Critical Propulsion
← Back to Insights
AI & Agents7 min read

Your Token Bill Isn't a Model Problem. It's a Delivery Problem.

Token bills aren't exploding because the technology got expensive. They're exploding because most teams skipped the operational discipline that keeps agentic workflows accountable.

50%
of GenAI projects will overrun their budgeted costs through 2028, due to poor architectural choices and lack of operational know-how

Gartner

5–30x
more tokens consumed per task by agentic AI workflows versus a standard GenAI chatbot

Gartner

90%
projected reduction in inference cost on a 1-trillion-parameter LLM by 2030 versus 2025

Gartner, March 2026

Visibility Helps. It Isn't the Fix.

Token costs are blowing up budgets across the enterprise. Bills surprising CFOs. CIOs throttling agents. A whole product category has spun up to catch the panic: cost gateways, model-switching dashboards, FinOps tools that promise visibility into the spend.

Prices per token are falling. Capability per token is rising. The companies getting burned aren't getting burned by the model. They're paying for the absence of operational discipline in how the model is used.

The third stat above is the punchline. Unit costs are crashing. Bills are still exploding. That gap is the discipline gap. Discipline isn't a product. It's how the work flows. It can't be bolted on after the bill arrives.

Pilots Lied. Production Told the Truth.

Most enterprises priced their AI rollouts off a chatbot pilot. One question in, one answer out, a few hundred tokens, the math felt benign.

Then the same teams shipped agentic workflows: multi-step agents that plan, retry, fetch documents, summarize, and write back. Each step spends tokens. Some spend twice because the agent re-loads context it could have cached. Others spend on skills the team forgot were loaded, or on recovering from a step the agent shouldn't have taken in the first place.

The pilot's per-task cost was honest about pilots. It said nothing useful about production. A workflow that cost two cents in pilot can cost fifty cents in production. Not because the model got more expensive. Because the agent does twenty things now where the pilot did one.

Four Levers. Not Four Models.

Teams running disciplined agentic delivery aren't running cheaper models. They're running the same models, on the same tasks, inside an operating model that bakes discipline in.

Observability First
Token usage logged at the call site. Costs attributed to the workflow, the skill, the agent, the user, and the customer outcome. Without per-agent, per-step attribution, every other lever is a guess. This isn't optional. It's the foundation everything else runs on.
Skill Bloat Audit
Every skill loaded into an agent's context costs tokens. Every invocation. Audit skill usage quarterly. Retire what isn't earning its slot. Split fat skills into thinner, composable ones the agent loads only when needed. Treat the skill manifest like a build dependency list.
Route the Work
Not every step needs the frontier model. Most don't. Classification, formatting, summarization, routing: cheap models handle these well. The expensive model gets called for the hard step: synthesis, novel reasoning, high-stakes generation. A well-routed workflow can cut spend in half without touching output quality.
Escalation on Cost, Not Just Uncertainty
Most teams escalate to a human when the agent gets uncertain. Disciplined teams also escalate when the agent gets expensive. A workflow burning tokens to recover from a bad plan doesn't need more tokens. It needs a consultant to step in, redirect, and let the agent finish on a cheaper path.

The Discipline Compounds

Each lever on its own is worth real money. Together, they change the math.

Without disciplineWith four levers
Per-task cost grows with usagePer-task cost flat or declining with usage
Bills surprise the CFO monthlyBills attributed to outcomes weekly
Skill manifest grows uncontrolledSkill manifest audited and pruned quarterly
Frontier model on every stepFrontier model only where it pays back
Humans escalate on uncertaintyHumans escalate on cost and risk

The teams that get this right aren't running fewer agents. They're running more agents, on harder problems, at a unit cost that lets the business say yes to the next workflow instead of pulling the plug on the current one.

That's the gap. Not model choice. Not vendor selection. Delivery discipline.

Who This Is For (and Who It Isn't)

This IS for You If:
  • Your monthly AI invoice is bigger than your forecast and nobody can explain why
  • You priced your rollout off a chatbot pilot and shipped agentic workflows
  • Your finance team wants per-workflow cost attribution and your platform can't provide it
  • You suspect skill bloat but don't have the instrumentation to prove it
This Is NOT for You If:
  • You're still running single-prompt use cases and your bill is rounding error
  • You think the answer is switching frontier providers
  • You're not ready to instrument before you optimize
  • You want a vendor to hand you a discount instead of an operating model
If you don't know what your most expensive workflow cost last week, you don't have a token problem. You have a delivery problem.
ShareLinkedInX

Your token bill is a signal. We help you read it.

We instrument agentic workflows so cost is a first-class signal alongside risk and quality, before the CFO asks.