The Question · ABA · Utilization review

Your ABA treatment plan is read by software first, a clinician second.

Most health insurers now use AI somewhere in utilization review. That changes who reads your treatment plan first. Here is what the software and the new CMS toolkit look for, the template trap that follows, and the plan structure that survives both readers.

By ASP-RCM Editorial, Healthcare RCM Practice · Updated

The question we asked on LinkedIn

If the first reader of your treatment plan is software, do you write for the machine or for the clinician who signs the denial?

84%

Health insurers using AI/ML

NAIC survey run by 16 states, published May 2025

7

New state AI review laws, 2026

Passed this year by Becker's count, effective June 2026 to January 2028

$10.1B

Medicaid + CHIP ABA spend, 2025

Up from $1.94B in 2021, per the CMS ABA toolkit

1+

Standardized outcome tool

CMS tells states to require at least one, and to refuse provider-made measures

How an ABA authorization request gets read now

Where a plan uses automated review, the path usually looks like this. The software never signs the denial, but it decides what the clinician looks at first.

  1. Plan submitted

    Assessment, treatment plan and hours request go in through the portal or fax.

  2. Software reads the fields

    It checks for diagnosis, instrument names, baseline scores, goals, hours and dates. Missing fields get flagged.

  3. Flags set the review

    Requests that match a pattern, such as maximum hours or a cloned plan, go to a reviewer with the flag attached.

  4. Clinician decides

    A licensed reviewer approves, cuts units or denies. New state laws require this human step.

The elements a reviewer looks for first

Taken from the treatment plan elements the CMS ABA toolkit (August 2026) asks states to require. If a field is missing, software flags it before a person reads a word.

Plan elementWhat a strong plan showsWhat gets flagged
Assessment and reassessmentResults from a named, standardized tool and the date of the next reassessmentNo instrument named, or no reassessment schedule
BaselinesScores taken at the initial assessment for every goalGoals with no starting point, so progress cannot be shown
GoalsMeasurable, functional goals for the child and the caregiver, with mastery criteriaGeneric goals copied across children
HoursDirect ABA hours per week, tied to severity and the goals, caregiver training shown separatelyMaximum hours by default, or hours with no link to severity
Duration and transitionAn expected treatment length and a step-down or discharge planOpen-ended plans with no end point
ProgressChange on the same instrument since the last period, written as numbersNarrative only, or a progress tool the provider made up

The hours question is now explicit

CMS calls 40 hours a week a risk for negative audit findings and suggests matching hours to DSM-5-TR severity level. Average Medicaid ABA hours per child rose 22% since 2021.

Severity level 1: suggested hours per week10 hrs
Severity level 2: suggested hours per week20 hrs
Severity level 3: suggested hours per week30 hrs
Average Medicaid hours per child per week, 202114.2 hrs
Average Medicaid hours per child per week, 202517.3 hrs

Source: CMS Medicaid and CHIP ABA toolkit, August 2026 (T-MSIS data; 2023 excluded by CMS for a data error). Suggested hours are a CMS best-practice example, not a federal cap.

The template trap

Writing only for the software pushes BCBAs toward templates. Identical documentation across children is on the CMS list of flags for improper billing.

Written for the machine only

  • Same goal bank and wording for every child
  • Maximum hours requested by default
  • Progress described in words, no numbers
  • Caregiver training folded into direct hours

Written for both readers

  • Structured fields complete, so nothing gets flagged as missing
  • The same standardized instrument at intake and every reassessment, so progress is a number
  • Hours justified against severity and goals
  • An individual clinical story a reviewer can sign off on

State laws on AI in utilization review

Seven states passed laws in 2026 that say AI cannot be the only basis for a denial, require clinician review, or require disclosure. Others were already in force.

  1. Oct 1, 2025

    Maryland HB 820 in effect.

  2. Jan 2026

    Texas SB 815 and Nebraska LB 77 in effect.

  3. Jun 11, 2026

    Washington SB 5395 in effect.

  4. Jul 1, 2026

    Iowa HF 2635 and Indiana HB 1271 in effect.

  5. Oct 1, 2026

    Alabama SB 63 in effect.

  6. Jan 1, 2027

    Colorado HB 1139, Georgia SB 444 and Utah SB 319 take effect.

  7. Jan 1, 2028

    Illinois law takes effect.

What these laws do and do not change

They require a human

A licensed clinician must make or review an adverse decision. Software can still triage.

They do not cover every plan

Self-insured employer plans under ERISA sit outside state insurance law. Check which plan type each family has.

They give you a question to ask

On any denial, ask which criteria were applied and who signed. Several of these laws make that answer your right.

Score one of your treatment plans

Pick a recent plan and tick what is true. It takes two minutes.

0 of 8Under 6? Your plans are giving software a reason to flag them.Get a free ABA denial audit
TAKEAWAY

Lead with structured data so the software finds every element, then write the clinical story for the human who signs. Same instrument at intake and every reassessment, hours tied to severity, and no copied paragraphs.

Questions operators ask next

Do payers use AI to review ABA treatment plans?

Many do somewhere in the process. A 2025 NAIC survey found 84% of health insurers use AI or machine learning, and most were using or exploring it for utilization management or prior authorization. ProPublica reported in December 2024 on internal Optum documents about reviewing ABA requests to approve fewer units.

Can a payer deny ABA hours based only on an algorithm?

In a growing number of states, no. Laws in Maryland, Texas, Nebraska, Washington, Iowa, Indiana and Alabama are in effect, with Colorado, Georgia, Utah and Illinois to follow. They generally require a licensed clinician to make or review the adverse decision. Self-insured ERISA plans are outside these state laws.

What hours does CMS consider reasonable for ABA?

There is no federal cap. The CMS toolkit says 40 hours a week is not best practice and gives an example of 10, 20 and 30 hours a week for DSM-5-TR severity levels 1, 2 and 3. It says higher intensity should be authorized only when the plan documents the need.

Which outcome tools should an ABA treatment plan use?

CMS tells states to require at least one standardized outcome instrument and not to accept measures the provider created. Use the same instrument at intake and at every reassessment so progress shows as a number.

Want a second read on your plans before the payer does?

Send us 90 days of ABA denials and authorizations. A senior ABA billing partner walks you through which plan elements are costing you units, in a free 30-minute call with a written gap report after.

Keep reading

Sources

  1. CMS, Medicaid and CHIP Applied Behavior Analysis Toolkit, August 2026
  2. NAIC, health insurer AI/ML survey report, May 2025
  3. ProPublica, UnitedHealthcare and autism therapy denials, December 2024

Checked November 4, 2026. State law dates are from public bill trackers and state legislatures; confirm current status and plan type before relying on them. General information, not legal advice.