For years the industry has talked about modernization. New interfaces. Better integrations. More automation. Mobile access. Digital forms. AI assistants. Improved dashboards.
Some of that work has genuinely helped. Some newer companies are pushing real ideas around AI, document intelligence and automated service.
But we’re mostly still avoiding the larger question.
If you were building an independent agency today, with the technology now available, would you design its operating system anything like the software agencies actually use?
I don’t think you would.
The problem isn’t that agency management systems are old. It’s that they were designed around a world where people had to perform nearly all of the work.
Employees entered the data. Employees read every document. Employees remembered what happened. Employees created the tasks, updated the stages, filed the attachments, watched the deadlines, compared the quotes, chased the carriers and documented the outcome.
The software was built to support that. It stored the work. The human understood and coordinated it.
That’s the part AI actually changes. Not the interface. The operating model underneath it.
Which means adding AI to existing software isn’t the same thing as building software for an agency where AI does the routine work by default.
A policy is not the client
Insurance software puts the policy at the center of the account. But a policy isn’t the client. It’s one instrument addressing part of their risk for a defined period.
I work with real estate investors. A typical one owns properties through several LLCs, runs management through another entity, carries loans from three or four lenders, and holds policies that were written at different times by different people for different reasons.
That’s not an account with policies attached. It’s a relationship network with changing risk, and the software has no idea it exists. The connections live in my head, in email, and in whatever notes somebody remembered to write.
The center of the system should be a working model of the relationship. The people, the entities, what those entities own, who lends to them, what they’re exposed to, what’s actually protecting them, and what’s uncertain.
The roof was replaced in 2007, 2015, and 2018
Here’s the problem in one field.
The current policy says the roof year is 2007. The client says it was around 2015. The inspection report says 2018.
Traditional software wants one answer. So somebody picks one, types it in, and the other two disappear.
But there was never one answer. There were three claims, supported by three different kinds of evidence, of three different strengths.
What should exist is all three, with the strongest one surfaced and the conflict visible. Roof replacement 2018, per the inspection report, page seven. Client confirmation pending. Conflicting evidence from the 2007 application and the client’s 2015 estimate.
That difference matters for underwriting. It matters for the accuracy of what we submit. It matters enormously when someone inherits the file and has no idea which number to trust or where it came from.
The software shouldn’t just store what we believe. It should preserve how we came to believe it.
Tasks are not the work
Most systems manage work through tasks, activities, pipelines and tickets. Those are administrative representations. They aren’t the actual objective.
The real work is protecting a property acquisition. Completing a renewal. Correcting a policy discrepancy. Satisfying a lender before a closing date. Getting a client through a claim.
A task list can be fully complete while the actual outcome fails. I’ve seen a file where every task was checked and the client still wasn’t covered correctly, because the tasks described activity rather than the thing we were trying to accomplish.
The system should understand what we’re trying to achieve, what’s known, what’s missing, what’s contradictory, what’s committed, and what’s actually been proven done. Not whether somebody dragged a card into a different column.
The part nobody wants to talk about
Here’s where this gets genuinely hard, and where most of the conversation goes quiet.
If a system reads the inspection report and concludes the roof is 2018, and the carrier later disagrees, who owns that?
I’m the licensed agent. My name and my NPN are on the file. The E&O exposure is mine. So “the AI extracted it” is not an answer I can give a client, a carrier, or a regulator.
That’s not an argument against building this. It’s an argument about how it has to be built. Authority has to be a real, recorded thing rather than a description in a prompt. It has to be clear which decisions a system may take on its own, which require a licensed person, and who actually made each call, permanently and after the fact.
Anyone who talks about autonomous agency operations without talking about who holds the license hasn’t thought it through. The judgment, the advice, the negotiation and the responsibility stay with a person. That isn’t a limitation to engineer around. It’s the actual product.
Why the incumbents haven’t done this
It’s easy to say legacy vendors are stuck with old data structures. That’s true and it’s the shallow version.
The harder version is that carrier connectivity is the real moat, and it took decades to build. Agencies are conservative buyers who won’t move a book onto something unproven. And no established vendor wants to be first to market with a system that takes consequential action in a regulated business, because the first serious error becomes their case study.
Those are rational reasons. They’re also why change in this category tends to come from outside it.
What people should be doing instead
The goal isn’t fewer people in insurance. It’s people doing the part that requires being a person.
Reading, extracting, sorting, filing, comparing, monitoring, chasing and reconciling are not why clients hire an agency. Understanding the client’s situation, seeing what actually matters, applying coverage judgment, negotiating with a market and advocating when something goes wrong are.
Most agencies have that ratio backwards, not by choice but because the tools require it. Better technology shouldn’t help people do more administrative work in less time. It should remove the administrative work.
The point
The industry will keep improving what exists. Better interfaces, more automation, more AI features. That work is worth doing and plenty of agencies will benefit.
But there’s a difference between software that stores what happened and software that understands what’s happening.
The system of record told us what happened. The next one should understand what’s happening, what it means, and what has to happen next.
I don’t think that gets built by adding features to what we already have.
Richard Sweet, owner of Vantage Point Risk and an independent insurance advisor.