Tokenomics - The Big Picture

Understanding token utilization at a company level.

This workshop so far has optimized tokens for one agent: finding the right model, having a workflow without waste, getting evals costs under control,… but that one agent doesn’t run in isolation. It runs inside a company where dozens of teams are also using AI; not just to run agents in production, but to build them. And to build all kinds of applications. Claude Code, Cursor, GitHub Copilot, Codex: they are writing code, generating tests, and shipping features today, billed per seat or by usage, on an invoice that most engineering organizations are not watching as closely as they should.

1. The Risk is as Big as the Opportunity #

Early in this workshop, we’ve seen a few horror tales of companies that learned the hard way that token spending can easily spiral out of control. The culture of ‘Tokenmaxxing’, short-lived as it was, established the idea that spending more tokens for the sake of spending more tokens was a good idea. Once the bill arrived, however, without any equally impressive increases in quality or speed, that notion was quickly reevaluated.

Per-token prices are expected to drop by 90% in the next few years. However, agentic workloads (especially) are expected to cause an increase in token consumption of 24x through the end of the decade. Cheaper tokens do not mean cheaper AI; instead, it means more of it, used in more places, by more people, faster than most budgeting processes could track. This pervasive approach will generate a myriad of new opportunities, and a new type of cost that needs to be tracked, controlled and scrutinized: token spending.

Engineering organizations already started to identify this risk: in more extreme cases, teams that burn an entire year’s AI budget in a few months, or that at first pushing engineers towards more token spending, only to retract and having to cap spend per engineer. We’ve also seen companies pulling back the majority of their coding-assistant licenses after a period of unmetered use. Unfortunately, these are not edge cases anymore, and they’re not isolated to high-tech companies. They’re becoming common enough that a “FinOps for tokens” discipline is starting to get its own name in the industry.

2. The Same Three Questions, at Enterprise Level #

This workshop has been asking three questions about our AI Shoping Assistant: 1. is the model right, 2. is the workflow optimized, 3. is the spend producing outcomes. Those questions also make sense at a high level, as we look into the Enterprise AI coding assistants spread across the company:

• Where is the spend going? Attributed by tool, by team, by individual; not just a lump sum on a vendor invoice.
• What spend is wasted? Idle seats, abandoned sessions, expensive models used for simple tasks; the coding-assistant equivalent of the workflow waste you found with spans and Agent Efficiency earlier in this workshop.
• What spend is working? Which teams are shipping more, faster, with fewer regressions, because of the assistant they’re paying for — not just how many tokens they burned.

A bill from the LLM provider answers none of these. It shows the total; it doesn’t show whether the total was worth it.

Right now, most companies have siloed control over these cost aspects: the agents that are running live, being monitored by an observability platform, while AI coding-assistant spend lives in a spreadsheet somewhere in finance or procurement, disconnected from any signal about whether that spend is producing good outcomes.

Tokenomics, powered by Splunk Agent Observability, aims to close that gap.

3. Introducing Tokenomics, Powered by Splunk Agent Observability #

Tokenomics is a new solution, and as such, it will see rapid evolution over time.

It works by connecting to the billing API of the service providers, such as Claude, Codex, Cursor, Copilot, etc, to retrieve spending data. In parallel, it can leverage HR or organizational data to correlate users and orgs with token spending, providing ways to filter and segment investigation by provider, department, team, user etc.

Trends are also calculated in a heatmap view, allowing us to understand the spending habits of different organizations, departments and teams, with the ability to further segment by provider. For instance, the Advanced Research department has increased their Codex spending by 20% this month, while lowering their Claude costs by 60% during the same period.

The vision and roadmap for this solution is very ambitious and exciting; we will share more information about it as new releases approach.

Next, we’ll see a quick walkthrough of the Tokenomics solution, provided by your presenter.

The bill tells you what was spent. Tokenomics tells you whether it was worth it — for one agent, or for every AI tool the company runs.

If your engineering organization adopted three different AI coding assistants tomorrow, what’s the first question you’d want a company-level Tokenomics dashboard to answer?

We’ve gone all the way from a single trace to a company-wide ledger. Time to close the loop.