AI vendors keep changing how they charge — per-token today, cached and batch tiers tomorrow, priority pricing and outcome pricing after that. Most companies discover their AI bill the way they discover a water leak: on the invoice. The Token Ledger is our framework for keeping AI spend measured, tiered, capped, and vendor-proof — no matter how the meter changes.
Three failure modes show up in nearly every stack we look at:
1. Spend is invisible until the invoice. Token usage isn't wired to features, so nobody can say what a conversation, a document, or a booking actually costs. You can't manage a number you can't see per-unit.
2. Everything runs on the flagship model. The demo was built on the most capable (most expensive) model, and it stayed there — including the 80% of calls a model a tenth the price would handle identically.
3. The pricing ground keeps moving. Providers have already shifted from simple per-token rates to prompt-cache discounts, batch tiers, priority premiums, and subscription seats — and repricing between model generations. A budget built on one vendor's current price sheet is a budget with an expiration date nobody wrote down.
Meter every AI call and roll it up per model and per feature, so the question "what does one customer conversation cost us?" has a number. Cost-per-outcome is the metric; tokens are just the plumbing.
Route each workload to the least expensive model that passes your quality bar, verified by test cases — flagship models for the high-stakes 20%, workhorse models for the routine 80%. Re-run the routing decision every time a vendor ships or reprices a model.
Layer the caches: provider prompt-caching for repeated context, semantic caching for repeated questions, and plain memoization for deterministic calls. Customer-facing AI answers the same twenty questions all day — those should approach zero marginal cost.
Hard monthly caps per feature, alerts at 60/85%, and graceful degradation — when a cap trips, the system falls back to a cheaper model or a canned response instead of silently running up the bill or falling over. Fail open for the customer, closed for the wallet.
One abstraction layer between your product and any model API, quality test-suites that make switching a config change instead of a rewrite, and a quarterly re-price ritual: re-quote your real workload against the current market. When a vendor changes its pricing model, you renegotiate from the driver's seat — because leaving is cheap.
Like everything we sell, the Token Ledger was built on our own books.
Our founder's business, Todo Culebra, runs customer-facing AI every day — so it runs this framework every day: per-model cost monitoring with budget alerts, workloads tiered across three providers by quality-per-dollar, an engineered semantic-cache layer for repeat guest questions, and rate limits + fallbacks on every public AI endpoint. The framework isn't a slide — it's the reason a one-person business can afford to keep AI answering 24/7.
Already running our AI concierge or booking build? The Ledger drops into it natively. Strategy-level question of which AI bets deserve budget at all? That's Moat & Meter — the two frameworks are designed to run together: one decides where the money goes, the other makes sure it isn't leaking on the way.
Tell us what you're running and roughly what it's costing. Two weeks later you'll know exactly where every token goes.
Book at islatechpr.com