The day the paperwork disappeared into the work.
No screen to hunt, no 9pm form, no note stuck in a queue for days. Follow one Tuesday through five people's eyes — a bilingual RBT, her BCBA, the QA lead, the biller, and the AI team working underneath them. The RBT talks between trials; an agent writes; specialist agents check every note; a ready-to-bill packet lands the same evening. Humans commit; agents draft and check.
Two hours, ten minutes — from the first trial to a claim that can go out. Compare that to the days of latency a note carries today.
The session starts before the notes do
Maria opens the client's card in the car. One authorization badge tells her everything that matters: the payer, the code she's billing, and how many units are left on this authorization. She taps Start session — and the three facts every payer audits are captured right then, at the door.
A server-side clock stamps the real start. Duration and units are computed, never remembered.
Florida Medicaid §6.2.4 requires location on every note. It's a tap, not a typed field.
Who was present, including whether a caregiver was there — the exact field auditors flag as missing.
EVV-ready by design. Time, location, and participants at check-in are five of the six data elements EVV requires — so if Florida re-expands its EVV mandate, the visit is already verified.
Mordió dos veces durante la transición al área de mesa. Redirigí con apoyo gestual y ofreció el token▍
She talks. In Spanish. In the moment.
Maria holds the mic for a few seconds between trials and describes what just happened — in the language she thinks in. The transcript streams live so she can see it caught the moment. Language is auto-detected across 90+ languages; the note that comes out later is compliant clinical English, with the original kept as evidence.
Every voice note is timestamped and attached to the session. Clear states — listening, processing, done — so she's never guessing whether it heard her.
Voice is an accelerant, never the only path. A quick typed note is always right there — noisy room, private detail, or just preference.
Dead-zone homes are handled. Many in-home sessions run where there's no signal. Capture is recorded locally on the device and uploads the moment signal returns — no lost notes, no re-recording from memory.
One data model, four ways in
Voice doesn't replace the session data — it fills it. Every path below lands in the same structured fields, in the database, validated against the client's treatment plan. The note is a rendering of that data — never the other way around.
Date, start/stop, place of service, participants, credentials, authorization — the device, the calendar, and the client record already know. The best capture is no capture.
Behavior counters, duration timers, prompt-level chips, Fair / Good / Poor — big one-hand buttons between trials. Numbers by tapping, not talking — a count beats a sentence.
ABC events, client response, environmental changes, observations — any language. The narrative-shaped facts that used to cost the 9pm hour, extracted into fields, not prose.
Whatever's still empty against the payer's required elements arrives as one-tap questions. A form that shrinks to just its gaps — and disappears when there are none.
The pre-filled form is the fifth path and the editing surface: everything the other four paths captured lands there, reviewable and editable until Maria signs. Professionals who prefer typing use it as capture, first-class, forever — voice is an accelerant, never the only way in.
Every behavior recorded must map to a target in the client's assessment and treatment plan; every replacement program must be a plan program. A behavior that isn't in the plan doesn't fail silently — it flags Dr. Chen that the plan may need updating. And every data row remembers its source: tap · voice 4:47pm · answered 5:33pm · edited before signing.
Structured session data is also what the note alone could never give the clinic: per-behavior graphs and data tables for reauthorization packets, progress analytics, and a capture schema that extends to other CPT codes as data — not code. See the data model →
The writer agent asks, instead of inventing
The writer agent reads the structured session data, the voice notes, and the client's treatment plan, and starts composing. When a required fact is missing, it does the one thing today's AI note tools don't: it stops and asks — two one-tap questions on Maria's phone, before the note exists.
A session note is a legal, billable record. If an agent guesses a prompt level or a caregiver's presence, it has fabricated clinical fact — the exact thing an OIG audit recoups on. Asking a targeted question is cheaper than a denial and safer than a guess.
"Ask, never invent — every fact in the note is Maria's, not the model's."
Which prompt level did the tacting program need today?
Was mom present for the transition work?
Two taps · the note builds itself from your answers
Only what's in the data. Safest for insurance audits.
A draft to approve, never a note to write
The note arrives as clean sections — the same structure QA and the payer expect. Each section carries a tiny source chip so Maria can see exactly where every line came from: a form field, a voice note at 4:47, an answer she gave at 5:33. Nothing is unattributed.
Session details · Behaviors observed · Interventions and response · Replacement programs · Narrative — the CASP-aligned shape that makes a note payer-defensible and QA-checkable.
"Only what's in the data. Safest for insurance audits." Set as clinic policy, not per-RBT whim — embellishment is a liability the QA agents would only have to catch later.
The ten one-tap actions — shorten, make objective, fix mistakes, and the rest — plus a custom instruction, plus edit-by-voice. A Spanish view for review; the stored note is English.
Credential-stamped (Maria R., RBT #XX-XXXX, 5:47 PM), always after session end, and it locks the version. The RBT signs every note — the BACB is clear the certificant owns its accuracy.
Draft, never commitment. Voice and AI produce an editable draft. The commit is a separate, deliberate act — Maria's signature — with the specifics read back and the version locked for the audit trail.
The BCBA reviews exceptions, not everything
The BCBA-lens agent has already read all fourteen of today's notes for clinical coherence with each client's plan. Dr. Chen opens her phone once: eleven are clean, three need her judgment — each flagged with a specific, citable reason, not a vague "please review."
11 clean — plan-aligned, all sections present Expand any one before you approve the batch
Response to intervention missing for ABC entry 2
Narrative doesn't reference the fading plan due this month
Narrative describes protocol modification — reads like 97155, not 97153
The BACB requires each RBT get supervision for ≥ 5% of monthly service hours. Review activity counts — so the tracker fills as Dr. Chen works, instead of a spreadsheet she reconstructs at month-end.
A missed cadence isn't a clerical slip — under TRICARE, a missed monthly 97155 recoups 10% of that client's claims for the whole six-month authorization.
Every note checked against the payer's rule pack
Today's QA is a sampling gamble — the industry reviews a sliver, and CentralReach's own data says 80% of notes fail at least one payer requirement. Dana's cockpit inverts it: the QA-lens agent checks 100% of notes against each payer's versioned rule pack in minutes, and hands her a short queue sorted by claim risk.
Not 5–10%. Every note, every rule, minutes after signing.
The queue Dana actually works — down from reviewing all 218.
Sunshine Health, Optum, TRICARE, Cigna — versioned, clinic-editable, no deploy.
| Note | RBT | Rule flagged · payer source | Risk |
|---|---|---|---|
| K.B. · 97153 | Sofia M. | Optum · stop-clock timestamp missing after 12-min break | High |
| M.L. · 97153 | Andre T. | Auth · units billed exceed units remaining | High |
| P.S. · 97153 | Maria R. | FL Medicaid §6.2.4 · caregiver-absence reason missing | Med |
| A.C. · 97153 | Sofia M. | TRICARE · narrative thin for a 3-hour session | Med |
Error patterns roll up by pod and by RBT — turning the QA rep's manual "track patterns for training" step into a live signal that fights a perpetually-junior workforce (RBT turnover runs 65–103% a year).
When Dana returns a note, it lands as a push on the RBT's phone with the specific fix — a one-tap correction during or right after the next session, not a rejection discovered next morning off the clock.
One note needs a quick fix: add the stop-clock time after the 4:12 break. Tap to open.
Eleven gates, all green, before the biller touches it
A claim line is only "ready to submit" when every one of eleven checks passes. By the time it reaches Robert, the agents have verified all of them and assembled the packet. His job is the one that has to stay human — pressing submit — not chasing eleven things across three screens.
member verified with payer on the date of service
covers 97153, 6 units billed ≤ 12 remaining
rendering RBT signature + date, all required fields
88 min → 6 units under the 8-minute rule
97153 rendered by an RBT — consistent
enrolled, and matches the claim (Optum 2026)
coded to highest specificity, supports necessity
12 (home) — matches the note's location
under the daily MUE and the 40 hr/week plan cap
six Cures elements on file where applicable
day 0 of 180 — comfortably inside the window, signed same day
The last step — an extension, not an API
Office Puzzle has no public API. We looked — there's no documented REST endpoint or webhook to push into.
So the loop closes with the PraxisNotes browser extension: Robert opens Office Puzzle, selects Maria's client, and the extension fills the session, note, and billing fields straight from the verified packet — field by field or all at once, acting as him, logged in as him. He checks the screen and presses submit. Office Puzzle generates the batch claim to the clearinghouse, exactly as it does today.
The extension ships in the billing-loop phase (months 2–3). Until then the packet works as a guided copy flow — and if Office Puzzle changes its screens, a versioned field map updates the extension without touching the platform.
The whole day, in one strip
The clinic owner doesn't operate a screen — they glance at one. This is Tuesday, closed out: the pipeline numbers the whole product exists to move, instrumented from day one.
Every one of those is a lever on the north-star metric — first-pass clean-claim rate — the number that turns a 30–60-day denial loop into a same-day fix. What we commit to measuring →
And for the 1,000+ solo RBTs
This isn't a separate product for clinics. Today's PraxisNotes users — the solo RBTs and BCBAs who already know the app — get the same core loop as an upgrade to the tool they use every night. Clinics unlock a team layer on top. One product, two tiers.
The solo tier keeps PraxisNotes' existing users on a faster, voice-first version of what they have; the clinic tier is where a 100+ RBT agency stops scaling QA with headcount. Same brain, same tools underneath — the agents that make both run →
By 6:10, Maria is home. The note is already billable.
No 9pm form, no queue, no rework loop. Five people each did the part only a human should — and a team of agents did everything in between. That team is the next page.
Meet the team that makes this run →