Home/AI Suite/AI for Credentialing
Credential OS · Production AI · 86/86 UAT Passing

The AI mechanics behind Credential OS.

Most credentialing AI is a marketing claim wrapped around a checklist. Credential OS carries production AI across document extraction, deterministic Pre-Flight scoring, dollarized Revenue-at-Risk projection, a single audited LLM Gateway, and an RS256-signed Credential Passport. This page shows the mechanics, not the brochure.

See Credential OS walkthrough → See the AI in action
Claude API via LLM Gateway 20 deterministic Pre-Flight rules RS256 Credential Passport
Deterministic by design

Pre-Flight runs 20 rules. Every save. Sub-100 ms.

AI is great at pattern recognition. AI is terrible at being deterministic about whether a malpractice policy is in force, whether an NPI is exactly ten digits, or whether a BCBA's BACB certification is active this month. So we did not use AI for Pre-Flight. We wrote twenty deterministic rules, run them on every profile save, and emit a 0-to-100 score with explicit BLOCKER, WARNING, and PASS levels per rule.

The result is a packet quality gate that does not hallucinate. Industry first-pass rejection rates run 15 to 25 percent. Our packets reject at low single digits because Pre-Flight refused to let them leave at less than 100. The card to the right is one provider's live readiness view.

Pre-Flight Validator · provider: Sarah Chen, BCBA
94/100
✓
NPI present, 10 digitsNPI-001
Pass
✓
DOB present, validBIO-001
Pass
✓
SSN encrypted, 9 digitsBIO-002
Pass
✓
CAQH attestation currentCAQH-001
Pass
✓
CA license active, unexpiredLIC-CA-001
Pass
✓
BCBA certification active (BACB lookup)CERT-001
Pass
!
Malpractice expires in 32 daysMAL-001
Warn
✓
Education history completeEDU-001
Pass
✓
Work history 10 years completeWHS-001
Pass
✓
All required docs attachedDOC-001
Pass
✓
Tax ID present, encryptedTAX-001
Pass
✓
DEA not required for BCBADEA-N/A
N/A
+8 more rules verified · last run 14s ago · 84 ms total
One choke point, one log

Every AI call routes through the LLM Gateway.

Most healthcare AI products are a sprawl of API key configurations across point integrations. Finance has no idea what AI is costing. Security has no idea what data left the building. Engineering has no idea which prompt version is live. We solved all three by putting every model call behind a single HTTP interface that emits a structured log row per call.

The view to the right is a real log slice. Each row is one model call. Each carries the timestamp, the purpose tag, the prompt summary, latency in milliseconds, and cost in cents. CFOs run cost queries against this log. Security runs PHI-scrubbed exports. Engineering rolls back prompt versions by config flip, not deploy.

Timestamp
Purpose
Prompt
Latency
Cost
14:32:08
extract
state-license / v3
5,841 ms
$0.014
14:32:14
assistant
tool-use / search-providers
812 ms
$0.003
14:32:22
extract
dea-cert / v2
4,103 ms
$0.011
14:32:31
assistant
tool-use / get-revenue-at-risk
684 ms
$0.002
14:32:42
extract
malpractice-coi / v2
7,229 ms
$0.018
14:32:51
classify
document-type / v1
418 ms
$0.001
14:33:02
extract
board-cert / v3
3,847 ms
$0.009
14:33:14
assistant
tool-use / run-preflight
741 ms
$0.003
8 calls · last 70 seconds · 4 purposes Total $0.061
Dollarize the lag

Revenue-at-Risk, computed per provider.

The hardest argument in credentialing operations is the budget request to fix the bottleneck. Cycle time charts do not move CFOs. Dollar numbers do. Revenue-at-Risk is the math that translates one provider sitting in enrollment limbo into the revenue cost of that limbo, projected against the closing window and the readiness score.

Adjust the sliders for one provider scenario. Pick the title (BCBA, RBT, MD, BCaBA) for the per-day revenue assumption, set the days remaining in the 90-day window, and the current Pre-Flight readiness score. The calculator shows the dollar amount at risk if enrollment does not close in window. Defaults are calibrated to anonymized client averages.

Revenue-at-Risk Calculator · single provider
Provider title BCBA
Days remaining in window 45days
130456090
Pre-Flight readiness 68/100
05080100
Revenue at risk · this provider
$11.5K
Daily run rate at title
$533
Risk if pushed to next window
$32K
Math: days × daily rate × (1 - readiness/100) · scales per title
The AI flow

From roster upload to signed Credential Passport.

Four AI moments in the Credential OS lifecycle. Each one shaves days, removes manual data entry, or surfaces a blocker before it costs money. Together they are the difference between 22 days and 90.

AI 01 · Discovery

NPPES auto-enrich.

NPI to live registry pull. Name, taxonomy, primary practice address, license number, board credential. Fills only blank fields. Auto-creates a state license from the taxonomy. Non-destructive every time.

AI 02 · Extraction

Claude reads the docs.

Confidence-routed per field. State licenses, DEA, board certs, malpractice COIs, certifications. High-confidence applies. Mid-confidence opens review. Low-confidence falls back to manual. Audit log per call.

AI 03 · Assistant

Tool-use assistant.

Claude with three RBAC-scoped tools: search-providers, get-revenue-at-risk, run-preflight. Credentialing managers ask natural-language questions and get structured tool output, not chatbot prose.

AI 04 · Passport

RS256 signed bundle.

JWS of all credentials. Asymmetric keys; private signing inside your environment, public verify endpoint published. Other organizations verify the passport offline without contacting your CVO. Tamper-evident.

What AI moves

Measured outcomes from AI-native credentialing.

Across the credentialing book. Anonymized client benchmarks; individual results vary by roster composition and how quickly clients respond to outreach.

87%
High-confidence auto-apply rate
Per-field confidence scoring on state licenses, DEA, board certifications, malpractice COIs, and certifications. The other 13 percent splits between mid-confidence review (9 percent) and low-confidence manual entry (4 percent).
84ms
Pre-Flight Validator full run
All 20 deterministic rules execute on every save. Sub-100 ms latency means coordinators see readiness scores live, not as a periodic batch job. Auto-runs on create, every update, and every NPI enrich.
22d
BCBA enrollment average
Industry baseline runs 90 to 120 days. Compression comes from parallel commercial and Medicaid submission, NPPES auto-enrichment, and Pre-Flight catching what payer reviewers would have rejected before submission.
Common questions

Frequently asked questions: AI for credentialing.

What model powers the AI document extraction?
We run Anthropic Claude through our own LLM Gateway. Every call is audited by tokens, cost, latency, prompt version, and purpose. The PDFBox heuristic extractor is a deterministic fallback when no model key is configured so the workflow degrades gracefully rather than blocking. Confidence scoring is per extracted field; high confidence auto-applies, mid confidence opens a review modal, low confidence falls back to manual entry.
How accurate is the AI extraction?
Across the document types in production (state licenses, DEA certificates, board certifications, malpractice COIs, certifications), the high-confidence auto-apply rate is consistently above 87 percent. Mid-confidence rate is around 9 percent. Low-confidence and unreadable documents are around 4 percent. Every extraction is paired with the source page so coordinators verify quickly when they need to.
What does Pre-Flight actually check?
Twenty deterministic rules. NPI present and 10-digit. DOB present and valid. SSN present and 9-digit. CAQH attestation current and unexpired. State licenses present, unexpired, and matching the state of practice. BCBA certification active for BCBA providers (BACB lookup). Malpractice policy in force with declared limits. Education and work history complete. All required documents attached for the provider's payer mix. Each rule emits BLOCKER, WARNING, or PASS with the rule ID. The 0 to 100 score is the weighted sum.
How is Revenue-at-Risk calculated?
For each provider currently in enrollment, the calculation is: (days remaining in the 90-day window) × (daily billable revenue for the provider's title) × (probability the enrollment closes within window, computed from current readiness score and historical TAT distribution). Per-title daily revenue defaults are: BCBA $533/day ($16K/month), BCaBA $300/day, RBT $233/day, MD $833/day. Defaults are calibrated to client-actual averages at engagement start.
What is the LLM Gateway and why does it matter?
Every AI call in Credential OS routes through one HTTP interface that logs tokens consumed, model cost in micros, end-to-end latency, the prompt version used, and the purpose tag of the caller (extraction, assistant, classifier). Three things this enables: finance has one cost report instead of vendor sprawl; security has one PHI boundary instead of N model integrations; engineering has one prompt-version registry, so when a model upgrade changes behavior, the rollback is a single config flip, not a deploy.
What is the Credential Passport?
A cryptographically signed JWS bundle of a provider's credentials (licenses, board certs, malpractice, NPI). The current key format is RS256 (asymmetric). Other organizations verify the passport offline against our published public key endpoint without contacting our CVO. Tamper-evident: any payload mutation breaks the signature. Use case: provider moves from your practice to a referring partner who can verify the active license, DEA, and board cert in seconds, not 14 days.
What does the conversational AI assistant inside Credential OS do?
The assistant uses Claude tool-use to run search_providers, get_revenue_at_risk, and run_preflight on RBAC-scoped data the user is authorized to see. A credentialing manager can ask 'which BCBAs are at risk this month' or 'what is the readiness score for Sarah Chen' and get an answer with structured tool output, not just a chatbot response. All calls go through the LLM Gateway.
What is the WhatsApp and SMS automation?
Twilio-backed outbound automation for missing documents and credential expirations. A daily 9am cron sweeps the roster and queues nudges. Two-way conversation: STATUS replies return current readiness; HELP replies return the missing items. Inbound webhook is permitAll-scoped with phone-number matching for security. End-to-end verified on production. Live with TWILIO_* environment variables; sandbox without them so the workflow degrades cleanly.
What about the MCP server?
Credential OS exposes a spec-compliant Model Context Protocol JSON-RPC 2.0 endpoint (protocol 2025-03-26). External AI agents authenticated with a JWT can call Credential OS tools subject to the same RBAC scope as a human user. Initialize, tools/list, and tools/call are all curl-verified. Use case: connecting an upstream AI orchestrator to credentialing workflows without writing custom integration code.

Want a CTO walkthrough? Bring the roster.

A 60-minute CTO walkthrough of Credential OS. Live extraction on a sample document. Pre-Flight run on your real provider roster. Revenue-at-Risk projection on the spot. No NDA wall, no sales deck. The same demo we run in serious diligence.

Get the CTO walkthrough See credentialing service