Notes · AI & Insurance

AI Isn't Going to Run Your Agency. It's Going to Hand You Back Your Day.

Where AI actually helps an insurance agency, where it doesn't, and why most of the tools being sold to you right now will disappoint.

Calvin Zeng  ·  Gasha Capital

There's a particular anxiety I hear when AI comes up with agency owners. It isn't quite fear. It's more like a low hum of am I about to be made obsolete, and is the thing that replaces me going to be worse at this than I am?

It's a fair worry, because the hype is exhausting and the real experience has mostly been disappointing. Every other week there's a new platform promising to automate your agency and a demo that looks magical for ninety seconds, then falls apart the moment you feed it a real submission. Most people's hands-on experience with AI is some version of this: you ask it to do a real task, it does 80% of it confidently, gets 20% wrong, and now you have to check all of it anyway. Sloppy. Not worth it.

So here's an argument that runs against most of what you're hearing, and I'll back it up rather than just assert it.

AI is not going to do a broker's job — not because the technology is weak, but because of what the job actually is. What AI is genuinely good at is a narrow, specific slice of the work, and that slice happens to be the exact busywork that keeps a broker from the parts of the job only they can do. The opportunity isn't replacement. It's subtraction: taking the drudgery off the desk so the hours go back to selling and serving clients, which is the whole point of the business.

I've spent real time inside agencies — sitting next to producers and account managers, watching the work get done. This is what I've come to believe.

01What a broker actually does all day

Strip a broker down to the work and they're doing four jobs at once.

They're a salesperson — finding business, building relationships, getting in front of prospects, closing. They're an advisor — teaching clients about coverage they don't understand, packaging a submission so it gets the best rate, knowing which carrier will look favorably on a risk and which one will choke on it. That knowledge is hard-won and lives mostly in the broker's head. They're a glorified executive assistant — taking information from the client, shaping it, routing it to carriers, waiting, getting something back, routing it to the client, and repeating, all day. And they're customer support — fielding a never-ending stream of servicing requests, which in something like commercial trucking means certificates, endorsements, and changes without end.

At a big agency these roles get split across producers, account managers, and service staff. At a smaller shop one person does all four in the same hour. But no matter the size, the same thing is true: there is too much busywork. Even with a full service team, there's always more that could be done — more selling, more calls, more attention to the relationships that retain clients and grow the book. And it's the small, manual tasks that keep getting in the way. When Salesforce surveyed sales teams across industries, reps reported spending less than a third of their time actually selling. Anyone who's worked in an agency doesn't need the survey.

So the real question isn't can AI do the broker's job. It's can AI take enough busywork off the plate that the broker can finally do the work that matters. To answer that, you have to be honest about what AI is.

02What AI actually is

AI is not a magic box where you hand over the company — here are the files, here's what everyone does, go run it — and walk off to the beach. Anyone selling that is selling a fantasy.

AI today is best understood as an extremely capable, extremely enthusiastic intern. It will do anything you ask, tirelessly, forever — but it needs constant guidance, it goes off the rails, and it isn't consistent. Give it a fuzzy instruction and it will confidently do the wrong thing and feel great about it. The skill isn't having AI. It's knowing exactly what to hand the intern, how to constrain the task, and where a human has to check the work before it goes out the door.

It helps to understand why AI has improved so fast in some areas and barely moved in others, because it explains everything about where it will help your agency.

AI improves fastest where you can cheaply tell the model whether it did a good job. Take software. When AI writes code, you run it instantly — it compiles or it doesn't, the tests pass or they don't. That tight, automatic feedback loop let the labs train models against millions of examples with a clear right-or-wrong signal, and the models got staggeringly good. A controlled study found developers using an AI coding assistant finished a task 55% faster than those without. Google's CEO has said roughly 75% of new code there is now AI-generated and engineer-reviewed, up from about 25% a year earlier.

Insurance is the opposite kind of problem. The data is proprietary and locked inside agencies and carriers, so the models were never trained on much of it. And the feedback is mushy — there's rarely a clean right-or-wrong answer to "did the broker handle this account well." Quality is judgment, relationships, and context nobody wrote down. You can still teach a model to do specific insurance tasks well, but you have to build the guardrails for it, tuned to the exact task.

03Why AI is winning for engineers and flailing everywhere else

AI is winning, at scale, for exactly one group right now: software engineers. Not because engineers are special, but because their work has four properties almost no other job has. It's bounded — a function takes inputs and returns outputs. It's checkable — a compiler or test tells you in seconds whether it worked. It's structured — code runs through a deterministic pipeline, so the same input gives the same output. And it's verifiable — a reviewer looks at the change and says yes or no. Point AI at work that is bounded, checkable, structured, and verifiable, and the leverage is enormous.

Now hold that against everything else the vendors promised. They swore AI would transform sales, finance, operations, and the results have been brutal. MIT's NANDA initiative found that about 95% of enterprise generative-AI pilots delivered no measurable return, despite tens of billions in spend. The conclusion wasn't that the models are bad. The models are fine. The tools were brittle and badly fitted to how the work actually happens.

That's not a model problem. It's a shape of the work problem — and it's the best news in this essay for a broker, because it explains exactly why the job is safe.

A broker's day is the opposite of engineering. It's unbounded — "advise this trucking client" has no clean edge; it spills into coverage judgment, relationship history, market conditions, and what the client said on the phone last Tuesday. It's hard to check — no compiler confirms the recommendation was right; that shows up at claim time, maybe years later. It's unstructured — the information lives in email, PDFs, the AMS, a carrier portal, and the producer's memory, none of which agree. And it's hard to verify — "did we place this well" is a judgment a senior person makes, not a green checkmark.

Every property that makes engineering perfect for AI, brokerage lacks. That isn't a weakness in the technology; it's a structural fact about the work. The relationship, the judgment, the trust, the read on a client across the table — those are the job, and they sit squarely in the category AI is worst at. What AI is good at is the bounded, checkable slivers buried inside the day: the certificate request that follows the same pattern, the data entry after a call, sorting which emails need a human. That's the busywork. And that's what AI can take.

So the framing flips. AI doesn't replace the broker. It deletes the parts of the day that were never really brokering, and hands those hours back to the work that grows and keeps the book.

04Why the "AI-native" insurance platforms keep underwhelming

There's a wave of AI-native insurance software being sold right now. Some of it helps. But there's a ceiling on how much, and the reason is worth understanding, because it should change how you buy.

The shape of a workflow is different at every agency, even when the workflow is nominally the same.

Take certificate of insurance generation. Every commercial agency does it, and on paper it's identical everywhere: a request comes in, you produce a COI, you send it out. But underneath, each agency's version is unique. It writes to a specific set of carriers. It uses specific templates. Its approval rules are its own — maybe anything naming a particular additional insured needs a second set of eyes, maybe anything over a limit loops in the producer. Its data lives in its AMS, whether that's Applied Epic, AMS360, HawkSoft, or EZLynx, and each one stores and exposes information differently. The task is the same. The shape belongs to that one agency.

A platform sold to a thousand agencies has to pick a shape. It builds for the average, demos beautifully, then meets your real workflow and collides with the fifty places yours differs. It handles maybe 70% of the volume and breaks on the other 30% — and that 30% is now more work than before, because someone is cleaning up after the tool on top of doing the job. That's the exact failure pattern in the MIT data. The pilots didn't fail because the AI was dumb. They failed because the work was specific and the tool was generic.

To be fair to the other side: the same research found internal builds succeed at roughly a third the rate of buying from a focused partner. But dig into why internal builds fail and it's almost always the same thing — a team that doesn't sit inside the workflow builds for the documented process instead of the real one. The lesson isn't "buy generic" or "build blindly." It's that whoever builds this has to deeply understand the specific work.

05The real unlock: building software got cheap

Here's what's genuinely new. The marginal cost of writing software has fallen to nearly zero. Building custom software for your own operation used to be insane — a team of expensive engineers, months of runway, and you'd still end up with something half-broken. So nobody did it; agencies bought off-the-shelf and bent the business to fit the tool.

That math has inverted. With AI-assisted coding, one or two capable engineers can now build what used to take a team. For the first time, it's rational to build software that fits a specific workflow exactly instead of forcing the workflow to fit someone else's software.

The key is who builds it. Not an engineer writing abstract systems they'll never see used. The model that works — and the research keeps confirming this — is to embed in the business, sit with the people doing the work, watch it for weeks, map where the real bottlenecks are, and only then build automation fitted to the exact shape of how that agency operates. Two principles separate the systems that work from the 95% that don't.

Audit before you build. Map what actually happens, not what the procedure manual says. There's always a gap — the "I always check this spreadsheet first" step, the unwritten rule that anything over a certain size goes to the owner, the seventeen exception types that come up every month. Build for the manual and you build the thing that breaks. Build for reality and you build the thing that sticks.

Use AI as little as possible. This sounds backwards coming from someone arguing for AI, but it's the most important principle in the field. Language models are expensive and non-deterministic — the same input can give two different answers, which you do not want running an entire workflow. Use the model only for the narrow steps that genuinely require judgment, like reading an unstructured email and figuring out which carrier a risk might fit. Everything else — the lookups, comparisons, routing, formatting — should be plain, deterministic code that does the same thing every time and costs almost nothing. The best production systems are mostly boring code with one or two model calls placed exactly where judgment lives. Even Anthropic, which builds these models, says the same thing: find the simplest solution possible and only add complexity when it's clearly needed. Boring is the goal. Boring is what you can trust. It's also what keeps the cost down, because you reserve the expensive models for the few steps that actually need them.

06Where it actually helps

The principle is abstract until you see it on real agency work. A few of the places it matters most:

Inbox triage. A producer drowns in email — spam, inbound leads, real to-dos, carrier replies, client questions, noise. A well-built system reads the inbox, understands what each message is, surfaces what needs the producer, and quietly handles or queues the rest. Not "AI runs your email" — more like a sharp assistant who pre-sorts the pile so you open it to the ten things that matter instead of two hundred.

COI generation, with a human at the end. A certificate request comes in. The system recognizes it, pulls the right information, drafts the certificate and the reply, and hands it to the producer to glance at and send. The human stays — in insurance, the person checking the work is also the person managing the liability. But their job shrinks from doing the whole task to checking a draft. That's an order-of-magnitude cut in effort on something done dozens of times a day.

Knowledge and follow-up from calls. Calls can be transcribed live and reliably now. That puts the collective knowledge of the best brokers behind every broker — the right questions for each coverage type, what to look for, which carriers tend to accept a kind of risk. And after the call, the resulting to-dos — the data that has to go into the AMS, the follow-up emails to draft and chase — can be prepped automatically. That data entry and chasing is pure drudgery, and exactly the bounded work AI handles well.

Servicing, billing, endorsements. Heavily email-driven, heavily repetitive, the same shapes of task over and over. That repetition is what makes it a fit — not to remove the service team, but to take the mechanical parts off them so they can handle the judgment and the relationships.

The through-line: in every case the human stays, the judgment stays, the relationship stays. What goes away is the motion — the manual, repetitive motion that was never the value the agency provides.

07Where it doesn't

I'd be doing the thing I'm criticizing if I only told the good parts. Knowing the limits is how you build trust in the tool.

It should not give final coverage advice unchecked. Coverage judgment is the heart of the advisory role and it carries real E&O exposure — a wrong certificate or a bad recommendation is a liability, not just an error. The human in the loop isn't a nice-to-have; it's the control that manages the risk.

It should not handle client data carelessly. Agencies hold sensitive information, and "we fed it all into a consumer AI tool" is how you end up with a compliance problem. Done right, this means controlled systems with proper data handling, not pasting client details into a public chatbot. If a vendor can't say clearly where the data goes, that's your answer.

And it is genuinely bad at real judgment under ambiguity. The model doesn't understand insurance the way a great account manager does; it pattern-matches. For the bounded, repetitive stuff that's perfect. For the novel, the weird, the high-stakes — that stays human, and should.

None of this is a reason to wait. It's a reason to be deliberate about where you point it. The agencies that win won't be the ones who pour AI over everything and hope. They'll be the ones who understand their own work well enough to know which 30% is busywork to automate and which 70% is judgment that shouldn't be.

08Where this leaves us

The "AI replaces brokers" story has the technology backwards. The work that makes a broker a broker — the relationships, the judgment, the advising, the selling — is precisely what AI is worst at, and that won't change soon. What AI can do, today, is take the manual busywork that's been stealing hours from that work and shrink it to a fraction of the effort. Less drudgery, more selling, more service, a stickier book, a more valuable agency. A quieter opportunity than the hype, and a far more useful one.

This is what we're building at Gasha. We acquire and operate owner-operated insurance agencies for the long term, and the operating work — sitting with the team, mapping the real workflows, building practical systems around them — is the whole point of how we do it.

Sources: Salesforce, State of Sales · MIT NANDA, The GenAI Divide (via Fortune) · GitHub Copilot productivity study · Google AI-generated code figures · Anthropic, Building Effective Agents · Sica Fletcher, agency valuation