Superbills & OON Claims: AI Agent vs Biller vs Manual

By Jude Lee · · Comparison

Office manager reviewing an itemized patient statement on a laptop at the front desk of a small practice

What’s actually hard about superbills

A superbill is an itemized receipt the patient submits to their plan for out-of-network reimbursement. Generating it is trivial. The work is everywhere else:

That last bucket — asynchronous, unstructured follow-up — is where most of the hidden hours go, and it’s the part templates never solved.

What an AI agent does here that a template doesn’t

An AI agent isn’t a chatbot that answers billing questions. It’s a system that carries out a multi-step task and takes actions: pull completed encounters for the month, assemble the statement, check it against a rules list, flag what it can’t verify, draft the patient email, log what it did. Connected to your practice-management or EHR data through something like MCP — the Model Context Protocol, an open standard for giving an assistant governed, scoped access to specific tools and records — it reads the same fields your biller reads instead of you pasting them in.

The practical difference from a template is variation. “This client had two 90837s and one no-show fee in March; the no-show fee doesn’t belong on a superbill” is a judgment a rigid template misses. A model can catch it if that rule is written into the skill — and will also miss it sometimes, which is exactly why a pilot tracks correction rate by failure type. An agent can also read a patient reply (“Aetna says the provider ID is invalid”), classify it, and route it to the right fix.

Option 1: the superbill feature already in your system

Many practice-management and EHR platforms used by therapists, chiropractors, and dentists can generate a superbill from a completed encounter — SimplePractice, TherapyNotes, and Jane App all document superbill output for out-of-network clients, and dental and specialty systems vary more. To find out what yours does, search your vendor’s help documentation for “superbill,” “CMS-1500,” “insurance reimbursement statement,” and “batch export.” It’s included, it’s inside your existing BAA, and it uses the identifiers already stored in your provider record.

If your service mix is narrow — one or two CPT codes, consistent session lengths, one location — this is almost certainly the right answer, and adding AI on top is overhead for no gain. Not every workflow needs an agent.

Where it falls short: batching across date ranges, reconciliation against payments actually received, and anything involving a patient’s reply. Built-in features generate; they don’t follow up.

Option 2: an outsourced billing service

A billing service or OON concierge will generate superbills and often file claims and chase reimbursement, converting a staff problem into a line item. Ask three concrete questions before signing:

Also ask specifically who answers the patient’s third email — services vary enormously in how much patient-facing communication they actually do.

Option 3: a custom agent over your own data

The build case is narrower than vendors suggest. It’s strongest when the superbill itself isn’t the problem and the surrounding exception work is. A custom setup typically means a small MCP server exposing a few read operations (encounters, payments, provider identifiers, contact preferences) and one or two write operations (log a note, queue a draft), plus a documented skill — a reusable, packaged instruction set telling the assistant how your practice builds a superbill, the same way every time. Same architectural questions as when you connect an assistant to an EHR via MCP: minimum-necessary scoping, audit logging, and what the agent may do unattended.

Know the failure shapes before you design the audit. In practice this agent breaks by: pulling stale identifiers from a cached provider record; inventing or misapplying a modifier that looks plausible; carrying non-billable line items (no-show fees, product sales) onto the statement; mishandling date-range boundaries around a deductible reset; and silently returning empty encounter sets after your PM vendor changes an API. Sample against each of those, not against “accuracy” in general.

Off-the-shelf superbill feature
Already under your BAA. Zero build cost. Deterministic output. Good when codes and locations are uniform. Weak on batching, reconciliation, and patient replies. Changes wait on a vendor roadmap.
Custom agent + MCP
Handles variation and exceptions. Can draft patient replies and triage rejection reasons. Requires real build and maintenance effort, a BAA with your model provider, and someone who owns it when your PM vendor changes an API.

Where a human has to stay in the loop

Code selection is clinical and compliance territory. An agent can flag that a documented session length doesn’t support the billed code; it should not change the code. Treat every AI-suggested coding edit as a draft for the rendering clinician or a certified coder to approve — the same permission logic as read-only versus approval-gated EHR access.

Keep a human on anything touching a Good Faith Estimate, too. Under the No Surprises Act, CMS requires providers to furnish good faith estimates to uninsured and self-pay patients; verify current requirements directly with CMS rather than trusting a model’s summary. We cover that side in patient cost estimate automation.

An agent that drafts is useful. An agent that silently changes a code is a liability.

Sizing the win without inventing numbers

Don’t accept anyone’s dollar headline, including ours. Build your own:

N × M ÷ 60
Monthly staff hours: superbills per month × minutes each
Worked example — use your own counts
hours × loaded rate
Recovered-time value
Worked example
R × avg reimbursement
Rejected superbills reissued correctly, if you track R
Worked example

Measure M honestly — include follow-up emails, not just generation. Subtract the cost of the option you’re considering (service fees, or build plus maintenance plus model usage). If the gap is thin, keep it manual. Recovered hours only count if they get reallocated to revenue or retention work; otherwise you’ve bought idle capacity.

Picking a tool, by the job in front of you

There’s no single best AI tool for medical professionals. The category matters more than the brand: clinical reference tools (OpenEvidence, Glass Health) answer clinical questions and don’t touch PM data; patient-facing voice and triage products target intake, not reimbursement; general assistants like ChatGPT and Claude are general-purpose language models, not FDA-cleared clinical decision support, and the FDA regulates certain decision-support software as a device. Administrative agents connected to your own systems are what superbill work actually needs. Free tiers are fine for drafting a no-PHI patient explainer — not for pasting encounter data without a BAA.

Automate only what you can verify

You’ll see a “30% rule for AI” cited in contradictory forms — cap automation at 30% of a process, expect 30% of output to need correction, limit AI to 30% of decisions. No standards body or regulator stands behind it. Our opinion: ignore the number, keep the principle. Automate only the share of a workflow you can independently check, and run a sampling audit after go-live. If you can’t verify the agent’s work, you can’t expand its scope.

What billing roles actually look like after this

The tasks that disappear are pure retyping. The work that survives and gets more valuable is exception handling, payer relationship work, coding judgment, and the patient conversations that require someone to explain, apologize, or negotiate — none of which an agent does well unsupervised. Our opinion: plan the role around that exception work rather than around headcount math. If you run a therapy practice, the same shift shows up across the rest of your stack, as we mapped in AI options for behavioral health practices.

A pilot you can run this month

  1. Count the real cost

    Log every superbill-related task for 30 days, including patient replies. You need M before you can evaluate anything.

  2. Exhaust the built-in feature first

    Confirm what your PM system already does in batch. Many office managers have never tried the bulk export.

  3. Write the rules down as a skill

    Document your superbill standard — fields, formats, what never goes on one, how to word the patient email. Useful whether you automate or not.

  4. Pilot on drafts only

    Produce drafts for one clinician’s panel for a month. Nothing sends without review. Track correction rate by the failure types listed above.

  5. Expand only where corrections are rare

    Grant unattended action for the narrow steps that proved reliable. Keep code-related suggestions permanently human-approved, and re-verify your HIPAA posture before widening data access.

Not sure where to start?

Get a free automation audit: we map your scheduling, intake, insurance, billing, and patient communication and show you what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides