Every preliminary estimate leaves money on the table. Test drive time. Masking liftgate undersides. Ceramic coating removal prior to refinish. OEM one-time-use rivets and clips. The Big Three estimating databases classify these as "Not-Included" by design, forcing the estimator into encyclopedic cross-referencing against IP P-Pages, DEG rulings, and OEM repair portals — work that nobody pays for. Agent 01 watches every save event in CCC ONE, Mitchell, and Audatex, scrubs the VIN and operation list against a vector database of P-Page exclusions, and returns a pending supplement queue with pre-written DRP-compliant justification. The shop stops forfeiting gross profit. The adjuster stops getting away with it.
The modern collision estimate is a contested legal blueprint. Insurers writing under DRP matrices reward severity suppression. OEMs publishing procedures reward liability deflection. And in the middle, the estimator has roughly four minutes to assemble a repair plan that satisfies both — plus a P-Page lookup the database will never surface automatically.
CCC ONE's MOTOR-sourced logic explicitly excludes air bag system diagnosis, pre/post scanning, cooling fan operation tests, research of OEM repair procedures, test drives, and refinishing roof rails from its standard labor times. Mitchell excludes masking liftgate undersides, substrate material preparation, manual entry of specific seam sealers, and anti-corrosion material application. Audatex excludes OEM data subscriptions, release bonding hot-air blower usage, template creation, and alignment bench setups. These are not edge cases. These are operations that appear on a majority of modern repair orders and get paid zero times out of ten if the estimator forgets to manually key them in.
The Database Enhancement Gateway catalog makes the exclusions exhaustingly clear. DEG inquiries routinely confirm that none of the Big Three databases auto-populate ceramic coating removal prior to refinish, one-time-use fastener replacement (a standard OEM requirement spelled out in Toyota CRIB Bulletin 188 and equivalents across Ford, Honda, and Hyundai), seam sealer application, or the test drive required to validate ADAS calibration completion. The estimator must consult the OEM Electronic Parts Catalog, cross-reference the required clips, copy the part numbers, and type them into the system as separate line items — for every repair order, every day.
When these lines finally make it onto a supplement, the adjuster rejects them by default. Internal carrier compliance metrics punish adjusters for approving manual labor without an audit trail. So the shop submits DEG inquiry numbers, OEM procedure screenshots, and P-Page page references, and the negotiation loop burns another 45 minutes on a $48 line item. Multiply that by 15 estimates a day. The shop is not running a body shop — it is running a P-Page research library with a paint booth attached.
Missed P-Page lines are not a $50 problem. They are a $28,000-a-month problem hiding inside 330 repair orders and 42 hours of unbilled estimator research.
Agent 01 fires on a single, atomic event: the estimator saves a preliminary estimate or supplement draft in the shop's primary estimating system. CCC Secure Share publishes a BMS JSON/XML payload on every workfile write. Mitchell's Transactional API exposes an equivalent OData REST endpoint. Audatex ships via OAS3 Swagger endpoints at /api/v1/Claim. No polling. No user interaction. A webhook reaches the agent within two seconds of the Save click.
The payload includes VIN, selected labor operations with their P-Page flags, parts list with OEM part numbers, damage photo attachments, DRP assignment metadata, and the carrier ID. That is everything needed to run the scrub.
// Trigger payload (CCC Secure Share BMS envelope, abbreviated) { "eventType": "estimate.saved", "workfileId": "WF-2026-041532-A", "shopId": "SHP-7741", "carrier": { "id": "STATEFARM-DRP-T3", "matrixVersion": "2026.Q1" }, "vehicle": { "vin": "5TDGZRBH8NS123456", "year": 2023, "make": "Toyota", "model": "Highlander XLE" }, "damage": { "primaryImpact": "REAR", "severity": "MODERATE" }, "lines": [ { "op": "R&R", "part": "Liftgate Assembly", "hours": 3.2, "pPageFlag": "INCLUDED" }, { "op": "REF", "part": "Liftgate Outer", "hours": 2.8, "pPageFlag": "INCLUDED" }, { "op": "R&R", "part": "Rear Bumper Cover", "hours": 1.5, "pPageFlag": "INCLUDED" } ], "attachments": [ "preliminary-photos-12ct.zip" ], "savedAtUtc": "2026-04-15T13:42:08Z" }
Agent 01 runs a deterministic five-stage routine. Each stage is bounded, auditable, and produces a structured artifact the next stage consumes. No hallucinated line items. Every recommendation is evidence-linked.
Parse VIN, carrier matrix version, damaged panels, selected operations, paint codes, and attached photo count. Decode VIN against NHTSA vPIC for trim-level accuracy (powertrain, ADAS package, paint process).
Run parallel semantic queries against: (a) the IP's current P-Page vector store, (b) DEG rulings indexed by keyword + year/make/model, (c) OEM EPC for the exact VIN to enumerate one-time-use fasteners, (d) I-CAR Repairability Technical Support bulletins, (e) OEM repair procedure portal (Toyota TIS, Ford MoTIS, Honda ServiceInfo).
For each panel on the estimate, compare "operations required by OEM procedure + P-Page exclusions" against "operations currently on the estimate." The difference is the candidate supplement list. Filter for DRP-matrix compliance — skip lines the carrier explicitly prohibits.
Every proposed line item is bound to its evidence: "CCC P-Page, Body Labor section, line 4.7" or "DEG Inquiry #17842 — Ceramic Coating Removal." The adjuster receives a pre-defended argument, not a request.
Write the enriched line items back to CCC Secure Share via a workfile update call, tagged as "Agent-Proposed — Awaiting Estimator Review." The estimator sees a small yellow badge in the workfile. One-click accept or dismiss.
The output does not live in a separate dashboard. It lives inside the workfile the estimator is already editing. Proposed lines appear highlighted in amber, each one annotated with the P-Page reference, DEG inquiry number, or OEM procedure section that justifies it. The estimator's workflow is unchanged — only enriched.
Five proposed supplement lines totaling $204.22 of recoverable gross profit on a single repair order. Zero adjuster debate without the agent.
A mid-sized independent shop writes roughly 15 estimates per working day. Industry data from Collision Advice's "Not-Included Operations" reference chart and repeated DEG submissions confirm that the average preliminary estimate omits between $70 and $110 of legitimate, P-Page-backed manual lines. We use $85 as a conservative midpoint.
At $45/hour admin labor rate, the reclaimed 42 hours per month of estimator research time is worth an additional $1,890 in soft cost recovery — time the estimator redirects toward supplement defense, teardown documentation, and customer-facing QC. The hard-dollar recovery is $28k/month. The soft-dollar recovery is a better-run estimating department.
Agent 01 lives on a thin integration layer that brokers three classes of data: the shop's estimating write-back channel (CCC Secure Share or Mitchell Transactional), the evidence corpus (P-Pages, DEG, OEM portals indexed into a managed vector store), and the VIN-decode reference (NHTSA vPIC + OEM EPC).
| Function | Endpoint / Source | Auth Model |
|---|---|---|
| Workfile event subscription | cccsecureshare.com/webhooks | OAuth 2.0 + shop-scoped API key |
| Estimate write-back (CCC) | POST /workfiles/{id}/lines | OAuth 2.0 bearer |
| Estimate write-back (Mitchell) | PATCH /api/v1/RepairOrders/{id} | OAuth 2.0 + client cert |
| Estimate write-back (Audatex) | POST /api/v1/Claim/{id}/Operations | OAS3 + signed JWT |
| VIN decode | vpic.nhtsa.dot.gov/api/ | Public, rate-limited |
| P-Page corpus | Internal vector store (pgvector) | Private network only |
| DEG rulings | degweb.org search API | Scraped + indexed nightly |
| OEM EPC | Per-OEM partner portal | OEM-issued service credentials |
The agent runs on Anthropic's Routines framework. The mission is narrow, the inputs are structured, the constraints are strict. No latitude for creative interpretation of P-Pages.
// Agent 01 — Routine: ppage_scrubber_v1.4 MISSION Identify labor operations and parts that OEM repair procedures and IP P-Pages require but the estimator has not yet added to the workfile. Return a strictly-cited supplement queue. INPUT - workfile.json (CCC/Mitchell/Audatex BMS payload) - vin_decode.json (NHTSA vPIC + OEM trim data) - carrier_matrix.json (DRP restrictions for this assignment) STEPS 1. For each damaged panel on the workfile, enumerate OEM- required operations from the repair procedure portal. 2. For each panel's refinish op, query the IP's P-Page corpus for "Not-Included" exclusions (masking, coating removal, seam sealer, anti-corrosion, test drive, scan). 3. For each R&R operation, query the OEM EPC for the VIN and enumerate one-time-use fasteners, clips, rivets, bolts. 4. For ADAS-flagged vehicles, append required pre/post scan and calibration verify lines with P-Page reference. 5. Exclude any line that the carrier_matrix.json explicitly prohibits. Do not propose lines that will be auto-rejected. 6. For every surviving candidate, attach exactly one primary citation (P-Page section OR DEG inquiry # OR OEM procedure section). No citation = no line. CONSTRAINTS - Never invent a DEG inquiry number. Cite only verified entries. - Never propose a labor op without a source document reference. - Never exceed DRP-allowed hours per op. Flag with "review" if OEM procedure exceeds matrix allowance. - Return empty queue if zero gaps are detected. Do not pad. - Output must validate against CCC/Mitchell/Audatex line-item schema before push. Reject invalid payloads at stage 05. OUTPUT SCHEMA { workfileId: string, proposedLines: [{ operation: string, hours: number, laborType: "BODY"|"REF"|"MECH"|"FRAME", partNumber?: string, dollars: number, citation: { source: string, reference: string, url?: string }, confidence: 0.0-1.0 }], suppressed: [{ operation: string, reason: string }], runtimeMs: number }
Agent 01 is the fastest of the five agents to go live because its integration surface is narrow and its success metric is unambiguous (dollars recovered per estimate). A typical deployment follows a 3-phase rollout over roughly 4 weeks.
Provision CCC Secure Share or Mitchell Transactional API credentials. Subscribe to workfile write events. Mirror the shop's DRP carrier matrix (pulled from the SMS) into the agent's configuration layer. Establish OEM EPC credentials for the top 5 makes in the shop's mix (typically Toyota, Ford, Honda, Chevrolet, Nissan — covers ~70% of US volume).
Run the agent in read-only mode for 80–120 real estimates. The agent logs its proposed supplement lines to an internal dashboard but does not write to the workfile. Estimating lead reviews the output daily and confirms the citation accuracy. This is where trust is earned.
Enable write-back. Proposed lines now appear inline in the workfile with a "Proposed by Agent" badge. Estimator accepts or dismisses with a single click. Instrumentation captures acceptance rate, adjuster approval rate on proposed lines, and dollars recovered per RO. Target: >75% acceptance rate by end of week 4.
Minimal. Agent 01 does not require the shop to restructure its SMS or retrain its estimators. Required inputs:
Week 1: credentials, config, carrier matrix mirror. Week 2: shadow mode, accuracy audit. Week 3: live on 1 estimator. Week 4: live shop-wide. First full month of measurable ROI lands in month 2.
The agent earns its seat at the estimating desk by being right. Shadow mode is non-negotiable. The adjuster already distrusts the shop — the estimator cannot afford to distrust the agent.