AI for CRE Certificate of Insurance Tracking: 4 Options

By Jude Lee · October 25, 2026 · Comparison

Two commercial real estate operations staff reviewing certificates of insurance at a conference table with a laptop

What a COI is, and why tracking it goes wrong

A certificate of insurance is a one-page summary of someone else’s insurance policy. In commercial real estate it shows up constantly: the sign installer on your listing, the photographer, the janitorial vendor at a managed property, a tenant whose lease sets a general liability floor — say $2M, purely as an illustration, since the actual number is whatever your lease negotiated — and the contractor doing a TI buildout.

The standard form most of these arrive on is the ACORD 25 “Certificate of Liability Insurance,” published by ACORD, the insurance industry standards body. Two things about that form matter operationally. First, it is a summary — the form itself carries language stating it is issued as a matter of information only and confers no rights upon the certificate holder. The policy and its endorsements are the actual contract. Second, the fields that matter most to a landlord or brokerage — additional insured status, waiver of subrogation, primary and non-contributory wording — are often buried in the description-of-operations box or referenced by endorsement number, not cleanly checkboxed.

So COI tracking breaks in three predictable places: certificates arrive by email and never get filed; expiration dates pass unnoticed; and nobody compares what the certificate says against what the lease or vendor agreement actually required.

Where AI genuinely helps on this task

Be precise about the capability. Modern AI models are good at reading a messy PDF and returning structured fields: insured entity, carrier, policy numbers, coverage types, limits, effective and expiration dates, certificate holder, and the free-text description block. That is document extraction, and it is the single most labor-intensive part of COI work.

An agent goes further: it monitors an inbox, extracts the fields, writes them to a system of record, compares them against stored requirements, drafts a deficiency email to the vendor’s insurance agent, and queues a renewal chase before expiry. That multi-step, action-taking behavior is the difference between a chatbot and an agent.

What AI does badly here: deciding whether blanket additional-insured wording satisfies a specific lease, interpreting an unusual endorsement, or judging whether a carrier’s financial rating is acceptable. Treat those as human checkpoints.

Extraction is the part worth automating. Interpretation is the part worth paying a human to do carefully.

Option 1: Spreadsheet and calendar reminders

One tab, one row per certificate, columns for entity, coverage, limits, expiration, and a date-based reminder. An ops coordinator keys in the fields as certificates arrive.

Wins when: you are tracking a few dozen active certificates, mostly listing vendors and a handful of tenants. Zero cost, zero integration risk, fully auditable.

Breaks when: volume crosses the point where data entry competes with deal work, or when multiple people touch the file and the “current” version becomes ambiguous. There is no shame in the spreadsheet. Plenty of brokerages genuinely do not have a COI problem big enough to buy software for.

Option 2: Dedicated COI compliance software

A category of purpose-built vendors exists for exactly this — companies such as myCOI, TrustLayer, and Jones market COI collection and compliance tracking products, often with document-reading features and automated vendor chasing. Check each vendor’s current product documentation for what their AI extraction actually covers, as feature sets move fast.

Wins when: COI compliance is a repeating, rule-heavy obligation across a managed portfolio, and you want someone else to own the vendor outreach workflow, the requirement templates, and the audit trail.

Breaks when: your COIs are entangled with deal-specific context living in your CRM or deal files, or when the software’s requirement model doesn’t match your lease language and you end up maintaining a parallel ruleset.

Option 3: A general AI assistant plus a documented skill

Here you use Claude, ChatGPT, or a similar assistant with a skill — a reusable, packaged instruction set that teaches the assistant to do this job the same way every time. The skill specifies the exact fields to extract, the output format (a JSON object or a CSV row), how to handle missing data (“return null, never guess”), and how to flag ambiguous endorsement language for human review.

The ops coordinator drops a batch of certificates in, gets back structured rows, pastes them into the tracker, and reviews the flagged items.

Wins when: you want the lowest setup cost and no procurement cycle. It is the fastest way to find out whether extraction quality is good enough on your documents before you commit budget to anything.

Breaks when: nothing is connected — a human still moves files and pastes output, and at volume the copy-paste becomes the bottleneck. Three quieter risks matter as much: there is no audit trail unless you deliberately build one, so you cannot later prove what the system read and when; there are no access controls on where certificates get uploaded, which is a real problem if they carry policy numbers and named insureds; and extraction behavior can drift silently when the underlying model updates, because nothing is re-testing it. If you run this option past a pilot, keep a small fixed set of certificates as a regression test and re-run it periodically.

Option 4: A custom agent connected through MCP

MCP (the Model Context Protocol) is an open standard for giving an AI assistant secure, governed access to your systems. A custom MCP server over your document store, property management system, and CRM lets the agent read the certificate, look up the requirement attached to that vendor or tenant record, write the extracted fields back, and open a task when something is short. Our walkthrough on connecting Claude to CRE systems via a custom MCP server covers the build pattern.

Wins when: COI data needs to live next to deal, lease, and vendor records you already maintain, and when you want one agent architecture serving several ops workflows rather than a point tool per task.

Breaks when: you underestimate maintenance. A custom build is software you now own — permissions, logging, model updates, and someone on the hook when it misreads a form.

The comparison below covers only the high-investment branch. Most readers should exhaust Options 1 and 3 first and only reach this fork once volume and integration pain justify real spend.

Buy a COI tool (high-investment branch)

Requirement templates, vendor chasing, and audit trails come pre-built. Good when COI compliance is a standalone obligation and you want a vendor accountable for it.

Build a custom agent (high-investment branch)

Your COI data joins your deal and lease data, and the same MCP plumbing serves other ops jobs. Good when integration with your records is the actual problem.

Which AI tool is best for this job

There isn’t a single best AI tool for commercial real estate, and any article claiming one is selling something. The honest answer is task-shaped: for reading documents and returning structured fields, a frontier general-purpose model with a tight skill definition is usually the strongest starting point. For chasing vendors through a compliance cycle, purpose-built software may beat a general model on workflow, not intelligence. Our task-by-task map of where AI helps in CRE is a better frame than a tool leaderboard.

How much should the agent do on its own

You’ll see a “30% rule” floating around AI discussions — the idea that an AI should handle only a bounded share of a workflow, with humans holding the rest. There is no authoritative standard behind that number, and we’d treat it as a rough instinct rather than a rule. The instinct is sound, though: pick your autonomy level deliberately. For COIs, a defensible split is extraction and data entry automated, renewal reminders automated, deficiency notices drafted but human-sent, and any judgment call on endorsement sufficiency escalated. If you want a structured way to pick, see our breakdown of AI agent autonomy levels in CRE.

Modeling the payback with your own numbers

Don’t take anyone’s headline savings figure, including ours. Build it from what you can measure this week:

COIs / month × minutes each

Step 1: time your current intake on ten real certificates

× loaded hourly rate

Step 2: convert to cost using your own ops payroll

× % the agent handles

Step 3: discount for review time — assume you still read every flagged item

Then add the part nobody models: what the recovered hours get redeployed to. An ops coordinator who stops keying certificates can run listing activity reporting or deal file QC.

On avoided exposure, don’t invent a figure. There are two places to look for something defensible: your own records — how many times in the past 24 months did a certificate lapse without anyone noticing, and what did remediating it cost in staff time, legal review, or a withheld vendor payment — and your insurance broker, who can tell you how your claims history and any prior incidents actually read to an underwriter. If neither produces a number you’d stand behind, leave it out of the model and treat it as the qualitative reason the task exists.

A two-week pilot

Pull 25 real certificates

Mix of ACORD 25s, carrier-specific formats, and at least three you know are deficient.
Write the extraction skill

Name every field explicitly. Instruct the model to return null rather than infer, and to quote the exact source text for any endorsement claim.
Score it against manual review

Have a human abstract the same 25. Count field-level mismatches. That error rate, not a vendor demo, is your decision input.
Decide where the human gate sits

Based on what it got wrong, set the review threshold before you scale anything.
Only then pick an architecture

If extraction works and the bottleneck is copy-paste, connect it. If the bottleneck is vendor chasing, look at purpose-built software instead.

If your 25-certificate test shows field-level error rates you can live with, the hard part shifts from reading to deciding — who signs off on a flagged endorsement, who sends the deficiency notice, and who owns the rules the agent checks against. If it doesn’t, keep the human doing the reading and fix the filing problem first. As of 2026, that is still a question you have to answer with your own documents, not with a vendor’s accuracy claim.

Not sure where to start?

Get a free automation audit: we map your deal pipeline, marketing, and back-office workflows and show you what's worth automating — before you spend a dollar.

Get a free automation audit