Skip to main content
The Practice
Billing operationsNovember 10, 2026

What AI Can Draft on a Speech Therapy Claim, and What You Still Own

AI can draft eligibility summaries, CPT suggestions, claim edits, status text, and denial-code maps for speech therapy. This article is the verify-before-submit workflow: what the tool may draft, what the biller or clinician must still check, and where a wrong suggestion becomes a denied or noncompliant claim.

Callie Editorial 20 min read
The billing issue
Verify

Action required

Denial recovery queue

01 · Classify

Eligibility, coding, documentation

02 · Correct

Fix the root record

03 · Respond

Resubmit or appeal on time

Reason → owner → deadline → evidence → outcome

At a glance

What you’ll leave with

  • AI speech therapy billing is a drafting layer on eligibility, CPT and units, claim edits, status tracking, and denial codes. The clinician or biller who submits the claim owns those codes.
  • The tool may summarize a 271, suggest 92507 or a timed unit count, flag a missing GN, or map a CARC. You still verify each suggestion against the note, the current CPT set, and the payer rule in front of you.
  • A wrong suggestion that ships is a denied claim or a noncompliant one. Run the same five-step verify-before-submit process on every AI-touched claim, including after the January 1, 2027 SLP code change.

Search “AI speech therapy billing” and you usually land on a product page, not a guide. The product page shows eligibility checks, superbills, and a claims queue. The question behind the query is different: which of those steps can a model draft, which steps still need a person, and where does a fluent wrong suggestion become a denied claim or a noncompliant one? This article is that guide. It is a billing-workflow piece, not a buyer scorecard and not an audit article about AI notes. The unit of work is the claim. The person who submits it owns the codes.

The definition

What is AI speech therapy billing?

AI speech therapy billing is software that drafts parts of the revenue cycle around a speech-language visit: eligibility summaries, CPT and unit suggestions, pre-submission claim edits, status tracking, and denial-code mapping. The tool may propose a code, a modifier, or a next action. The clinician or biller who submits the claim owns those codes. A suggestion that ships without a check becomes the practice’s claim, and a wrong one becomes a denial or a noncompliant submission.

That is a narrower job than “AI for speech therapy.” An ambient scribe drafts a note. A billing tool drafts a claim. The note and the claim have to tell the same story, but they fail in different places. A fluent invented percentage in a SOAP note is a documentation problem. A fluent invented unit count on 92507 is a claim problem. Both become the practice’s when someone signs or submits. The review pass for AI-drafted notes covers the first. This article covers the second.

Most of the useful work is not generative glamour. Summarizing a 271 eligibility response, checking a claim against National Correct Coding Initiative edits, and translating a 276/277 status or an 835 remittance are mechanical steps with a thin review. Suggesting a CPT code from a note is generative, and the review is not thin: the submitter is attesting that the code describes the service. The AI features scorecard sorts those families for buyers. Here the question is operational: for each step, what may the tool draft, and what must you still verify?

Five billing steps, split into draft and verify

Billing stepWhat AI may draftWhat you still verify
EligibilityA plain-language summary of the 271 response: coverage, cost sharing, prior authorizationThe therapy-specific benefit, habilitative limits, and that a quote is not a payment guarantee
CPT and unitsA suggested code and unit count from the note, schedule, or a trained code setThe current CPT descriptor, timed versus untimed rules, and the minutes in the signed note
Claim scrubbingFlags for missing modifiers, NCCI pairs, medically unlikely units, and clearinghouse field editsWhether the flag is a real edit, and whether the note supports any modifier that bypasses it
Status trackingA 276/277 or clearinghouse status in plain EnglishThe actual status on the transaction, not a guessed narrative
Denial codingA map from Claim Adjustment Reason Codes and Remittance Advice Remark Codes to a next actionThe payer’s reason, deadline, and path on the remittance itself

The centerpiece

What does a verify-before-submit process look like for an AI-touched claim?

Run the same five checks on every claim a tool has drafted or flagged. The order matches the revenue cycle: coverage, codes, edits, attestation, then follow-through. Skip a step and the next one is guessing. The person who clicks submit is the owner, whether the draft came from a model, a template, or a clearinghouse scrubber.

  1. 01

    Read the 271, not only the summary

    Run the HIPAA 270/271 eligibility transaction. If a tool summarized it, open the underlying response and confirm active coverage, in-network status, therapy-specific benefits, visit limits, habilitative versus rehabilitative buckets, and whether prior authorization is required. Log the date, payer, and reference. A summary that says “covered” is not a payment guarantee; the eligibility verification workflow is still the source of the quote you can defend.

  2. 02

    Match CPT and units to the signed note and the current code set

    Confirm the suggested code’s current descriptor, whether it is timed or untimed, and that the signed note supports it. Through December 31, 2026, individual SLP treatment is untimed 92507, one unit per day. Group treatment is 92508, also untimed. Timed OT and PT minutes still follow the Medicare 8-minute rule when that is the payer’s method. From January 1, 2027, ASHA reports that 92507 is replaced by a timed family; 92508 stays untimed. If the tool proposes a deleted code, extra units on an untimed code, or minutes the note does not contain, reject the suggestion.

  3. 03

    Scrub modifiers and edits against CMS and the payer, not against the model’s confidence

    Check GN on SLP always-therapy lines, NCCI procedure-to-procedure pairs, medically unlikely unit counts, place of service, and rendering provider. A scrubber that flags a missing GN is doing mechanical work. A scrubber that proposes modifier 59 to bypass an NCCI pair is proposing a clinical assertion the note has to support. The modifiers guide is the scenario map. False positives stay on the desk; they do not go out “because the tool said so.”

  4. 04

    Name the person who attests, then submit

    Record who verified the draft: clinician, biller, or both. The claim is the practice’s certification that the codes describe the service. There is no AI co-signer. If the person who would attest cannot explain why this code, this unit count, and this modifier belong on this date of service, the claim is not ready.

  5. 05

    Track status and denials from the transaction, then feed the lesson upstream

    Read the 276/277 or clearinghouse status and, after adjudication, the 835 remittance. Use Claim Adjustment Reason Codes and Remittance Advice Remark Codes as the reason, not a model’s paraphrase. Classify, correct, appeal, or write off using the denial workflow. If the same AI suggestion caused the denial, change the prompt, the template, or the stop-ship rule so the next claim does not repeat it.

“The tool drafts. The submitter attests. A suggestion that ships is no longer a suggestion.”

Eligibility

Can AI check eligibility before a speech therapy visit?

An AI layer can summarize a standard electronic eligibility response (the HIPAA 270 inquiry and 271 response) into coverage status, cost sharing, and whether prior authorization is required. It cannot guarantee payment. Payers treat a benefits quote as information, not a promise, and therapy-specific facts such as habilitative limits often sit outside the generic 271. Treat the summary as a first pass. Confirm the therapy benefit, visit limits, and authorization rules against the plan before you book or bill.

HIPAA-covered health plans must support the 270/271 transaction. That is the right first check: active coverage, plan dates, deductible progress, sometimes a service-type snippet for speech-language pathology. What it often does not answer well is the set that actually denies pediatric SLP claims: habilitative versus rehabilitative visit limits, developmental-diagnosis exclusions, visit counters shared with OT or PT, and authorization triggers after a visit count. Those still take a benefits call and a written log. CMS is also clear that eligibility operating rules exist so providers can see coverage and cost sharing electronically. They do not convert that response into an adjudication.

CPT and units

Can AI assign CPT codes or units for speech therapy?

No. A tool may suggest a code or unit count from the note or a trained code set. CPT selection and unit math remain the submitter’s. Through December 31, 2026, individual SLP treatment is untimed 92507, billed once per day. On January 1, 2027, ASHA reports 92507 is replaced by timed codes; 92508 stays untimed. A model that still proposes 92507 after that date, or applies the 8-minute rule to an untimed code, proposes a claim the payer can deny.

ASHA’s timed-versus-untimed guidance is the sorting rule: a timed code names a time in its descriptor (“each 15 minutes,” “initial 30 minutes”); an untimed code does not, and it is generally reported once per date of service. Most SLP evaluation and treatment codes have been untimed. That is why a 45-minute 92507 visit is still one unit, and why an AI that “helps” by multiplying 45 by the 8-minute chart is not helping. The 8-minute rule is a Medicare Part B method for timed therapy codes, documented in CMS Pub 100-04, Chapter 5, Section 20.2. It does not mint extra units on an untimed code. When a session is group treatment of two or more, the code is 92508, not extra units of 92507. The group-therapy documentation article covers that split.

1 unit

untimed 92507, through 2026

Individual SLP treatment is billed once per day regardless of session length (ASHA timed/untimed guidance; CPT descriptor has no time unit).

Jan 1, 2027

92507 is replaced

ASHA: a family of timed base and add-on codes replaces 92507. 92508 remains the untimed group code. Other payers set their own adoption.

GN

on SLP always-therapy lines

Medicare returns always-therapy codes billed without the matching plan-of-care modifier (CMS therapy modifier edits; CARC 4).

Claim edits

Can AI scrub claims and reduce denials?

A scrubber can flag missing modifiers, NCCI procedure-to-procedure pairs, medically unlikely unit counts, and fields that fail a clearinghouse edit before the claim leaves. That can catch mechanical errors. It cannot promise fewer denials, and this article publishes no savings percentage. The flags still need a person who can tell a real edit from a false positive, and who can check the note against the claim. Medical-necessity denials live in the chart, not in the edit engine.

CMS’s National Correct Coding Initiative maintains procedure-to-procedure edits so two codes that should not both pay on the same date for the same patient do not both pay, unless a clinically appropriate modifier is allowed and reported. Medically Unlikely Edits cap the units of one code on one date. Always-therapy SLP codes billed to Medicare without GN are returned, not paid, using Claim Adjustment Reason Code 4. Those are the mechanical catches a scrubber is good at. They are also the catches a stale or generic model is bad at: it may miss this year’s NCCI pair, invent a 59 that the note cannot support, or treat a medical-necessity denial as a coding typo. The Federal Trade Commission has told businesses that AI performance claims need substantiation like any other advertising claim. “We reduce denials” is such a claim. Measure it on your own remittances, or do not repeat it.

Follow-through

How should you track claim status and denial codes with AI?

AI can translate a HIPAA 276/277 claim-status response, or a clearinghouse status, into plain language, and it can map Claim Adjustment Reason Codes and Remittance Advice Remark Codes on an 835 remittance. Use that translation to classify the next action. Do not let the model invent a reason the remittance did not state. The denial workflow still starts from the payer’s actual codes, deadline, and submission path.

HIPAA adopted the 276/277 as the claim-status inquiry and response, with operating rules so plans furnish that status electronically. After adjudication, the 835 remittance reports each adjustment with a group code (who is financially responsible), a CARC (why the dollars moved), and often a RARC (more specific explanation). X12 maintains those code lists. A tool that says “denied for medical necessity” when the remittance shows a missing modifier has sent you down the wrong path, and you can burn an appeal deadline correcting a claim that only needed GN. Keep the payer’s codes in the work item. Plain English is a layer on top, not a replacement. The denial workflow is how those codes turn into a correction, a resubmit, or an appeal.

Compliance

Is it compliant to let AI suggest codes on a therapy claim?

Suggesting is compliant only if a named person verifies the suggestion before the claim is certified and submitted. ASHA’s generative-AI guidance places evaluation of any technology under Principle II of the Code of Ethics, and states that AI cannot replace the SLP. HIPAA requires a business associate agreement before a vendor receives protected health information. The claim that goes out is the practice’s attestation. There is no AI asterisk on that signature.

ASHA also says any claim about an AI tool’s efficacy must be substantiated with evidence, which is the same posture the FTC takes on AI advertising claims. A vendor that receives member IDs, diagnoses, notes, or claim files on your behalf is a business associate; HHS is unambiguous that the written agreement has to exist before that information flows, trials included. The diligence checklist for tools that hear sessions is in the AI scribe HIPAA article. Billing tools that never record audio still receive PHI. They still need the agreement. What they do not get is the right to be the author of the claim. Medicare documentation rules still require that the record support the billed service, including treatment time that supports timed codes. An unverified suggestion that inflates units or names the wrong service is not an AI error on the remittance. It is the practice’s claim.

Worked application

Where a wrong suggestion becomes a denied claim

The failure modes are easier to see on a real-shaped visit than on a feature list. The case below is fictional. The rules it trips are not.

Fictional worked case

A 45-minute pediatric language session, three AI drafts

A fictional outpatient SLP sees Maya, age 6, for individual language treatment under a certified plan of care. The signed note documents 45 minutes of one-on-one treatment targeting two-word combinations, with cue data and a skilled assessment. Date of service: a Tuesday in November 2026. The practice bills Medicare Part B and a commercial plan that follows Medicare coding conventions for this family. An AI billing assistant reads the note and the schedule.

Draft 1: units on an untimed code

The tool sees “45 minutes” and proposes 92507 × 3 units, applying the 8-minute chart as if 92507 were a 15-minute timed code. That is a claim the payer can deny or recoup. 92507 is untimed in 2026: one unit for the day, however long the visit ran. The minutes still belong in the note as total treatment time. They do not mint extra units. Reject the suggestion. Bill 1 unit.

Draft 2: missing GN

The same draft omits the plan-of-care modifier. 92507 is an always-therapy code. On a Medicare professional claim it needs GN. CMS instructs contractors to return always-therapy lines billed without the matching modifier, with Claim Adjustment Reason Code 4. The claim does not adjudicate; it comes back as a rejection. The scrubber that flags the missing GN is doing useful work. The submitter still has to apply GN, not a stack of GN and GP on the same line.

Draft 3: the 2027 codebook lag

The same tool, run on a January 2027 note, still proposes 92507 because its code set was trained on 2026. ASHA states that 92507 is replaced on January 1, 2027 by timed base and add-on codes selected by clinical service and reportable time, and that other payers set their own adoption. If Medicare has deleted 92507 and the commercial plan has not, the two claims for the same visit need two different code lines. The person who submits each claim checks that payer’s current file. The model does not.

What should have shipped

November 2026: 92507, 1 unit, modifier GN, diagnosis and rendering provider matching the note, eligibility re-checked against the 271, authorization current. The signed note already supports the service. The AI drafts did not. Two of the three drafts would have been a rejected or noncompliant claim. The third would have been a denied or unprocessable claim on day one of the new code set. None of those outcomes is an AI problem on the remittance. Each is a verification skip.

Field checklist

08 items

Stop-ship checks before an AI-touched speech therapy claim goes out

  • Eligibility: the 271 (or the benefits call log) was read, not only the AI summary, and habilitative versus rehabilitative coverage was named.
  • Code: the CPT matches the current descriptor and the signed note, not last year’s codebook.
  • Units: untimed codes are 1 unit per day; timed codes use this payer’s minute method, with minutes in the note.
  • Group versus individual: 92508 for group of two or more, never extra units of individual treatment.
  • Modifier: GN is on SLP always-therapy Medicare lines, and no line carries more than one of GN, GO, or GP.
  • Edits: NCCI pairs and MUEs were checked, and any 59 or X modifier is supported by distinct time or site in the note.
  • Attestation: a named clinician or biller can explain the code, units, and modifiers out loud.
  • After submit: status comes from the 277 or clearinghouse, and denials are classified from the CARC and RARC on the remittance.

Quick answers

AI speech therapy billing FAQ

What is AI speech therapy billing?

It is software that drafts parts of the speech-language revenue cycle: eligibility summaries, CPT and unit suggestions, pre-submission claim edits, status tracking, and denial-code mapping. The tool may propose a code or a next action. The clinician or biller who submits the claim owns those codes. A suggestion that ships without verification is the practice’s claim, not the vendor’s.

Can AI assign CPT codes or units for speech therapy?

It can suggest them. It cannot assign them in a way that removes your review. Through December 31, 2026, individual SLP treatment is untimed 92507, one unit per day. Group treatment is untimed 92508. On January 1, 2027, ASHA reports that 92507 is replaced by timed codes selected by clinical service and reportable time; 92508 stays untimed. Unit math for timed OT and PT codes still follows the payer’s method, often the Medicare 8-minute rule. The submitter checks the current descriptor, the signed note, and the payer file before the claim goes out.

Can AI scrub claims and reduce denials?

A scrubber can flag missing GN modifiers, NCCI procedure-to-procedure pairs, medically unlikely units, and clearinghouse field errors before submission. That catches mechanical mistakes. It does not guarantee fewer denials, and no savings percentage belongs in a guide without a primary source. Medical-necessity denials live in the chart. False positives and invented modifiers still need a person who can read the note against the edit.

Is it compliant to let AI suggest codes on a therapy claim?

Yes, if a named person verifies the suggestion before the claim is certified and submitted. ASHA’s generative-AI guidance puts evaluation of any technology under Principle II of the Code of Ethics and states that AI cannot replace the SLP. HIPAA requires a business associate agreement before a vendor receives protected health information. The claim is the practice’s attestation. There is no AI co-signer.

Does AI replace a biller for speech therapy claims?

No. It can shorten mechanical steps: summarizing a 271, flagging a missing modifier, translating a 277 status, mapping a CARC. It cannot attest that the code describes the service, decide whether a 59 is supported, or classify a medical-necessity denial. Practices still need a person who owns the queue. Software handles the repeatable checks; people handle the judgment calls.

What happens if an AI tool suggests 92507 after January 1, 2027?

Treat it as a stale codebook, not as a code. ASHA states that 92507 is replaced on that date by a timed family, while other payers decide if and when they adopt the new codes. Confirm the payer’s current file. If the plan still accepts 92507, that is the plan’s rule, not the model’s. If the plan has deleted it, shipping 92507 is an unprocessable or denied claim. Timed minutes, once they apply, have to appear in the signed note.

Primary sources

Bibliography / 15
  1. 01Generative Artificial Intelligence (AI) for Clinicians in Audiology and Speech-Language PathologyAmerican Speech-Language-Hearing Association
  2. 02Understanding Timed and Untimed CPT CodesAmerican Speech-Language-Hearing Association
  3. 03Speech-Language Pathology Treatment Codes Replacing 92507American Speech-Language-Hearing Association
  4. 04Medicare Claims Processing Manual, Chapter 5, Section 20.2 (Pub 100-04)Centers for Medicare & Medicaid Services
  5. 05Medicare Benefit Policy Manual, Chapter 15, Section 220.3 (Pub 100-02)Centers for Medicare & Medicaid Services
  6. 06Therapy ServicesCenters for Medicare & Medicaid Services
  7. 07NCCI for MedicareCenters for Medicare & Medicaid Services
  8. 08Health Plan Eligibility and Benefits Transaction Basics (270/271)Centers for Medicare & Medicaid Services
  9. 09Operating Rules for Eligibility and Claims StatusCenters for Medicare & Medicaid Services
  10. 10Claim Status Request and Response (276/277)Centers for Medicare & Medicaid Services
  11. 11Health Care Payment and Remittance Advice (CARC and RARC)Centers for Medicare & Medicaid Services
  12. 12Complying with Outpatient Rehabilitation Therapy Documentation Requirements (MLN905365)Centers for Medicare & Medicaid Services
  13. 13Business Associates (HIPAA guidance)U.S. Department of Health and Human Services
  14. 14Keep Your AI Claims in Check (business guidance)Federal Trade Commission
  15. 15Habilitative/Habilitation Services glossaryHealthCare.gov

Written by Callie Editorial

Published November 10, 2026

Educational content, not legal, billing, or patient-specific clinical advice.