AI in Community-Based Adult Support Services

Executive findings

This report concerns community-based adult support providers, not hospitals. The in-scope organizations deliver residential, supported-living, personal-support, day/community, employment, behavioral, and care-coordination services to adults with intellectual or developmental disabilities and related support needs. Medicaid HCBS and state systems are normally the principal public-program context; Medicare applies only when a person or service separately triggers it. [S-003, S-004]

The closest consistent official measure of provider-company scale is the Census Bureau industry for Residential Intellectual and Developmental Disability Facilities. In 2022 it contained 7,304 employer firms, 35,807 establishments, and 530,422 employees nationwide. This is a narrow proxy: it excludes nonemployer organizations and many supported-living, in-home, day, employment, and case-management providers classified elsewhere. [S-001, S-002]

AI adoption in this exact sector is not measured representatively. The best nearby survey found that 91% of 403 U.S. home-care agency leaders were currently using or planning to use AI, but the numerator combines current behavior with intent and the population is broader home care, not IDD residential firms. [S-008]

The most documented applications are administrative and assistive: scheduling and matching, shift briefing, draft notes, completeness checks, sensor alerts, communication access, training, policy retrieval, EVV reconciliation, incident chronology, and supervisor exception queues. Evidence of functionality is substantially stronger than evidence of outcomes. [S-006, S-007, S-009, S-010]

No independent end-to-end evaluation was located for a DSP Personal Assistant. Clinical ambient-scribe studies show that documentation time can fall in physician workflows, but they do not establish the safety, quality, or effectiveness of the project PA in adult residential support. [S-012, S-013]

Industry definition and payer boundary

The report uses person-respecting terms: people receiving services, participants, adults with intellectual or developmental disabilities, and adults with behavioral-health or mental-health support needs. “Mental challenges” is not used as a technical category because it can blur distinct needs and can be stigmatizing.

In-scope provider work includes group homes and other community residential settings, supported living, personal support, day and community participation, employment support, respite, behavioral support, service coordination, and related administrative operations. Hospitals, skilled nursing facilities, and acute inpatient psychiatric care are out of scope.

CMS describes HCBS as Medicaid services that support people in homes and communities rather than institutions. Personal care commonly supports activities of daily living and instrumental activities of daily living. Medicare may intersect for dual-eligible adults or covered clinical services, but it is not the default payer for ordinary DSP residential support. [S-003, S-004]

Provider-company scale: what can be counted

There is no single public national registry of every unique parent company delivering IDD HCBS. State licensing, Medicaid enrollment, waiver-provider directories, nonprofit structures, and multi-state corporate families do not use one consistent identifier.

The table therefore uses the Census Bureau SUSB “firm” count for NAICS 623210, Residential Intellectual and Developmental Disability Facilities, enterprise-size total, 2022. A firm is the closest official company measure. An establishment is a physical business location. State firm counts show firms operating in each state; they are not additive because one multi-state firm can appear in more than one state. [S-001, S-002]

50-state firm table — NAICS 623210, 2022

National unique-firm total: 7,304; establishments: 35,807; employment: 530,422. State firm counts are not additive. [S-001]

AI adoption: exact populations and denominators

Home-care leader survey: 91% of 403 U.S. respondents reported current AI use or plans to use AI in May 2026. This is executive-reported intent plus use, not current adoption, and the surveyed population is home-care agency leaders rather than IDD residential firms. [S-008]

Qualitative agency study: at least 6 of 8 agency participants worked at agencies using AI shift matching. The eight agency participants were part of a 22-person purposive interview study that also included 11 home-care workers and 3 advocates. This is evidence that the application exists—not a prevalence estimate. [S-007]

IDD residential firms: no credible source located establishes what percentage of the 7,304 national NAICS 623210 employer firms currently use AI. DSPs: no representative personal-use rate was located. Bounded agentic AI: no representative adoption rate was located.

Daily workflow application map

Technology classes

Conventional automation follows predefined rules or transactions. In this sector it includes schedules, reminders, EVV, eligibility and authorization checks, form validation, and record reconciliation. EVV remains conventional automation unless a separate model component is documented. [S-005]

Predictive AI estimates a score, class, risk, or likely event. Examples include worker-service matching, vacancy or capacity forecasts, and sensor-based change or fall alerts. These systems require scrutiny for biased inputs, false alerts, workforce surveillance, and response capacity. [S-006, S-007]

Generative assistants create or transform language and audio. Relevant uses include shift briefs, draft service notes, summaries, translation, policy answers, and incident chronology. Outputs remain proposals until an authorized person verifies sources, meaning, and completeness.

Bounded agentic AI can select and sequence approved tools or actions toward a goal. In the project PA, bounded means least-privilege tools, explicit stop points, source attribution, reversible actions, auditable state, and human authority before documentation, reporting, service, billing, safety, or employment consequences. No representative adoption or independent sector outcome was found. [S-015]

Measured benefits and limits

No independently measured benefit was located for an end-to-end DSP Personal Assistant, AI shift briefing, plan-alignment checker, incident-packet assistant, policy assistant, EVV reconciliation agent, or supervisor exception queue in the target sector.

Clinical documentation analog: a pre/post quality-improvement study of 45 U.S. physicians, 17,428 encounters, and 9,629 AI-assisted encounters associated ambient AI with 0.57 fewer minutes per note, 6.89 fewer documentation minutes per day, and 19.95 fewer total EHR minutes per day over three months. It did not study DSPs or IDD services. [S-012]

Time-motion analog: a prospective study of 9 clinicians and 169 consultations in Singapore found documentation time decreased from 5.3 to 4.5 minutes per consultation, while consultation and total cycle time did not significantly change. It demonstrates that a local documentation gain need not reduce the total workflow. [S-013]

Vendor home-monitoring and workforce cases report operational improvements, but missing sample sizes, comparator construction, adjustment, and adverse-event detail prevent acceptance as independently measured benefits. [S-016-S-018]

Direct Support Staff and the project Personal Assistant

Start of shift: confirm assignment and identity; retrieve the current approved person-centered plan; summarize changes since the last shift; show scheduled supports, risks, preferences, open incidents, referrals, authorizations, and travel context. Missing or conflicting information remains visible.

During support: maintain consent state; optionally assist with transcription, translation, anonymous speaker separation, communication access, observations, reminders, and safety escalation. The PA does not diagnose, determine capacity, identify a speaker biometrically, or infer consent from silence.

After support: prepare a source-grounded draft note, surface missing evidence, propose tasks and referrals, and assemble an attributed handoff. The DSP corrects and attests; configured supervisor or clinical review remains separate.

Shift close: unresolved work transfers to a named accountable person or authoritative system. Sending a message does not close a referral or incident. Source audio is retained only through transcript validation, then follows the approved secure-deletion lifecycle. These are specified project rules, not production-deployment evidence.

Workforce context

The 2023 Relias/ANCOR survey had 763 DSP respondents. Thirty-eight percent reported not feeling fairly compensated and 21% reported insufficient support as job dislikes; 54% had been with their current organization four years or less. These measures describe a self-selected workforce sample and are not AI impacts. [S-011]

The evidence makes workforce design consequential. Scheduling or performance systems can redistribute hours, visibility, and discipline. Worker notice, understandable reasons, correction, appeal, accessibility, and protection against hidden productivity scoring are material risks. [S-007, S-011]

Company, product, and website examples

CareConnect documents workforce scheduling and matching functions for home-care organizations. It establishes capability, not representative adoption or independent benefit. [S-016]

Sensi.AI and CarePredict document ambient monitoring and alerting in home or residential settings. Their public customer cases are directional because core samples and comparison methods are incomplete. [S-017, S-018]

Voiceitt documents speech-recognition support for people with nonstandard speech. This is relevant to accessible communication, but no eligible target-sector outcome estimate was located. [S-019]

Conference and product discussions identify case-management visibility, AI-assisted care notes, administrative automation, intelligent scheduling, compliance, and staff training as active categories. These examples document market activity, not effectiveness. [S-010, S-020-S-023]

Social-media and conference scan

The 2025 ANCOR conference program included predictive analytics for resource planning, intelligent scheduling, back-office AI for compliance and training, and discussion of keeping DSPs involved in organizational decisions. [S-010]

Public 2026 LinkedIn discussion among IDD and HCBS technology professionals emphasizes documentation burden, case visibility, scheduling, payroll, billing, compliance, and human judgment. Another recurring theme is continuity risk when providers cannot export person-centered records or change vendors. [S-020-S-023]

These are qualitative signals from vendors, professionals, and conference organizers. They are not sampled populations, and no post, reaction count, or repeated claim is used for adoption or measured benefits.

Barriers, risks, and governance

Data and integration: person-centered plans, notes, incidents, EVV, schedules, authorizations, training, billing, and state systems are fragmented. A fluent summary can conceal missing or stale inputs.

Rights and person-centered care: surveillance, false behavioral inference, inaccessible interfaces, poor translation, and automated scheduling can reduce autonomy or shift burdens onto people receiving services and DSPs.

Workforce: hidden matching, productivity scoring, or discipline based on model outputs can create unfair allocation and undermine trust. Human review alone is insufficient unless staff can understand, correct, and appeal consequential outputs.

Privacy and security: plans, audio, location, routines, behavioral observations, health data, and communications may be highly sensitive or PHI. Applicability depends on the entity and activity, but minimum-necessary access, retention, vendor contracts, and incident response remain essential. [S-014]

Governance: model, prompt, tool, plan, policy, and report-template versions must be traceable. NIST AI RMF provides a voluntary risk-management framework; it is not a safety certification. [S-015]

Evidence gaps and conclusion

No representative percentage of U.S. community-based IDD residential firms currently using AI.

No complete national provider-company registry covering residential, supported-living, in-home, day, employment, behavioral, and coordination services under one identifier.

No independent end-to-end DSP Personal Assistant evaluation covering briefing, encounter support, draft documentation, task handoff, and supervisor approval.

No representative bounded-agentic AI adoption rate and no stable market definition of agentic autonomy.

Sparse outcome evidence for adults with IDD, nonstandard speech, multilingual households, rural services, participant autonomy, continuity, equitable workload, appeals, adverse events, and staff retention.

The evidence supports a narrow conclusion: AI is entering administrative and assistive workflows around community-based support, but direct sector adoption and effectiveness remain poorly measured. Conventional automation is established; predictive and generative functions are present; bounded agentic workflows remain emerging and unvalidated at sector scale.

Method and evidence rules

Research cutoff: 2026-08-12. The update reviewed the project’s governing IDD/DSP specification and prior industry research, then independently verified public quantitative claims and refreshed the sector boundary, Census firm data, product examples, and social priorities.

Adoption percentages retain the sampled population and denominator. Measured benefits require a baseline or comparator, sample, geography, period, and evidence type. Social sources inform themes only. Vendor capabilities are separated from adoption and outcomes.

This research report contains no product proposal, pilot, implementation plan, ROI model, sales strategy, or call to build or buy. It is not legal, clinical, billing, or compliance advice.

Source register