In April, Uber's CTO admitted the company's entire 2026 AI budget was gone, just four months into the year. Agentic coding tools had spread from a third of its 5,000-engineer organization to 84% in a single quarter, and the forecast collapsed under the consumption. He wasn't alone.

Microsoft reportedly canceled most of its internal agentic coding licenses over runaway token bills. Priceline's renewal reportedly came back four to five times the prior year.

One healthcare enterprise consumed a trillion tokens in six months, racking up over $6 million in unplanned costs before its finance team understood what was driving the number.

Here is the paradox that makes these stories interesting rather than merely expensive: per-token prices have fallen an estimated 98% since late 2022, while enterprise AI bills have risen roughly 320% over the same period. Cheaper units, exploding totals.

Gartner's analysis points at the multiplier: agentic workloads consume 5 to 30 times more tokens per task than a chatbot query. But the multiplier itself isn't uniform.

Two organizations running the identical model on comparable workloads can see order-of-magnitude different bills. The variable isn't the intelligence. It's what intelligence has to work against, which is why legacy system modernization has quietly become the highest-leverage line in the AI budget.

An agent operating on clean data and documented interfaces completes a task in a handful of efficient steps. The same agent pointed at a fifteen-year-old ERP spends most of its tokens fighting the system: re-reading tables that contradict each other, retrying calls against undocumented APIs, and reasoning through a schema nobody has touched since 2011.

One company is paying for intelligence. The other is paying interest. Which one a company becomes usually depends on whether anyone applied the kind of staged modernization strategies that make systems legible before the agents arrive.

Technical Debt Now Has a Market Price: Metered Debt

For thirty years, technical debt was the safest thing in business to ignore. Everyone agreed it was real; nobody could put it on an income statement. Deferral was free. That era just ended, and the industry hasn't yet named what happened. AI agents are metered. They are priced per token, for every read, every retry, and every reasoning step, and when an agent operates inside a messy system, the mess stops being invisible. It gets itemized, at frontier rates, on a monthly invoice. Technical debt has acquired a market price for the first time in the history of software. Call it metered debt.

Why AI Agents Cost More on Legacy Systems: Where the Tokens Actually Go

The mechanics explain why two companies on the identical model can see order-of-magnitude different bills. An agent's cost per task is roughly (context tokens + output tokens) × steps × retries, and legacy systems inflate every variable in that equation.

Context bloat

An agent querying a well-designed system pulls a 3,000-token schema-aware response. The same query against a legacy database with no data contracts returns raw dumps, ambiguous column names, and duplicate records: 40,000 tokens the model must carry and re-read at every subsequent step. At frontier pricing, that single design difference compounds across every step of every task, every day.

Retry loops

Undocumented APIs force exploratory behavior. An agent that cannot know whether an endpoint is idempotent must verify state after every write, which means extra calls, extra reasoning, and extra tokens. Engineering teams tracing production agent tasks have found the majority of total token spend going to error handling and re-verification against systems that cannot confirm their own state.

Reconciliation reasoning

When three systems hold three definitions of "customer," the agent doesn't fail. It burns reasoning tokens adjudicating. The most expensive computer in the stack ends up doing work a data contract would have done for free.

Tool sprawl

Agents interact with systems through tool definitions, and every tool's description and response schema ride along in the context window. A legacy estate wrapped in dozens of ad-hoc tools taxes every call, whether or not those tools get used.

Why AI Pilots Fail to Scale: Pilots Die Where the Debt Lives

This is why "cost control," the folder the industry filed the blowups under, misses the point. Independent audits of engineering teams running agents in production found re-sent context alone accounting for the majority of the bill, which is a system-design cost, not a model cost. The token bill is a direct function of how hard a system is to read.

Firms that specialize in modernizing legacy systems see the pattern hold without exception: agents deployed on modernized foundations reach production, while agents bolted onto illegible stacks stall in pilot as their costs balloon.

Gartner now forecasts that 40% of AI agent projects will be cancelled by 2027 on cost overruns alone, not technical failure but simple economics. Two years of debate treated the adoption-to-scale gap as a maturity problem, a talent problem, a model problem. The invoice says otherwise: pilots die where the debt lives.

How to Make Legacy Systems Agent-Ready: The Engineering That Pays Down the Meter

None of this argues for the two-year replatforming projects that made "modernization" a dirty word. The debt gets paid down incrementally, and the playbook is established: rehost, replatform, refactor, rearchitect, and replace, all sharing one principle. Never bet the business on a cutover. Applied to agent-readiness specifically, the sequence runs in four moves.

1. Build an API facade over the legacy core

A thin, documented layer over the legacy core gives agents clean, typed endpoints without touching the system of record. The typical route is the strangler fig pattern, where old and new run in parallel until traffic cuts over module by module. This alone routinely cuts context consumption by an order of magnitude, because the agent reads a contract instead of a dump.

2. Establish data contracts

Explicit schemas, single definitions for core entities, and versioned interfaces. This kills reconciliation reasoning, the silent tax, at the source.

3. Instrument event-level audit logging

Every automated action writes to an append-only trail a regulator or CFO can follow. This is the governance layer analysts now call "the new cybersecurity," and it is a property of the system, not the model. The stakes sharpened in May, when security researchers documented a fully autonomous intrusion in which an agent compromised an exposed server, harvested credentials, and walked internal directories in under an hour, with no human at the keyboard. Illegible systems have always leaked, but they leaked at human speed. Agents, both a company's own and an attacker's, now operate on them at machine speed, and a system without an audit trail cannot say what either one did. In an agentic environment, unauditable is uninsurable.

4. Keep human sign-off on irreversible actions

Human sign-off comes last, and it stays permanently, wherever actions are irreversible. The teams capturing real value are not the ones automating the most. They automate one bounded process, instrument it, prove it against evals, and then widen.

What Metered Debt Predicts: Three Shifts to Watch

If metered debt is real, three developments follow, each specific enough to be proven wrong.

• First: within eighteen months, token-per-transaction becomes a standard finance metric, owned by CFOs rather than CIOs, because it is the first number that expresses system quality in dollars.

• Second: "agent-ready" replaces "cloud-ready" in board decks, with a specific meaning of readable data, traceable actions, and defined human sign-off.

• Third, the one that will feel obvious in hindsight: technical-debt disclosure enters M&A due diligence.

When debt has a metered price, acquirers will ask for the meter reading. A company running agents at five times a competitor's token cost is carrying a liability that never appeared on any balance sheet. Now it does.

The Bottom Line: The Meter Is Already Running

The past two years of enterprise AI were a race to buy intelligence. The next two will be a reckoning with what that intelligence runs on. The meter is already installed, and it reads history, not ambition. Every company is about to find out what its shortcuts actually cost, either on its own schedule or on the invoice's.

Media Contact
Company Name: Code District
Email: Send Email
Country: United States
Website: https://codedistrict.com