The 14 agents that run a dental practice on Alvero

Alvero runs 14 purpose-built AI agents organized into three tiers: Bookkeeper AI (7 agents), the Economic Layer (4 agents), and the Meta-Tier (3 agents). Each agent has a fixed scope, defined escalation conditions, and a human-in-the-loop classification. No agent guesses; ambiguous cases surface to a human queue.

Agent count — canonical

Bookkeeper AI: 7  ·  Economic Layer: 4  ·  Meta-Tier: 3  ·  Total: 14. These numbers are fixed in the platform schema. Custom agents are not supported in the current release.

Agent tier overview

The three tiers are hierarchical by data flow, not by importance. Bookkeeper AI produces the financial events that the Economic Layer interprets. The Meta-Tier learns from patterns across the economic signals and proposes improvements that feed back into practice configuration.

Tier Agents Primary function Human-in-loop model
Bookkeeper AI 7 Financial data ingestion, posting, reconciliation, and month-end close Ambiguous postings and failed integrity checks surface to review queue; close gated by Bookkeeper Supervisor
Economic Layer 4 Spend analysis, revenue recovery, operational efficiency, and daily owner brief Outputs are read-only advisory; no Economic Layer agent modifies practice data autonomously
Meta-Tier 3 Cross-tenant pattern detection, per-practice improvement coaching, and agent quality monitoring Green/Yellow/Red safety ladder; Red-tier changes require explicit owner approval before applying

Tier 1 — Bookkeeper AI

The Bookkeeper AI is the financial nervous system of the platform. Seven agents handle the full accounting lifecycle: data ingestion from three source systems, transaction posting, three-way reconciliation, month-end close, and a supervising meta-agent that gates every QuickBooks export. The tier runs continuously; nothing reaches QuickBooks until the Bookkeeper Supervisor passes integrity tests.

7 Bookkeeper AI agents

1. Dentrix Ingestor

Purpose
Pulls production, collections, adjustments, claims, and patient ledger events from the Dentrix API in near real-time and emits normalized canonical events downstream.
Inputs
  • Dentrix API (production ledger, collections, adjustments)
  • Dentrix claims feed (submitted, paid, denied, adjusted)
  • Patient ledger change events (new charges, payments, writeoffs)
Outputs
  • Normalized canonical financial events (JSON, versioned schema)
  • Ingestion-lag telemetry to the monitoring dashboard
  • Alert on any Dentrix API failure lasting >5 minutes
Escalation
Stops and pages the platform when the Dentrix API returns a 4xx/5xx for more than 5 consecutive polling cycles, or when a schema mismatch occurs in the canonical event model that cannot be auto-mapped.
HITL Tier
N/A — autonomous within bounds

2. Bank/Card Ingestor

Purpose
Connects to practice bank and credit card accounts via Plaid; pulls transactions, balances, and statement metadata daily and tags every record by tenant.
Inputs
  • Plaid Transactions API (connected bank accounts)
  • Plaid Investments / Liabilities feeds where applicable
  • Credit card account feeds via Plaid Link
  • Statement-cycle metadata (statement dates, cut-off times)
Outputs
  • Tagged transaction records per tenant (tenant_id, date, amount, merchant, account)
  • Daily balance snapshots for reconciliation input
  • Plaid webhook event log
Escalation
Escalates when Plaid Link tokens expire (user re-authentication required), when a connected account returns fewer than expected transactions for a statement period, or when balance drift exceeds configurable thresholds.
HITL Tier
N/A — autonomous within bounds

3. Document Ingestor

Purpose
Ingests bills, invoices, statements, and vendor PDFs arriving via email forwarding or manual upload; extracts line items using OCR and LLM with structured output validation before passing records downstream.
Inputs
  • Inbound email attachments (PDF, TIFF, JPG) forwarded to the practice's Alvero inbox alias
  • Manual uploads via Staff Portal or Bookkeeper AI surface
  • Vendor portal integrations (where available)
Outputs
  • Structured invoice records (vendor, date, line items, totals, payment terms)
  • Confidence score per extracted field; low-confidence fields flagged for human review
  • Raw document stored in R2 with tenant-scoped access path
Escalation
Any document where the aggregate confidence score falls below threshold, where a total cannot be parsed, or where the vendor is unrecognized is placed in the document review queue. The agent does not post unvalidated records.
HITL Tier
N/A — autonomous within bounds (escalates low-confidence items)

4. Posting Agent

Purpose
Applies posting rules against the practice's chart of accounts; routes ambiguous postings to a human review queue rather than making a best-guess entry.
Inputs
  • Normalized canonical events from the three ingestor agents
  • Chart of accounts (QBO-synced, per tenant)
  • Posting rule set (configured at onboarding, editable by owner)
Outputs
  • Posted journal entries (account, debit, credit, memo, source reference)
  • Posting review queue items for ambiguous transactions
  • Posting-rule match log for auditability
Escalation
Any transaction that matches no rule, matches multiple conflicting rules, or has a value above the configurable high-value threshold is placed in the review queue. The owner or designated bookkeeper resolves ambiguity before the entry is committed.
HITL Tier
Always supervised (ambiguous items)

5. Reconciliation Agent

Purpose
Performs three-way matching across Dentrix payment records, bank deposits, and merchant settlement reports; flags variances by amount, date, and counterparty.
Inputs
  • Canonical payment events from the Dentrix Ingestor
  • Bank deposit records from the Bank/Card Ingestor
  • Merchant settlement files (Stripe, POS terminals)
Outputs
  • Reconciled match report (matched, unmatched, partial-match items)
  • Variance flags with amount, date delta, and probable cause classification
  • Reconciliation summary for the Close Agent
Escalation
Unmatched items older than 3 business days, variances above configurable dollar thresholds, and any batch where the matched percentage falls below 95% are escalated to the review queue with diagnostic detail.
HITL Tier
N/A — autonomous within bounds (escalates variances)

6. Close Agent

Purpose
Owns the month-end close sequence: accruals, deferred revenue from prepayments, payroll allocations, and depreciation; produces a pre-close checklist and a post-close trial balance.
Inputs
  • Reconciliation Agent output (cleared and uncleared items)
  • Payroll data (payroll provider export or manual upload)
  • Prepayment and treatment plan ledger (from Dentrix Ingestor)
  • Asset register (depreciation schedule, configured at onboarding)
Outputs
  • Pre-close checklist (open items requiring resolution before close)
  • Draft accrual and deferral journal entries
  • Post-close trial balance (per GAAP / ASPE standards)
  • Close package handed off to Bookkeeper Supervisor for gating
Escalation
Any unresolved reconciliation item, any payroll variance, and any depreciation schedule anomaly blocks the close. The agent does not force-close; it surfaces a pre-close checklist and waits for human sign-off on outstanding items.
HITL Tier
Always supervised (close sign-off required)

7. Bookkeeper Supervisor

Purpose
Meta-agent that monitors the other six Bookkeeper AI agents, gates all QuickBooks Online exports, and only releases a close period when every integrity test passes.
Inputs
  • Output health signals from all six subordinate agents
  • Integrity test suite (configurable thresholds per tenant)
  • Human sign-offs from the review queue
  • Close package from the Close Agent
Outputs
  • Integrity test report (pass/fail per test, blocker classification)
  • QBO export authorization (or block with reason)
  • Closed-period lock record (immutable audit entry)
Escalation
The Bookkeeper Supervisor never overrides a failed integrity test autonomously. Any failure that cannot be resolved by re-running a subordinate agent is escalated to the practice owner with a structured diagnostic report. QBO export is blocked until resolved.
HITL Tier
Always supervised (controls QBO export gate)

Tier 2 — Economic Layer

The Economic Layer receives the verified financial events produced by Bookkeeper AI and transforms them into actionable operational intelligence. Its four agents are advisory: they surface findings, rank opportunities by dollar value, and compose the daily owner brief. No Economic Layer agent modifies practice data autonomously.

4 Economic Layer agents

8. Cost Cutter

Purpose
Analyzes supply spend, vendor pricing, and lab costs against network benchmarks; surfaces actionable savings opportunities with vendor-switch recommendations and ROI math.
Inputs
  • Posted supply and lab expense transactions (from Posting Agent)
  • Vendor invoice line items (from Document Ingestor)
  • Network benchmark data (anonymized, opt-in cross-tenant)
  • Vendor pricing reference database (maintained by Alvero)
Outputs
  • Ranked savings opportunities (vendor, current cost, benchmark, annual delta, switch difficulty)
  • Vendor comparison cards with contact and pricing detail
  • Savings summary for Owner Brief AI
Escalation
Cost Cutter surfaces findings; it does not execute vendor changes. If a finding requires price verification (e.g., a benchmark data point is more than 90 days old), it flags the finding as "unverified" rather than presenting it with full confidence.
HITL Tier
N/A — read-only advisory

9. Revenue Hunter

Purpose
Identifies unbilled production, missed insurance billing opportunities, dormant patients with active treatment plans, and recall gaps; ranks each opportunity by recoverable dollar value.
Inputs
  • Dentrix production and collections ledger (via canonical events)
  • Treatment plan status (Dentrix Ingestor)
  • Insurance eligibility and payment history (Dentrix claims feed)
  • Patient recall and appointment history
Outputs
  • Unbilled production list (patient, procedure, date, dollar value)
  • Dormant patient list with last-visit date and open treatment plan value
  • Insurance opportunity flags (underpaid claims, missed pre-auth)
  • Ranked opportunity summary for Owner Brief AI
Escalation
Revenue Hunter does not contact patients or submit claims. Any opportunity requiring patient communication routes through the Messenger surface with human approval. Claim re-submissions require explicit staff action.
HITL Tier
N/A — read-only advisory

10. Efficiency AI

Purpose
Measures chair utilization, provider production per hour, no-show rate, and same-day production rebooking performance; identifies the single highest-leverage operational improvement per practice per day.
Inputs
  • Appointment schedule and completion data (Dentrix)
  • Chair and operatory configuration (onboarding data)
  • Provider hours and production records (Dentrix Ingestor)
  • No-show and cancellation history (Dentrix)
Outputs
  • Chair utilization rate (per operatory, per provider, per day)
  • Provider production-per-hour ranking
  • No-show rate trend and financial impact estimate
  • Single "top lever" recommendation surfaced to Owner Brief AI
Escalation
Efficiency AI surfaces analysis and recommendations. Schedule changes, rescheduling communications, or staffing adjustments require human action via the Staff Portal or Messenger. Efficiency AI does not modify appointments.
HITL Tier
N/A — read-only advisory

11. Owner Brief AI

Purpose
Composes the morning briefing—yesterday's collections, today's chair load, open alerts, and ranked action items—and delivers it at 3:00 AM PT so it is ready when the owner wakes.
Inputs
  • Previous day's collections (Bookkeeper AI canonical events)
  • Today's appointment schedule (Dentrix)
  • Findings from Cost Cutter, Revenue Hunter, and Efficiency AI
  • Open review-queue items and unresolved alerts
Outputs
  • Structured daily brief (delivered to Flight Deck and optionally to email)
  • Ranked action item list with agent attribution and estimated impact
  • Alert summary (any blocked close, reconciliation variance, or Plaid issue)
Escalation
If no data is available from any Bookkeeper AI agent (e.g., Dentrix API down overnight), the brief notes the data gap explicitly rather than extrapolating. Owner Brief AI does not send the brief via SMS containing PHI; brief delivery uses the Messenger surface with human approval on any PHI-containing message.
HITL Tier
N/A — read-only advisory (delivery channel subject to Messenger approval rules)

Tier 3 — Meta-Tier

The Meta-Tier operates above the practice-level agents. Pattern Miner learns from anonymized cross-tenant data (opt-in). Per-Tenant Coach applies improvements per practice using a three-tier safety ladder. Agent Evaluator serves as the quality circuit breaker for every deployed agent.

3 Meta-Tier agents

12. Pattern Miner

Purpose
Runs cross-tenant pattern detection on anonymized, opt-in data to find practice-improvement signals that replicate across the Alvero network; never exposes or infers individual practice data.
Inputs
  • Anonymized economic signals from participating tenants (opt-in consent required)
  • Aggregated operational metrics (no patient-identifiable data)
  • Network-wide efficiency and cost benchmarks
Outputs
  • Validated improvement patterns (e.g., "practices scheduling hygiene re-care at 24h post-appointment have 12% higher recall rate")
  • Pattern confidence score and supporting sample size
  • Pattern candidates passed to Per-Tenant Coach for application
Escalation
Pattern Miner does not apply anything. All findings are proposals handed to Per-Tenant Coach. A practice that has opted out receives no cross-tenant patterns.
HITL Tier
N/A — read-only analysis (no write access to any tenant)

13. Per-Tenant Coach

Purpose
Applies validated improvements from Pattern Miner to individual practices using a three-tier safety model that governs autonomy by risk level.
Inputs
  • Validated patterns from Pattern Miner
  • Current practice configuration (posting rules, recall intervals, reminder settings)
  • Tenant-specific regression metrics (baseline performance for comparison)
  • Owner approval decisions (for Red-tier proposals)
Outputs
  • Applied configuration changes (Green and Yellow tier, post-validation)
  • Shadow-mode test reports (Yellow tier, 7-day observation window)
  • Owner approval requests (Red tier, with full change description and rollback plan)
  • Regression metrics report post-application
Escalation
See HITL Safety Ladder below. Red-tier changes never apply without explicit owner approval. Yellow-tier changes auto-apply only if no regressions appear after 7 days of shadow-mode observation.
HITL Tier
Green — auto-applies  Yellow — 7-day shadow, then auto  Red — owner approval required
HITL Safety Ladder — Per-Tenant Coach

Green tier (low-risk, reversible): changes auto-apply immediately. Examples: adjusting a recall reminder interval by ±1 week, updating a posting rule for a known vendor category.

Yellow tier (moderate impact): runs in shadow mode for 7 days. The proposed change executes in parallel with the existing configuration; outcomes are compared. If no regressions appear, the change auto-applies. If regressions appear, the proposal is discarded and a diagnostic is surfaced.

Red tier (high-impact or PHI-adjacent): requires the practice owner to review a full change description, expected impact, and rollback plan before any change applies. No timeout; the proposal waits indefinitely for an explicit approval or rejection.

14. Agent Evaluator

Purpose
Monitors every Alvero agent's output quality for 14 days following each deployment or configuration change; automatically rolls back any agent whose regression metrics breach defined thresholds.
Inputs
  • Regression metric baselines (established before each deployment)
  • Live agent output samples (post-deployment)
  • Human-resolved queue items (used as ground-truth signal)
  • Deployment manifest (which agent version, what changed)
Outputs
  • 14-day regression report per agent per deployment
  • Auto-rollback command (reverts to prior agent version if threshold breached)
  • Rollback notification to owner and Alvero platform team
  • Cleared status record (agent version certified stable after 14-day window)
Escalation
Auto-rollback executes without human approval when regression metrics breach threshold—this is a circuit-breaker, not a trust decision. The rollback notification is immediate and includes the specific metrics that triggered it. If the prior version also fails regression tests, the Agent Evaluator escalates to Alvero's platform team and suspends the affected agent pending manual review.
HITL Tier
Rollback is autonomous (circuit-breaker); escalation to humans on double-failure
← Architecture Security Model →