I recently sat on a MarshBerry FirstChoice webinar about putting AI to work inside an agency. The host asked the last question the way these always get asked: what advice would you give an owner who wants to deploy AI but does not know where to start? My answer surprised me as I said it. Stop trying to deploy AI. That framing is the mistake almost everyone is making.
The problem: shiny object syndrome with a new coat of paint
Walk into most businesses right now and the AI conversation sounds the same. Someone saw a demo. Someone posted on Facebook asking whether Platform A is better than Platform B. Someone is about to sign an annual contract for a tool they cannot yet explain the purpose of. I know the pattern because I have lived it. I am the owner who sees something new and thinks, that looks cool, I want to try that, I will figure out how to make it work later.
That instinct gets the order backwards. You end up buying a tool and then rearranging your business, sometimes even the kind of client you go after, to fit the tool. You get locked into a contract. Half the team adopts it, half ignores it, and six months later you cancel it and everyone is left wondering why you keep changing things. The next time you bring something in, they have already decided not to bother.
The root cause is simple. People are starting with the tool instead of the problem.
The Vantage Point: AI is a layer, not a replacement
Here is the line I keep coming back to with my own team. AI is a layer, not the replacement. Humans keep the judgment, the trust, and the coverage by design. AI carries the routine.
That is not a soft, feel-good statement. It is an operating boundary. It tells you exactly what to point AI at and what to keep your hands on. We are a person-to-person business. Clients want automation when it is convenient, and they will bind online and self-serve all day, across every age bracket. But when something matters, they want to know a person actually looked at it. They trust a person more than they trust the AI, because they know the AI still makes mistakes. If someone wants to buy insurance from an AI agent with no human in the loop, that is fine, but they are not my client and I am not building for them.
So I do not ask AI to solve judgment, trust, or knowledge problems. I ask it to solve process problems and friction points, the things that slow the business down and do not need a human touch. Once you draw that line, the right first question stops being “what tool should I buy” and becomes “where is the friction.”
The framework: the Friction-First AI Filter
This is the order of operations I now run before I will even look at a piece of software.
One: name the friction, not the tool. Walk your business from attract to capture to intake to quote to bind to service to renewal, and find the specific places where work slows down, leaks, or gets dropped. Not “I want a CRM.” Instead, “we lose two days getting a commercial quote out because the producer only goes to the two carriers they like.” That is a friction point. Tools come after you can name it.
Two: decide what stays human. For each friction point, ask what part genuinely needs a person and what part is just repetition. The repetition is the candidate for AI. The judgment is not.
Three: check the connection. The fastest way to create new friction is to add another disconnected platform. Before you buy, ask how the tool connects to the systems your team already lives in. My team lives in our CRM, so the question is always how something flows in and out of there, not whether it is impressive on its own. One more login that does not connect is a step backward dressed up as progress.
Four: prove it small. Try it with one person, on one task. If it works, expand it. If it does not, drop it and move to the next friction point. No annual contracts you cannot escape, no all-hands rollout of an unproven idea.
Five: earn the buy-in. Do not deploy AI down from the top. Go to a team member and ask what the most boring, repetitive thing they do every day is. Then go solve that. Do it department by department. When the first win removes something they hated, the team stops bracing against change and starts bringing you the next problem to solve.
The business application: where this has actually paid off, and where it has not
The honest part of this is that the filter tells you to turn things off as often as it tells you to turn things on.
I have rebuilt our entire website with AI as the layer, not the author. I described the friction in plain language, that I needed a fast, connected site with smart intake forms and a real learning center, and the AI built it on a platform most web developers will not touch. Work that one vendor quoted me at twenty-eight thousand dollars, and that five vendors over six years could not deliver, got built for a few hundred. But notice what stayed human. I put in the information and the judgment about who we serve and how we talk to them. I did not let the AI decide that.
Now the counter-example. I tried an AI phone system four separate times, and I turned it off every time. The technology kept improving, but a pattern kept showing up: my team knew the AI would pick up if they were busy, so they stopped answering the phone. And how do you make revenue in this business? You pick up the phone and talk to people. The tool was technically working and quietly costing me money. The friction it removed was less valuable than the behavior it created. The filter is what let me see that and walk away without feeling like I had failed at AI.
The risk lesson: your data is the exposure nobody reads the contract for
There is a risk angle here that owners in any industry should sit with, and it is squarely in my world as an insurance professional.
When you connect an AI tool or open an API to a vendor, you are deciding where your data goes and what is allowed to happen to it. Some systems will sell your data or use it to train their own models. In an agency, that data is client names, policy details, and personal information. Releasing that into a model that trains on it is not a tech convenience question, it is a liability and trust question. Before you turn anything on, your data needs to be clean and you need to know exactly how the vendor maps it, stores it, and uses it. Read the contract for the data clause, not just the feature list. The most expensive AI mistake is not the subscription you forgot to cancel. It is the client trust you cannot get back.
That is the same discipline we bring to coverage. You do not wait until after the loss to find out what the policy actually said. You do not wait until after the breach to find out what the vendor was allowed to do with your data.
Questions to sit with
What are the three places in your business where work most reliably slows down, leaks, or gets dropped? Be specific enough to name the step, not the department.
For each one, what genuinely needs a human, and what is just repetition you have been treating as if it needed you?
If you handed your team your current AI tool list, how many of those tools could each person explain the purpose of? And which one would they quietly be relieved to see go?
What is the single most boring, repetitive task someone on your team does every day, and what would it take to remove it this month?