I built my own AI agent. Read this before you build yours.

Software is dying, the feeds say. I'm the proof it's partly true.

Every feed says the same thing now. Software is dying. A SaaS product is just a database with business logic on top, the logic is moving into AI, and soon nobody will buy tools they can build. Microsoft's CEO said the very notion of business applications will probably collapse in the agent era. The influencers took it from there.

I can't dismiss the take, because I'm the proof it's partly true.

I'm a commercial insurance broker. I don't have a computer science degree. And I built a working AI agent. It reads policies, drafts submissions, fills applications. The building part is real. It has never been cheaper or faster, and it will get cheaper and faster still.

So when a broker tells me they're thinking of building their own, I don't laugh. I ask one question.

Can you verify every answer before your name goes on it?

The software was never the expensive part

Building it took me far longer than the influencers admit. Still a fraction of what custom software used to cost, and the tools get better every month, so the direction of their argument holds. But the part nobody warns you about came after the build worked. Four things, none of which can be prompted into existence.

Being wrong quietly

An agent that reads a policy will be right most of the time. The tenth time, it reports a limit that an endorsement changed three pages later, and it reports it with full confidence. In most industries that's an embarrassing email. In ours it's a proposal with your name on it, and E&O practice is blunt on this point: professionals remain responsible for mistakes caused by the tools they use. The carrier defending you will not care that the sentence was generated. Your client's lawyer will care even less.

Someone else's data

A brokerage's file isn't your data. It's your client's operations, financials, loss history, and payroll, held in trust. Every DIY build has to answer where that file goes, who can see it, and what it trains. Most don't. Security researchers keep finding the same thing in AI-built apps: databases left open, access controls skipped, secrets exposed. One audit found that 45% of AI-generated code fails basic security tests. The breaches making headlines weren't built by careless people. They were built by fast ones.

The regulator already has an opinion

More than half of U.S. states have adopted the NAIC's model bulletin on insurers' use of AI. It expects a written governance program, documentation, testing, and oversight of third-party AI tools, and regulators can ask for all of it during an examination. Canadian regulators and privacy law are pointed the same direction. Buy a tool and you audit a vendor. Build one and you become the vendor you now have to audit, with none of the compliance staff a vendor has.

The demo is the easy five percent

MIT studied hundreds of enterprise AI deployments in 2025 and found 95% of pilots produced no measurable business impact. The detail that matters more: tools from external vendors succeeded roughly twice as often as internal builds. And the best-funded builder on earth already ran this experiment. Klarna shut down Salesforce and Workday to replace them internally, and within a year its CEO called the fallout tremendously embarrassing, said they never actually replaced SaaS with an LLM, and doubted other companies would follow. If a fintech with hundreds of engineers couldn't carry it, a brokerage running on producers and CSRs won't either. The build ends. The maintenance never does.

The question that replaces build vs buy

None of this means don't build. I'd tell any broker to build internal tools. A script that renames files, a prompt that drafts a first email, an agent that summarizes your own notes. If it breaks, you lose an afternoon. Build those freely, and you'll learn more about AI in a month than any webinar will teach you.

The liability never moves. Build it or buy it, the broker answers to the client for what goes out. So the real line isn't build versus buy. It's whether the output touches a client file. The moment it does, the standard stops being "does it work" and becomes "can you defend it." Can you show where every answer came from. Can you prove the client's data never left your control. Can you hand a regulator the documentation. Can you point at someone whose full-time job is keeping it accurate as forms, carriers, and models change.

Software isn't dying. Software that was just a database with a login is dying. What's replacing it, in any industry where being wrong has a price, is software built to be defended: citations, audit trails, permissions, and a team whose only job is keeping it accurate.

Full disclosure

I kept building. It turned into Broker Agentx, an AI workspace for commercial brokers, and it exists precisely because of everything above — every answer cited to the policy language it came from, client data encrypted, isolated, and never trained on, and a broker approving each step. It's free to start, so you can test it against your messiest tower and judge for yourself.

I built my own agent so you don't have to find out what it really costs.