--- title: "Automating Subcontractor COI Compliance Tracking with AI Document Parsing" url: https://ishchuk.eu/blog/automating-subcontractor-coi-compliance-ai-document-parsing published: 2026-10-02T01:15:00.000Z updated: 2026-10-01T23:37:52.708Z tags: [construction, COI tracking, AI document parsing, Make.com, n8n, insurance compliance, ACORD 25] --- # Automating Subcontractor COI Compliance Tracking with AI Document Parsing The fastest way to stop chasing subcontractor certificates of insurance is to stop reading them yourself. An AI workflow that parses every incoming ACORD 25, checks the limits against your contract requirements, and flags non-compliant certs automatically cuts a 15-20 hour weekly admin burden down to under 3 hours, based on numbers the COI software vendors themselves publish. I've built versions of this for construction clients and I'll be honest about what it catches and what it doesn't. ## Why Manual COI Tracking Fails Volume kills you, and the failure mode is boring: certs arrive wrong, expire quietly, and someone has to re-chase them by phone. Quick definitions, since the acronyms pile up fast. A certificate of insurance (COI), usually the ACORD 25 form, summarizes a subcontractor's liability coverage at a point in time. The general contractor (GC) collects it before the sub starts work, verifies the limits, and tracks expirations. The workflow below assumes commercial general liability (CGL) as the anchor coverage. Industry numbers, mostly from COI tracking vendors, so treat them as directional: - A comparison by Billy Insurance puts manual review at roughly 30 minutes per certificate. In my experience an experienced compliance admin reads a clean ACORD 25 in 3-5 minutes; the rest of that half hour is broker follow-up, which is exactly the part automation eats. A GC with 200 active subs still burns around 100 hours per review cycle on that basis. - BCS, a COI tracking vendor, estimates spreadsheet-based tracking costs firms 15-20 hours weekly, about $36,400 a year in labor at $35/hour, before you count any actual loss. - Over 70% of certificates submitted on construction projects are initially non-compliant, per PrequalPilot's 2026 roundup. Most certs arrive wrong, so you're not skimming, you're re-requesting. - illumend cites research that 23% of construction certificates contain errors or coverage gaps that survive manual review. The risk side is worse than the labor side. If a sub's coverage lapses mid-project and someone gets hurt, liability flows uphill to the GC and the owner, since GC policies typically exclude independent-contractor work. A single uninsured incident can run past $500,000, a figure Certificial uses in its case-study math, and construction litigation keeps pushing required limits upward. One missed renewal can halt a pour or hold up close-out. ## What You're Actually Parsing: The ACORD 25 The ACORD 25 is the standard certificate of liability insurance form published by ACORD, the Association for Cooperative Operations Research and Development. One page, dense, and harder for software than it looks: - Checkbox-driven coverage types: general liability, auto, umbrella, workers' comp, employer's liability - Limit fields scattered across per-occurrence, aggregate, and products/completed-operations columns - Handwritten policy numbers and broker notes on scanned copies - Additional insured and waiver of subrogation flags that live in the description box or on endorsements, not in structured fields - Layout drift between carriers and brokers, plus the growing problem of AI-doctored certificates, which the National Insurance Crime Bureau flagged as fueling a new wave of insurance fraud One more thing a lot of people miss: a COI is not a contract. It's informational. The actual policy is the binding document. That's why endorsement verification (CG 20 10 and CG 20 37 additional insured forms, primary and non-contributory status) matters more than the front-page numbers, and why pure-OCR workflows underperform, they read text but skip structure and endorsements. ## The Workflow: Make.com + an LLM Vision Model Here's the setup I'd build for a mid-size GC, using Make.com because construction admins tend to already know it, though n8n works identically if you want to self-host. ### Step 1: Inbound capture A Make email webhook watches a dedicated compliance@ address. Every attachment gets saved to Google Drive or S3, and the sender's email domain gets matched against your vendor list. Anything from an unknown domain routes to a human before parsing, because you don't want to parse a phished cert from someone impersonating a broker. ### Step 2: Parse with a vision LLM Pass the PDF to a vision-capable LLM (GPT-4o, Gemini, or Claude) with a strict extraction prompt: return JSON with carrier, policy numbers per line, effective and expiration dates, per-occurrence and aggregate limits per coverage type, additional insured status, waiver of subrogation, and the named insured exactly as printed. Vision models handle handwritten policy numbers and skewed scans far better than classic OCR. Honestly, the parsing is the easy part in 2026. ### Step 3: Compare against the requirement matrix This is where the value is. Your matrix per contract type; typical commercial baseline in 2026: - CGL: $1,000,000 per occurrence, $2,000,000 aggregate, per-project aggregate preferred - Auto liability: $1,000,000 combined single limit - Workers' comp: statutory; employer's liability commonly $500K each accident / $500K disease-policy-limit / $500K disease-each-employee - Umbrella: $2M-$5M on commercial and public work, higher on federal/infrastructure jobs - Additional insured via CG 20 10 (ongoing operations) and CG 20 37 (completed operations), primary and non-contributory endorsement, occurrence-based coverage, waiver of subrogation - Carrier quality: an A.M. Best rating of A- VII or better is the usual commercial bar, so add an A.M. Best lookup to the workflow rather than trusting the carrier name the LLM read The workflow compares parsed JSON against the matrix and outputs a verdict: compliant, deficient (with the specific fields that failed), or unreadable (needs human eyes). I always keep that third state. And here's the trap: off-the-shelf vision LLMs are unreliable at reading endorsement edition dates, the difference between a gold-standard CG 20 10 11 85 and a newer, more restrictive edition is fine print that models hallucinate about. Never auto-pass a cert on endorsement language alone. Route endorsement verification to a human, every time, until you've audited your parser against a few hundred real packets. ### Step 4: Act on the verdict Compliant certs push into Procore's vendor record or a compliance sheet, dated. Deficient ones trigger an auto-email to the subcontractor, CC'd to their broker, listing the specific gaps. Brokers work for the sub, not for you, so emailing the broker directly without the sub in the loop is how automated follow-ups die in someone's inbox. Unreadable ones go to a review queue. ### Step 5: Renewal tracking A scheduled module checks expiration dates against a 30/14/7-day window and fires renewal reminders, demanding fresh certificates and endorsements before the policy term lapses. Don't overthink this. Endorsements ride the annual policy term; you don't "re-verify" a standing endorsement quarterly, you replace it at renewal. Most compliance failures are expiration failures, not bad initial certs, and strict renewal tracking is what actually catches them. ## What About Off-the-Shelf Tools? Fair question. BCS, myCOI, Jones, SmartCompliance, CertFocus, and Certificial all do COI tracking as a service: - CertFocus runs $6-$29 per vendor per year on the enterprise end, per Vertikal RMS - BCS quotes 24-48 hours with analyst review and claims 99.7% compliance accuracy - myCOI's platform review runs a day or two; the "two weeks" horror stories are broker lag, not software lag, so don't blame the tool for a sub who won't answer email I think there's an honest middle path here. If you're a 50-sub GC, buy the software, the math works without touching an automation platform. If you're larger, or your compliance rules are contract-specific and weird, or you already run Procore and want the verdict inside your existing vendor records, a custom Make/n8n + LLM workflow wins because you own the requirement matrix and the audit trail. And a hybrid, AI parses and flags, a human spot-checks the flagged ones and every endorsement, gets you most of the 80-90% time reduction the vendors advertise without pretending the machine is infallible. ## What This Costs and Returns Build cost for the custom route: roughly $5K-$15K as a consulting project, which matches the broader range SMBs pay for AI consulting work (most land at $10K-$15K for a 4-6 week engagement). Run cost: Make or n8n plus a few hundred LLM calls a month, maybe $50-$100/month total for a mid-size roster. Return math, using the vendor numbers with the caveat that they're vendor numbers: reclaiming 12-17 of those 15-20 weekly hours at $35/hour is $21,000-$31,000 a year in labor, and a single prevented uninsured-sub incident avoids a six-figure exposure. The labor saving alone pays back a $15K build inside a year. The risk avoidance is the part your insurance carrier actually cares about, and carriers increasingly ask about sub-compliance processes at renewal. ## Where I'd Caveat the Whole Thing The extraction accuracy numbers vendors publish (99%+) are on clean, standard ACORD 25 PDFs. Scanned faxes, broker-modified forms, and multi-page packets with endorsements are where it drops, hard. Design for the failure: keep a human queue, log every automated verdict with the parsed JSON attached, and never let the workflow pass a cert on endorsement language without human sign-off, because a cert saying "additional insured" on the front page doesn't mean the endorsement exists on the policy. That distinction is exactly what manual review misses too, so this isn't automation being worse than humans, it's both being bad at it and only one of them being cheap. ## FAQ ### What is a COI in construction? A certificate of insurance, usually the ACORD 25 form, is a one-page document summarizing a subcontractor's liability coverage at a point in time. General contractors collect it before a sub starts work, verify the limits against contract requirements, and track expiration dates so coverage never lapses mid-project. ### How much time does manual COI tracking take? Vendor research puts spreadsheet-based COI tracking at 15-20 hours per week, roughly $36,400 a year in labor at $35 per hour. A single certificate takes about 30 minutes to process manually including broker follow-up, and a GC with 200 active subs spends around 100 hours per review cycle. ### Can AI accurately read ACORD 25 certificates of insurance? Vision-capable LLMs like GPT-4o, Gemini, and Claude extract structured data from ACORD 25 PDFs, including handwritten policy numbers, with high accuracy on clean documents. Accuracy drops on scanned or broker-modified forms, and they are unreliable at reading endorsement edition dates, so a human review queue for endorsements is recommended. ### What insurance limits should a general contractor require from subcontractors? The 2026 commercial baseline is $1 million per occurrence and $2 million aggregate for general liability, $1 million auto liability, statutory workers' compensation, and $2 million to $5 million umbrella on larger projects. GCs should also require additional insured endorsements (CG 20 10 and CG 20 37), primary and non-contributory status, and carriers rated A- VII or better by A.M. Best. ### How much does it cost to automate COI tracking? A custom Make.com or n8n workflow with an LLM vision model costs roughly $5,000 to $15,000 to build plus $50-$100 per month to run. Off-the-shelf COI tracking services like CertFocus run $6 to $29 per vendor per year, and managed services like BCS review certificates in 24-48 hours. ### Should a GC buy COI tracking software or build a custom AI workflow? For a GC with around 50 subcontractors, off-the-shelf COI software is usually the better buy since the pricing works without any integration work. A custom Make.com or n8n plus LLM workflow makes sense for larger rosters, contract-specific compliance rules, or firms that want compliance verdicts inside Procore vendor records with a fully owned audit trail.