Underpayments are silent. Denials are loud.

Most practices cannot answer a simple question without opening a PDF: what should this payer have paid us for this code, on this date, for this provider? If you cannot answer it quickly, you cannot tell an underpayment from a correct payment — and underpayments do not announce themselves.

A denial is loud. An underpayment is silent: the claim is accepted, money arrives, the line closes. Nothing in the workflow flags that the amount was 12% below contract.

Why contracted rates are harder than they look

The mental model of “a fee schedule” — one code, one rate — almost never survives contact with a real behavioral-health contract. The same CPT code can pay differently depending on:

• Modifier — telehealth, extended session, or a required combination that changes the payable amount.
• Credential tier — MD, NP, PA and licensed clinical staff frequently carry different percentages of the same base rate.
• Place of service — office, telehealth from home, facility.
• Rate basis — a flat dollar amount, or a percentage of a specific Medicare benchmark year, which quietly changes what you are owed when the benchmark updates.
• Effective dates and escalators — annual increases and renegotiation dates that shift the rate mid-year.

Multiply that across a dozen payers and you have thousands of legitimate rate combinations, described in prose across dozens of PDF pages, none of it machine-readable.

The terms that are not rates

Rates get the attention, but the operationally expensive terms are usually elsewhere in the contract:

• Timely filing windows — frequently shorter in behavioral health, and the single most avoidable category of lost revenue.
• Appeal levels and deadlines — how many attempts you get and how long each window runs.
• Prompt-pay terms — what the payer owes you for paying late, which almost nobody claims.
• Recoupment lookback — how far back a payer can claw money back, which determines your real exposure.
• Behavioral-health carve-outs — whether claims route to the payer or to a delegated entity, and under which terms.
• Telehealth terms — audio-only allowance, rate parity, permitted places of service.

These terms govern behavior every single day, and they live in a document nobody reads after signing.

Turning a contract into something a system can enforce

The useful end state is that contract terms become checks that run against real claims. Getting there has one hard requirement: every extracted term must be traceable to the sentence it came from.

This is where most automation attempts fail. A model reads a contract and asserts a rate. The rate is plausible. Nobody can verify it without re-reading the PDF, so either the team trusts it blindly or ignores the tool. Both outcomes are worse than no tool.

The alternative is provenance as a hard constraint: each extracted fact carries the verbatim quote and its location in the document, verified against the source by code rather than asserted by the model that read it. Facts that cannot be verified cannot be accepted. Ambiguous tables stop for a human instead of guessing. Rules arrive as drafts that your team approves before they affect anything.

What this changes

Once contract terms are structured and enforceable, three things become possible that were not before: underpayments become visible as a variance rather than invisible as a payment; timely-filing risk becomes a countdown rather than a discovery; and renegotiation becomes a conversation backed by your own data rather than the payer’s.

None of that requires trusting a model. It requires that every number you act on can be traced back to the page it came from.