Billing tells you what happened. Pricing intelligence helps you decide what should happen next when the world moves: provider invoices shift, usage mix tilts, or margin targets tighten.
Pricing Advisor is Chargly’s answer to that gap — built for teams who want suggestions, not surprises. This deep dive is for operators and engineers who need to trust the system enough to wire it into a real release process.
You will see how recommendations are formed, why determinism matters, and how version history turns pricing into something you can govern.
The problem space: moving costs, static menus
Credit menus look stable to users; underneath, provider economics move. If you only react when finance screams, you are either subsidizing usage or over-charging relative to competitors — often both on different events at the same time.
Advisor’s job is to surface those tensions early, with reasons attached.
- Raw token dashboards do not map to product events users understand.
- Spreadsheet models go stale the week after you ship them.
- Autonomous price bots destroy trust the first time they misfire.
We chose explainable rules over opaque optimization because AI billing is still a trust product.
What a recommendation contains
A useful recommendation is not “change 8 → 10.” It bundles context:
What you should see before you click Apply
- Which event (product-facing name) is affected
- Current credit cost vs suggested cost
- Signals: cost drift, margin vs target, usage volume or mix
- Plain-language rationale operators can repeat in Slack
Human in the loop
Advisor proposes. Your team disposes. Apply and reject are both first-class outcomes.
Deterministic and explainable
“Deterministic” means the same inputs yield the same suggestion — no randomness, no hidden model refresh changing outputs overnight. That makes recommendations testable and disputable in a healthy way: if you disagree with the premise, you fix the inputs or the policy — not argue with a black box.
Version history as the audit trail
When you apply, Chargly creates a new immutable pricing rule version (v1 → v2, and so on). History captures what changed and when. That is how you answer:
- “What were we charging in April for
image.generate?” - “Did we intentionally accept this margin compression?”
Rejections matter too: a recorded no is evidence of judgment, not drift.
MCP and agents (without removing accountability)
Some teams use MCP tools so internal agents can fetch recommendations or apply after human approval. The architecture still assumes someone owns the decision — agents accelerate paperwork, they should not silently reset pricing.
Where Advisor does not help yet
Advisor will not fix bad event naming, broken metering, or pack economics that do not match usage. It amplifies signal from a clean loop. If deductions are noisy, recommendations will be noisy — which is another reason to invest in billable events and ledger clarity first.
Pricing Advisor is how credit-first billing grows up: from static tables to governed iteration — with receipts.