AI Can't Fix Collision Estimating Until the Shop Logic Is Built First

Key findings

1. 1AI is not the foundation of collision estimating automation. It is the top of a stack that begins with structured shop logic. Without the lower layers, a language model produces confident, well-written, frequently wrong recommendations.
2. 2The RocketPros Shop Logic Layer is a named five-layer estimating intelligence foundation: Source Truth, Rule Logic, Procedure Evidence, Shop Memory, and AI Explanation. Each layer has a specific job and a specific dependency on the layers beneath it.
3. 3Estimate data alone is insufficient. CCC, Mitchell Connect, EMS, BMS, and PDF exports each carry different limitations. A defensible recommendation has to preserve source line numbers, totals, labor categories, and evidence pointers across the entire pipeline.
4. 4Canada, particularly provinces with structured public-insurer programs such as MPI and SGI, exposes the importance of rule-driven estimating. That discipline is a strong proving ground for any estimating automation that intends to be more than a chat interface.
5. 5Shop Memory has to learn from outcomes without converting bad habits into rules. Approvals, denials, edits, and supplements adjust how the system understands a carrier or program. They do not override published rules or OEM procedures.
6. 6The estimator stays in the seat. The Shop Logic Layer is built to be reviewed, not auto-applied. The metric of success is files that close cleaner, with fewer supplements and faster cycle time, not abstract AI usage statistics.

Body

1. The wrong starting point

Most pitches for AI collision estimating start with the model. A language model reads the estimate, suggests missing operations, summarizes the file, and writes the supplement. On a slide it sounds correct. In a real shop it usually is not. The reason is that an LLM can describe what is on the page, but it cannot, on its own, tell you what should be on the page. It does not know that the vehicle is on an SGI accredited program, that the bumper R&I on this VIN triggers a specific OEM calibration, that the appraiser at this carrier has rejected the same blend operation three times in the last six months, or that the part you replaced is non-reusable under the program rules.

When the model has no logic layer beneath it, it does the only thing it can. It guesses confidently. In estimating, a confident wrong answer is worse than no answer. It walks past the estimator, ends up on the supplement, gets pushed back by the insurer, and erodes trust in the tool until it is quietly turned off. The shops that have tried generic AI tools almost all describe the same arc: an interesting demo, a few useful summaries, then a slow drift back to the human-only workflow because the recommendations could not be defended on review.

2. What shop logic actually is

Shop logic is the practical operating layer that sits between raw estimate data and a defensible recommendation. It is everything an experienced estimator has internalized over years of writing files, fighting supplements, and learning what each carrier accepts on this VIN, in this province, on this program, with this part, in this paint code. None of it is in the estimate by itself. All of it has to be present before a model is asked to read the file.

In a serious estimating system, shop logic is not a vibe or a prompt. It is structured data and structured code. Jurisdiction rules. Insurer and program rules. Estimating platform behavior. Labor categories. Refinish logic. Non-reusable parts logic. Repair versus replace contradictions. Overlap and included-operations checks. ADAS and calibration triggers. Documentation requirements. Shop-specific approval history. Estimator habits. Supplement patterns. Repair order lifecycle. Each one of those is a layer of detail the model cannot infer reliably from a CCC ONE or Mitchell Connect export. Each one has to be modeled deterministically before any AI is allowed to draft language on top of it.

3. Introducing the RocketPros Shop Logic Layer

The RocketPros Shop Logic Layer is the named foundation we build estimate review automation on. It is a pre-AI estimating intelligence stack composed of five layers, in a specific order. Each layer has one job. Each layer depends on the ones beneath it. AI is the last layer, not the first. That is the entire point.

1. 1Source Truth Layer. Preserves the estimate exactly as received.
2. 2Rule Logic Layer. Applies jurisdiction, insurer, program, labor, refinish, overlap, non-reusable parts, and supplement rules deterministically.
3. 3Procedure Evidence Layer. Connects recommendations to OEM procedures, ADAS triggers, repair operations, and required documentation.
4. 4Shop Memory Layer. Learns from accepted, dismissed, edited, denied, and approved recommendations without converting bad habits into rules.
5. 5AI Explanation Layer. Uses a language model to explain, draft, summarize, and help the estimator move faster, only after the layers below have done their work.

We are introducing this stack with a name because the industry needs one. The conversation around AI collision estimating has been dominated by tools that collapse all five jobs into a single prompt. That is the part that does not hold up in production. The Shop Logic Layer is how RocketPros separates the jobs, so each one can be built, tested, and trusted on its own.

4. Layer 1, Source Truth

The Source Truth Layer preserves the estimate exactly as it arrived. Line numbers, labor categories, totals, parts, operations, notes, attachments, and source format. CCC exports look different from Mitchell Connect exports. EMS drops fields. BMS keeps more, but not all. PDFs lose structure entirely. Photos and supplements arrive on different timelines. Source Truth is the part of the system that refuses to lose any of that.

It matters because every recommendation downstream eventually has to be cited back to a specific line in a specific estimate, on a specific platform, written by a specific estimator on a specific date. Without that audit trail, the estimator cannot verify the finding, the appraiser cannot review it, and the shop cannot defend it. Most generic AI tools normalize early and lose the source. The Shop Logic Layer normalizes alongside the source, never instead of it. Mitchell Connect automation done well is not a parser. It is a structured representation of the file that survives every later step and can always be cited back.

5. Layer 2, Rule Logic

The Rule Logic Layer applies the deterministic rules of estimating before any AI is involved. These are the rules that should never be hallucinated and never be guessed. Jurisdiction rules. Insurer and program rules, including MPI estimating rules, SGI estimating rules, US DRP program rules, and per-carrier expectations. Labor category rules. Refinish logic across blend, full, edge to edge, two-tone, and adjacent-panel handling. Non-reusable parts. Overlap and included-operations checks. Supplement-eligibility logic. Repair versus replace contradictions on the same panel.

These rules are code, not prompts. They run on the structured representation produced by the Source Truth Layer. They produce findings with specific line citations, specific rule references, and specific financial impact. Rule Logic is also where most of the value of estimate compliance lives. Files that pass through a clean Rule Logic Layer arrive at the appraiser with the contradictions already resolved, which is the part that actually shortens cycle time and reduces supplements.

6. Layer 3, Procedure Evidence

The Procedure Evidence Layer connects findings to the underlying OEM procedure, ADAS trigger, repair operation, and documentation requirement. This is the layer that turns a rule finding into a defensible position. A bumper R&I on a VIN with a forward-facing radar is not just a Rule Logic alert. It is a specific OEM procedure ID, a specific calibration trigger, a specific list of required pre- and post-repair scan operations, and a specific documentation package the program expects.

ADAS calibration estimating is the clearest case for this layer. Calibration should never be guessed by AI. It should be tied to vehicle data, the specific operation that triggered the requirement, the OEM procedure that defines it, and the evidence on file. The Procedure Evidence Layer maps the trigger to the procedure and surfaces the calibration line, the scan operations, and the documentation requirements before the estimate locks. The OEM remains the source of truth. The system makes that source visible at the moment the writer needs it.

7. Layer 4, Shop Memory

The Shop Memory Layer is what makes the system specific to your shop, your carriers, and your estimators. Every accept, dismiss, edit, denial, approval, and supplement outcome is signal. Shop Memory captures it and uses it to weight future recommendations. Over time, the system understands which lines this carrier has been approving at this shop, which lines have been historically negotiated, and which estimators have a specific pattern of supplements that points back to a missing item at write-up.

The trap in this layer is bad-habit learning. If a shop has been historically suppressing a calibration line because a particular appraiser pushed back, a naive feedback loop will learn to suppress it too. The Shop Logic Layer is designed to refuse that. Shop Memory adjusts how the system understands a carrier or program. It does not override an OEM procedure or a published program rule. The Procedure Evidence Layer continues to surface the required line, with evidence, even when past outcomes show it was dropped. That distinction has to be designed in deliberately. It does not appear by accident, and it is the difference between a system that improves and a system that quietly drifts.

8. Layer 5, AI Explanation

The AI Explanation Layer is where a language model finally gets to do what it is good at. By the time the estimate reaches this layer, the deterministic findings, the OEM procedure references, the documentation requirements, and the shop-specific context are all present. AI is asked to explain the finding in plain language, draft the supplement note, summarize the file for a manager, write a cover letter for a carrier, and surface patterns across files that a single estimator would not catch on a Tuesday afternoon.

This is also where AI earns its place on speed. A model that has been handed a structured finding with attached evidence can produce a clean, cite-backed explanation in seconds. A model asked to do that work without the layers beneath it produces a paragraph that sounds right and falls apart on review. Same model. Different inputs. Entirely different outcomes. That is the practical case for the Shop Logic Layer. It changes what AI is being asked to do.

9. Why the order of these layers matters

The order is not aesthetic. It is causal. Source Truth has to come first because every later layer depends on a faithful record of the file. Rule Logic has to come before Procedure Evidence because procedures are mapped against rule findings, not against raw text. Procedure Evidence has to come before Shop Memory because Shop Memory weights findings, and a finding without a procedure reference cannot be weighted responsibly. Shop Memory has to come before AI Explanation because the model needs to know which findings to emphasize, in plain language, for this shop and this carrier. AI Explanation has to come last because language is the layer that should never decide what is true.

Skip a layer or invert the order, and the failure mode is predictable. Skip Source Truth and recommendations cannot be cited back. Skip Rule Logic and the model invents rules. Skip Procedure Evidence and ADAS calibration estimating becomes a guess. Skip Shop Memory and every estimator gets a generic system that ignores their carrier mix. Put AI first and you get the demo that does not survive contact with the queue. The Shop Logic Layer exists in this order because that is the order in which a defensible estimate gets built.

10. Why Canada is the right proving ground

Canadian collision repair is a strong place to build this kind of foundation because the rules are more legible. In provinces with public insurer programs such as Manitoba Public Insurance and Saskatchewan Government Insurance, the program documents are published, the tariffs are explicit, and the scoring frameworks are written down. Shops that operate on these programs already treat estimating as rule-driven and documentation as part of the line. The discipline is not optional.

That discipline is exactly what the Rule Logic Layer and the Procedure Evidence Layer have to encode. A platform that can write a clean file on an MPI claim or an SGI claim, with the right tariff lines, the right non-reusable parts, the right calibration triggers, and the right documentation package, can usually be adapted to a US DRP program with less friction than the reverse. We are not claiming Canadian collision repair is harder. We are saying it is more legible. That makes it the right place to build.

11. What this looks like in practice

The framework is easier to see in concrete cases. Each of the following is something the Shop Logic Layer can resolve before AI writes anything. Each is also something a generic LLM, given only the estimate, will get wrong a meaningful share of the time.

• A vehicle marked not drivable on intake. Rule Logic and Procedure Evidence determine whether the program triggers a specific labor operation, a tow allowance, storage charges, or a documentation requirement. The model does not invent the rule.
• A bumper R&I on a VIN with a forward-facing radar. Procedure Evidence maps the operation to non-reusable clips and retainers, possible bracket replacement, and OEM-required calibration regardless of fault code. None of that is in the bumper line itself.
• A repair and replace operation on the same panel. Rule Logic flags the contradiction. AI Explanation drafts the plain-language note so the estimator can verify or correct it before the estimate locks.
• A refinish operation on an adjacent panel. Rule Logic checks blend versus full, edge to edge, overlap with body labor, and program rules on adjacent-panel refinish. The output is consistent across estimators because the rules are code.
• A recurring supplement pattern at this shop. Shop Memory notices when the same category of items appears on supplement repeatedly. AI Explanation surfaces the pattern back to the manager, with the underlying findings attached, so the gap can be closed at write-up rather than re-litigated on every file.
• A line a carrier has historically pushed back on. Shop Memory records the outcome. Procedure Evidence keeps surfacing the line when the OEM requires it. The system does not learn to drop a required calibration. It learns to attach better documentation to it.

12. The estimator stays in the seat

The point of the Shop Logic Layer is not to remove the estimator. The estimator is the most expensive and most valuable component of the workflow. They carry the relationships with appraisers, the memory of past disputes, the judgment about which fights are worth having, and the credibility that makes the file land cleanly on review. A system that tries to replace them ignores most of what makes the trade work.

The right framing is a second set of eyes. The Shop Logic Layer reads every line on every estimate, applies the rules, checks the OEM procedures, compares to prior approvals at this shop with this carrier, and surfaces the things the estimator would have caught on a calm day with no queue. On a queue of thirty files, that calm day does not exist. The system does not get tired and does not skip a line because the next file is overdue. It hands the estimator a ranked list of items to look at, with evidence attached, and lets the estimator decide. Body shop estimate review done well makes estimators faster and more confident, not redundant.

Evidence is the reason this works. A finding with no evidence behind it is just an opinion in a nicer format. A defensible finding has to show, on the same screen, what was found, why it matters, what rule or OEM procedure supports it, where in the estimate the issue appears, what the financial impact is, what documentation is needed, and whether similar items have been approved at this shop before. That is the difference between a tool that produces commentary and a tool that produces decisions. It is also the difference between AI that is impressive and AI that is defensible. Collision repair supplement automation that does not carry that evidence package is just faster commentary.

13. The RocketPros position

RocketPros and Accur8 are built from the shop floor outward. The team has spent years inside Canadian collision shops, MSO operations, Mitchell Connect workflows, MPI and SGI claim files, US DRP programs, and the supplement and reinspection cycles that determine whether a file is profitable. The product reflects that. It is not a chat box on top of an estimate. It is the Shop Logic Layer, with AI added where it belongs.

• Real estimating workflows. The system runs alongside how shops actually write files in CCC ONE and Mitchell Connect, not against a clean lab dataset.
• Canadian rule awareness. MPI estimating rules, SGI estimating rules, and other provincial program logic are first-class, not afterthoughts.
• Insurer and program differences. The same line behaves differently across carriers and programs. The system treats that as data, not noise.
• Platform realities. EMS, BMS, PDF, photos, supplements, and parts decisions are all part of the file, with structure preserved through every layer.
• Estimator review. Every recommendation is built to be reviewed, not auto-applied. The estimator is the decision-maker.
• Operational ROI. The metric is files that close cleaner, with fewer supplements, faster cycle time, and stronger program scores, not AI usage statistics.
• Evidence-first automation. Every recommendation carries the rule, the procedure, the location in the estimate, the financial impact, the documentation requirement, and the prior-approval reference where one exists.
• AI only where it earns its place. Speed, explanation, and pattern recognition are good uses of a model. Replacing the rules layer is not.

We do not claim approval, endorsement, certification, or formal integration with CCC, Mitchell, MPI, SGI, ICBC, or any private carrier. RocketPros is independent software. We are aligned to the published rules of the programs we support, and we are explicit about where we sit. The Shop Logic Layer is a collision intelligence layer built by people who understand the messiness of estimating. That is the entire pitch.

14. Final takeaway

Most AI estimating tools fail or disappoint because they assume the model is the system. In collision repair, the model is the last layer, not the first. The truth lives across OEM procedures, insurer rules, program documents, jurisdiction requirements, platform exports, photos, supplements, parts decisions, and the judgment of the people who write the files. The RocketPros Shop Logic Layer organizes that work into Source Truth, Rule Logic, Procedure Evidence, Shop Memory, and AI Explanation. The order is the product.

The future of estimating is not AI writes the estimate. The future is structured shop intelligence, with AI assisting where it actually belongs. That is the only version of this that holds up on Monday morning, when the queue is long and the appraiser has questions.

Frequently asked

Citations

1. [1]Manitoba Public Insurance, Body Shop and Glass Information portal. Published Light Vehicle Tariff, accredited repair program documents, and RPS materials.https://www.mpi.mb.ca
2. [2]Saskatchewan Government Insurance, Accredited Repair Program documentation, RPS scoring guidance, and published bulletins.https://www.sgi.sk.ca
3. [3]I-CAR Repairability Technical Support, aggregated OEM position statements and procedure references for North American collision repair.https://rts.i-car.com
4. [4]CCC Intelligent Solutions, Crash Course Report and ongoing analyses of severity, calibration frequency, and estimating workflow trends.https://www.cccis.com
5. [5]Mitchell International, Industry Trends Reports and analyses of ADAS, severity, supplement frequency, and platform behavior.https://www.mitchell.com/insights/industry-trends
6. [6]Society of Collision Repair Specialists, guidance on OEM procedure adherence, documentation standards, and estimating discipline.https://www.scrs.com
7. [7]Insurance Bureau of Canada, industry statistics on Canadian auto claims, severity, and collision repair economics.https://www.ibc.ca
8. [8]Repairer Driven News, ongoing coverage of OEM position statements, calibration policy, DRP program changes, and post-repair litigation.https://www.repairerdrivennews.com

What this looks like inside RocketPros

The audit logic, scoring, and documentation patterns in this paper map directly to four RocketPros modules. If you want this applied to your shop's real estimates, start with the module that fits the workflow you're trying to fix.

• RPS ComplianceTrack MPI, SGI, and DRP program risk before it affects scorecards.
• Estimate AnalysisCatch missed labor, materials, parts, and documentation gaps before submission.
• AutomationRead saved Mitchell, CCC, and Audatex files without manual upload.
• ADAS CalibrationSurface calibration triggers tied to sensors and OEM procedures.

← Back to all research