AI agents for MGAs: from four weeks to four hours on claims and submissions
By Vishal Kumar, Founder and CEO · Last updated · 8 min read

In short: For a US managing general agent (MGA) we work with, AI agents read incoming claims, assign the codes and chase missing documents by email. Work that used to take about four weeks now takes about four hours once the agents are running. The agents work alongside the existing system, Applied Epic, rather than replacing it. Compliance rules decide what they may and may not do, and a person signs off anything that needs judgement.
Four weeks to four hours.
That's the line I used at an Austrade business matchmaking session last weekend, when someone asked what we do in financial services.
Here is how I said it:
"They use something called Applied Epic to manage all of their insurance contracts. So we deploy agents to read all the claims and come up with codes. There's a whole bunch of compliance things of what we can and cannot do. But broadly, what used to take about four weeks, we do in about four hours once the agents are running."
This article explains what sits behind that line. What the agents actually do. What they are not allowed to do. And where people stay in control.
What does an MGA back office look like?
An MGA is a specialist firm that a carrier authorises to underwrite and sell on its behalf. The MGA sits in the middle of a long chain.
On one program we are building, the chain runs like this: a carrier in the UK, its partner in the US, a wholesaler, a broker of record, and the business selling to the end customer. Almost everyone in that chain works by email and attachments. Only one party had real technology.
Most MGAs run on an agency management system. Applied Epic is a common one. It holds the policies, the clients and the history.
What it doesn't do is read.
Every claim and every submission arrives as a pile of documents. Someone opens each one, works out what it is, checks what is missing, writes back, waits, and then keys the result into the system.
So where does four weeks go?
Not in the software. In the reading, the chasing and the waiting.
| Way of working | Elapsed time |
|---|---|
| People reading, chasing and keying | About 4 weeks (672 hours) |
| Agents running, people approving | About 4 hours |
What do the AI agents actually do?
An AI agent is software that uses a large language model (LLM) to read, decide on the next step and take an action, within rules you set.
For claims and submissions, the flow looks like this:
- An email arrives with documents attached.
- The agent reads every attachment and works out what each one is.
- It checks the pack against the list of what this type of claim or submission needs.
- If something is missing, it replies to the sender and names exactly what to add.
- It keeps that email thread going until the pack is complete.
- It assigns the claim codes and prepares the record.
- It writes the result into the system of record, or queues it for a person to approve.
- Anything it isn't sure about goes to a person, with its reasoning attached.
I described the email part to a manufacturing investor last week like this:
"You email your documents; it'll reply back saying, hey, these documents are missing, please add these ones. And it maintains that entire email chain to the point where, okay, now I've got everything. Now I will go do something on the other end."
None of these steps is clever on its own.
The time saving comes from removing the gaps between them. A person handles a pack when they get to it. An agent handles it when it lands, at 2 am, on a public holiday, in the middle of month-end.
| Step | Who | What happens |
|---|---|---|
| 1. Email arrives | In | documents attached |
| 2. Reads every attachment | Agent | works out what each one is |
| 3. Checks the pack | Agent | against what this claim needs |
| 4. Assigns codes | Agent | prepares the record |
| 5. Approves | Person | anything that needs judgement |
| 6. Applied Epic | Out | record written back |
| If something is missing | Agent | Replies naming exactly what to add, until the pack is complete |
| If the agent is unsure | Person | Gets the file with the agent’s reasoning attached |
Do you have to replace Applied Epic first?
No.
rip out the legacy system and buy a new "AI-native" platform. It feels like progress. It also means a migration, retraining every user and a year of risk before anything gets faster.
leave the system of record where it is and put agents around it. They read what comes in, prepare what goes in, and write back through an integration.
We have a working Applied Epic integration, built in Venus Pro, our own insurance software. That lets agents read from and write to Epic, so staff keep working in the screens they already know.
The rule I follow: change the work before you change the platform. If the platform still needs replacing later, you'll replace it with a clear list of what the agents need from it.
What are AI agents not allowed to do in insurance?
This is the part people skip, and it's the part that decides whether a project survives its first audit.
In the US, insurance is regulated state by state. The National Association of Insurance Commissioners (NAIC) adopted a model bulletin on insurers' use of AI systems in December 2023. It reminds insurers that decisions or actions impacting consumers that are made or supported by AI "must comply with all applicable insurance laws and regulations", including the laws on unfair trade practices and unfair discrimination. It also expects insurers to "develop, implement, and maintain a written program" for the responsible use of AI systems. [1] States choose whether and how to adopt it, so check the rules in each state you write in.
In practice, that turns into a short list of limits you agree before any agent goes live. For example:
- Agents may classify documents, extract data, check completeness, draft replies, assign suggested codes and prepare records.
- Agents may not make a decision that affects a policyholder's cover, price or claim outcome on their own.
- Agents may not send anything that reads as advice or a coverage position without a licensed person approving it.
- Every action is logged: what the agent saw, what it decided, and who approved it.
When I said "there's a whole bunch of compliance things of what we can and cannot do", this is what I meant. The list is written down, the agent is built to it, and the logs prove it.
- December 2023
- NAIC adopted its model bulletin on insurers’ use of AI systems
- Source: NAIC [1]
Where do people stay in control?
Two places: when the agent isn't sure, and when the decision needs judgement.
The first one is a design choice most teams get wrong. If you ask an AI to sort an email into "complete" or "incomplete", it will always pick one. Even when it's only 60% sure.
So we always give it a third door:
"Always have a way for escalation where the system itself is not sure. Never assume. In fact, whenever you're writing a prompt, always give a way out. Otherwise, it will pick one of the options even if it is not sure."
In my experience, 80 to 90% of a well-defined flow can run fully automated. The rest goes to a person, with the agent's notes attached, so they decide in a minute instead of starting from scratch.
The second place is judgement. Claims outcomes, coverage questions and anything unusual stay with your people. The agent's job is to make sure that when a person looks at a file, everything they need is already there.
| Outcome | Share of a well-defined flow |
|---|---|
| Fully automated | 80 to 90% |
| To a person, with the agent’s notes attached | 10 to 20% |
Why do agent projects stall after the demo?
Edge cases.
I'm honest with clients about this, because it's where the time really goes:
"I can show you a demo in like 20 minutes that will be like, yes, we want it tomorrow. And then four weeks later, we're still fixing edge cases. Whereas an AI is like, please tell me what to do now, or else I will do the wrong thing, and I will not even tell you."
A person who meets an odd document either knows what to do or asks their manager. An agent needs to be told. So the first weeks of any rollout are spent finding those cases, writing the rule for each, and widening the escalation path.
Once that's done, it runs. And the escalation path catches what's left.
How should an MGA start?
Pick one flow. Measure it. Then automate it.
- Choose the flow with the most waiting. First notice of loss, claims coding or new submissions are common starting points.
- Measure today's baseline. Elapsed time from email received to record complete, and staff minutes per file. Without this, "four weeks to four hours" is just a story.
- Write the allowed and barred actions with your compliance lead before any build starts.
- Run the agent in shadow mode next to your team for two to four weeks. Compare its output to theirs.
- Turn on automation for the easy majority and route everything else to people.
- Review the escalations weekly. Each one is either a new rule or a reason to keep a person in the loop.
If you want a broader view of where AI fits in an agency or MGA, start with our guide to AI for insurance agencies, our US insurance page and the commercial insurance case study. If token costs worry you, read how to reduce AI costs.
- 2 to 4 weeks
- in shadow mode next to your team before automation goes live
- Recommended practice in this article
Your next step
Where does a claim sit waiting longest in your office: in the inbox, in the chase for documents, or in the queue for keying?
If you know the answer, that's your first agent. Get your AI roadmap.
Frequently asked questions
What is an MGA?
A managing general agent is a specialist insurance business that a carrier authorises to underwrite, and often bind and service, policies on its behalf. MGAs usually work through brokers and wholesalers, so their back office handles a high volume of documents and emails.
Can AI agents work with Applied Epic?
Yes. Agents can read from and write to Applied Epic through an integration, so the system of record stays in place. We have a working integration built in Venus Pro, our own insurance software.
Is it legal to use AI agents to process insurance claims in the US?
Using AI to support claims work is allowed, but the decisions still have to comply with existing insurance law. The NAIC model bulletin of December 2023 expects insurers to keep a written program for responsible AI use. Rules vary by state, so get advice for each state you operate in.
Will AI agents replace claims staff?
In our work, agents take the reading, sorting and chasing. People keep the judgement calls, the unusual files and the conversations with customers. Most teams use the time to handle more volume without hiring.
How long does it take to set up?
A working demo can take days. Getting edge cases handled and compliance sign-off usually takes several weeks. Plan for a shadow period before you switch anything to automatic.





