AI in Production · Case Study
AI-Powered Certificate of Insurance Automation Case Study
How a US commercial insurance brokerage turned a mailbox full of certificate requests into an automated pipeline: one that reads the email, drives the agency system, and knows when to stop and ask.
The result: an AI pipeline replaces the 11 manual steps (about fifteen minutes) behind every certificate, produced 28 certificates from a single email request, and navigates 273 holders on one master certificate automatically.
Key Metrics of the AI Certificate Automation
- active insureds in the book
- 740
- manual steps to issue one certificate, before
- 11
- certificates produced from a single email request
- 28
- holders on one master certificate, navigated automatically
- 273
The Customer
A commercial brokerage whose busiest document is a certificate
A US commercial insurance brokerage writing commercial lines for roughly 740 active insureds: contractors, property managers, technology firms and service businesses across several states.
- Runs on Vertafore AMS360. Certificates are issued through eForms, a desktop application. There is no API behind it.
- Every contract asks for one. Every lease, subcontract and vendor onboarding needs a certificate naming the right holder, policies and endorsements.
- A small desk carries it. Constant, unglamorous, and unforgiving: a wrong holder address comes straight back.
At a glance
- Customer
- US commercial insurance brokerage
- Book
- Roughly 740 active insureds
- Agency system
- Vertafore AMS360, certificates issued through eForms
- Document
- ACORD 25 certificate of liability insurance
- Solution
- AI pipeline that reads requests and drives the agency system
The Problem
Every request arrived as an email. Every one was typed in by hand.
- 01Open the email.
- 02Work out which insured it belongs to.
- 03Find them in AMS360.
- 04Open their master certificate in eForms.
- 05Add the holder.
- 06Tick the endorsements.
- 07Print a PDF.
- 08Save it.
- 09Log an activity.
- 10Draft a reply.
- 11Attach and send.
Eleven steps, about fifteen minutes, repeated all day. Little of it needs judgment. All of it needs attention.
The AI Solution
API where there is one. Drive the desktop where there is not.
1. Read everything
The body, the attachments, and the attachments inside forwarded attachments. A prior certificate in the thread is usually the best source for the holder block, so it is unpacked and read alongside the request.
2. Match, then decide
The insured is matched on distinctive words, not prefixes. Every field carries a confidence score: a clear request goes forward alone, an unclear one stops and names the doubt.
3. Do the work
Open the master certificate, add the holder, tick the endorsements line by line, print that holder's PDF, and log the activity back to the customer file.
Every step is written down. What was read, what was matched, what was typed, what a person changed, and what the reasoning cost to run, visible per request.
See the AI in Action
1. Intake: the request arrives.
Requests land in the certificates mailbox and are read, classified and queued without anyone opening them. The panel underneath shows what the system did with this one: the request type, the matched customer with its confidence, and the attachments it read, not just the email body.
2. Review: what was understood, before anything is typed.
Every extracted field carries a confidence score. Certificate holder, as it will be typed shows the four boxes exactly as they will be entered, with counters against the printed line width; the care-of line sits on its own box so the company name cannot clip. The grid sets additional insured and waiver of subrogation per coverage line. An agent can correct any of it here.
3. Execution: the agency system, driven step by step.
Where the work actually happens. The automation opens the customer's master certificate in eForms, right-clicks into the holder form, types the holder, and ticks the endorsements line by line. The plan on the right ticks green as each step completes, so the desk can see what has been done and what has not.
4. Output: the certificate, and the reply that carries it.
The finished ACORD 25, traceable back to the request that asked for it. Alongside it: the audit trail for that certificate, and a reply drafted to the original sender with the document attached, ready for the desk to approve and send.
5. Judgment: knowing when to stop.
This request did not go through. Two fields came back below the confidence threshold: a holder address that appeared only in an attachment, and umbrella wording that implied an endorsement worth checking. So the system refused to issue and handed it to a person, naming the specific doubt. The button to run it stays disabled until someone confirms.
Challenges Faced Building AI Certificate Automation
The reading was the easy part.
No API to issue a certificate
Certificates exist only inside a Windows desktop application. Anything that issues one has to drive it the way a person does.
The holder list hides itself
273 holders collapse behind a single link. Printing one means expanding the list and finding the right leaf in a tree that truncates names.
Names do not fit
The printed certificate clips past a fixed width. A care-of line has to be split across two boxes, or the company name goes out cut off.
One email is not one certificate
A single request asked for 28 separate venues, each needing its own holder, certificate and PDF, from one block of prose.
Evidence hides in attachments
Forwarded requests bury the prior certificate inside another attachment. Until unpacked, the most useful document was invisible.
Knowing when not to act
Two insureds with similar names are easy to confuse, and a certificate issued against the wrong one is a real liability. It had to be able to stop.
The AI pattern travels further than certificates
Any back office running a system with no API for the thing that matters has the same shape of problem: a queue of requests, a document to produce, a desktop application in the way, and work that needs judgment often enough that nobody can let it run unwatched.
- 01
Read the request
Email, attachments, and the documents inside them.
- 02
Score the confidence
Act alone where it is clear, stop where it is not.
- 03
Work the system
API where there is one, the interface where there is not.
- 04
Write it all down
Every step, every correction, every cost.
Get In Touch