with New Relic today.