At the end of every month, the owner of a small independent insurance agency sits down with her office manager and reads through a stack of conversations. Not all of them. There are too many for that. But enough to get a feel for what people called, texted, and chatted about, and what the agency did in response.
This October was the first full month the agency ran with AI receptionists on its website chat and phone line through the WeKinnect Marketing Platform. The agency has four licensed agents, one office manager, and a part-time customer service representative. It writes home, auto, and small business coverage for people in a handful of towns, and most of its customers have been with the agency for years.
The owner expected the AI to answer the phone. What she didn’t expect was what it would show her about her own office.
The calendar topic for today is simple: share what AI helped uncover this month. This is that story, told honestly. It’s about four patterns that were sitting in the agency’s own conversations, what the team did about each one, and the human steps that stayed exactly where they were. It’s also about one pattern that turned out not to be a pattern at all.
Before the patterns, it helps to be clear about the job the AI was given, because that’s what made the patterns visible.
On the website chat, the AI receptionist answers common questions from the agency’s own knowledge sources: its website, its office hours, the lines of coverage it writes, and a short list of plain answers the office manager wrote and uploaded. When it doesn’t have enough information to answer, it says so and collects a name and phone number or email so a person can follow up.
On the phone, the voice receptionist answers from the same knowledge sources. If it can’t find an answer, it says so, offers to take a message, and captures contact details. It can transfer calls to set numbers based on what the caller needs and on conditions like the time of day. After-hours behavior is configurable, so the agency chose to take messages overnight rather than transfer. Recordings, transcripts, and summaries are saved for review, depending on the edition.
What the AI never does at this agency is give coverage advice. It doesn’t explain what a policy covers, recommend a deductible, or say whether a loss will be paid. Those questions go to a licensed agent, every time. The agency wrote that rule into the AI’s role instructions on the first day, and the office manager tests it with simulated conversations whenever the instructions change.
That boundary isn’t only good practice. The National Association of Insurance Commissioners, the organization of state insurance regulators, has a public page on artificial intelligence in insurance that makes two points worth repeating. First, tools that generate language can produce information that sounds accurate but is wrong, so AI-generated information should be reviewed carefully, especially for important decisions. Second, companies that use AI remain responsible for following insurance laws and consumer protection rules, and human oversight remains an important part of insurance decisions. The page is written mostly about insurance companies, but the principle applies just as much to a four-agent office on Main Street.
So the AI’s job was narrow: answer what it can from approved sources, capture what it can’t, and route it to a person. Doing that job consistently produced something the agency had never had before: a readable record of what people actually asked, when, and how.
The first pattern was the clearest. In the days after each batch of renewal notices reached customers, the number of calls and chats climbed sharply, and most of them asked a version of the same question: “Why did my rate go up? I didn’t file a claim.”
The agents had always felt this. Renewal weeks were busy. But they’d never seen it laid out across a month, with the timing lined up against the mailing dates. In the conversation history, it was obvious. The questions arrived in a wave, and the wave started a day or two after the notices landed.
What the AI did: It couldn’t answer the question, and it was set up not to try. It told each person that a licensed agent would review their renewal with them, collected the best number and time to call back, and created a follow-up for the agents.
What the people did: This is where the pattern became useful. Instead of reacting to the wave, the agency planned for it. The owner blocked out agent time in the week after each renewal batch, so callbacks happened the same day instead of piling up. And the office manager wrote a short, plain note that now goes out before renewal notices arrive, explaining that rates can change even without a claim.
That explanation isn’t something the agency made up. The Insurance Information Institute has a consumer article on renewing a homeowners or car insurance policy that explains the same idea: rates can rise because of broader regional and national trends, such as more frequent disasters and higher repair costs, even if you haven’t had a claim. It also encourages people to keep an open conversation with their insurance professional. The agency’s note says that in a few sentences and ends with an invitation: “If you’d like to go over your renewal, call us and one of our agents will walk through it with you.”
The note doesn’t answer anyone’s specific question. It isn’t supposed to. It prepares people for the conversation, so that when they call, the agent’s time goes to their policy instead of to the general explanation.
The second pattern was a small one that had been costing the agency for years without anyone noticing.
When the office manager looked at the missed calls and the calls the voice receptionist handled during business hours, there was a cluster in the middle of the day. Between roughly noon and one, calls that should have gone to a person were going to voicemail or to the AI. The reason was obvious once someone saw it: the whole office took lunch at the same time. Nobody had planned it that way. It was just habit.
What the AI did: During that hour, it answered, took messages, and, where the agency had it turned on, sent a missed-call text so the caller knew someone would get back to them. Missed-call text-back depends on the edition, and texting in the U.S. requires a registration step for business messaging, so that’s worth confirming before you count on it.
What the people did: They staggered lunch. One agent and the customer service representative now take an early lunch, the others a later one. Nobody is asked to skip lunch, and the phone is answered by a person through the middle of the day.
This is a good example of what AI is useful for in a small office. It didn’t solve the problem. It made the problem visible, and the solution was a scheduling change any office manager could make in an afternoon.
The third pattern came from the evening and overnight messages.
Most of what came in after hours wasn’t urgent. It was routine: people asking for a copy of their auto insurance ID card for a new car, a proof of insurance for a landlord or a lender, or a change of address. A few were more serious, such as people reporting a loss, and those were already handled by the agency’s process of directing people to their insurer’s claims line.
What the conversation history showed was that the routine requests had a rhythm. They clustered on weeknights, when people were dealing with paperwork after work, and on Sunday evenings.
What the AI did: It collected the details for each request (name, policy, what document was needed, and where to send it) and placed them in the shared inbox for the morning. It didn’t send documents or make changes. Those actions stay with staff.
What the people did: The customer service representative now starts each morning with the overnight queue, and routine requests from the night before go out by mid-morning. The agency also added one sentence to its after-hours greeting so people know when to expect a response. The result isn’t faster in a dramatic way. It’s predictable, and predictable is what customers notice.
One smaller change came out of this pattern too. The AI’s instructions can be different for each channel, so the office manager set text replies to be short and plain, and email replies to be a little fuller, with the list of what the customer service representative would need. Instructions like these change how the AI communicates, not what it’s allowed to do. That distinction is useful to remember: you can make the AI friendlier or briefer on a channel without accidentally giving it a new job.
The fourth pattern is the one every business should look for, because it’s about the AI’s own gaps.
The platform lets you review conversation logs to find common questions the AI couldn’t answer well, and on any AI response, you can open an explanation that shows its reasoning, which knowledge sources it used, and any tools it called. The office manager reviewed a sample of conversations each week, which is the rhythm the platform’s own guidance suggests as a starting point.
Two gaps stood out. People kept asking whether the agency could insure a boat or a small camper, and the AI kept saying it wasn’t sure, because the knowledge sources didn’t mention it. The agency does write those policies. It just had never put that on the website. And a handful of people asked, in different words, whether the agency could help with coverage for a home-based business. The AI answered generically, because the uploaded answers only mentioned “small business.”
What the AI did: It told people it didn’t have enough information and collected their details, which is exactly what it should do when its sources are thin.
What the people did: The office manager added two short pages to the website and two plain answers to the knowledge sources, one at a time, and tested each change before moving on. That one-change-at-a-time approach matters. If you change five things at once and something goes wrong, you won’t know which change caused it.
Here’s the honest part of the story.
Midway through the month, one of the agents noticed that several of the people who asked about canceling their policy seemed to live in the same town. It looked like a pattern: maybe a competitor was advertising there, or maybe the agency had done something wrong with those customers. There was some talk of a targeted campaign to win the town back.
The owner asked a simple question before anyone acted: how many customers does the agency have in that town to begin with? The answer was a lot. It’s the town where the office has been for twenty-five years. Of course more cancellation questions came from there. More of everything came from there.
Psychologists have a name for this. An illusory correlation is the appearance of a relationship that doesn’t exist, or the overestimation of one that’s weak. It’s one of the easiest traps to fall into when software makes patterns easy to see. The patterns that come out of your data are starting points for a question, not conclusions. A person still has to ask whether the pattern makes sense, what else could explain it, and whether it’s worth acting on.
The agency didn’t run the campaign. Instead, the owner asked each agent to call the customers who had asked about canceling, which was the right thing to do anyway. Most of them had questions about their renewals.
It’s worth listing the steps that didn’t change, because those are what make the AI trustworthy for a business like this.
1. Who answers coverage questions? A licensed agent, always. The AI routes these and never answers them.
2. Who approves messages that go to customers? The office manager approves templates and the renewal note before they go out. Any new automated message is reviewed by a person first.
3. Who owns each follow-up? Every routed conversation lands with a named person, and the daily digest of new messages goes to the office manager so nothing sits unassigned.
4. Who decides what the patterns mean? The owner and her team. The AI shows what people asked. The team decides what to do about it.
5. Who reviews the AI’s work? The office manager, weekly, by reading a sample of conversations and the explanations behind the AI’s answers.
If your agency uses any kind of scoring to sort follow-ups, by how ready a person seems to talk or how complete their request is, treat it as a sort order for your team’s attention, not a verdict on the person. Scoring isn’t available in every account setup, so ask before planning around it. And even where it is, the person on the other end of the call is a customer with a question, not a number.
To see how the same tools look in a different kind of organization, picture a small foundation that raises money for a community hospice.
Its October patterns might look similar on the surface. Messages arrive in waves after a mailing. Some questions come in after hours. Some come in when the office is at lunch. But the path a donor needs is very different from the path an insurance customer needs.
For the foundation, the AI can capture a message, route it, and remind a staff member to follow up. What it shouldn’t do is thank a donor. Many of the foundation’s gifts come from families who have just lost someone. A thank-you from a person who knows that, and who writes or calls personally, is part of the relationship. A form message, however well worded, isn’t. Stewardship language stays personal. The AI can make sure no one is forgotten. A person still does the thanking.
That’s the difference between a trade or service scheduling path and a nonprofit donor path. Scheduling is about getting the right person on the phone at the right time. Stewardship is about making sure the right person says the right thing, in their own words, at the right time.
There’s a difference between a chatbot and what the agency set up, and it’s worth naming.
A chatbot on a website captures a name and a question and sends it somewhere. That’s useful, but it stops there. What the agency set up does more of the work around the conversation. Calls are transferred based on what the caller needs and the time of day. Messages from chat, phone, text, and email land in one shared inbox. Follow-ups get assigned to a person. The office manager gets a daily digest. And the conversation history becomes something the team can read for patterns at the end of the month.
None of that replaces judgment. It prepares the people who still make the call.
If you’re thinking about adding AI to your phones or chat, here’s what to check first. These are the same things the agency checked.
1. What will it answer from? Read your website and your uploaded answers as if you were a customer. If something is outdated, the AI will repeat it. The AI uses this information only when it needs to answer a relevant question, so the information has to be right.
2. What must it never answer? Write those boundaries into the AI’s role instructions. For an insurance agency, that’s coverage advice. For a clinic, it might be medical questions. For a nonprofit, it might be anything about a specific donor or family.
3. Where does each kind of request go? Decide who gets transfers, who gets messages, and who covers the inbox when someone is out.
4. What happens after hours? Choose whether calls transfer or take messages, and tell callers when to expect a response.
5. What does your plan include? Some features, like text-back, voice, and call recordings, depend on the edition. Texting customers in the U.S. and Canada has its own requirements, including a registration step in the U.S. Confirm before you build a process around a feature.
6. Who reviews the AI’s work, and how often? Pick a person and a day. Weekly is a good start.
At the end of the month, the owner and the office manager went through their list. Renewal weeks get planned agent time and a heads-up note. Lunch is staggered. Overnight requests are handled first thing each morning. Two gaps in the website and knowledge sources are fixed. And one pattern turned out to be a coincidence, which was a lesson too.
None of those changes came from the AI deciding anything. They came from a team that could finally see what their customers were asking, and that took the time to decide what to do about it.
If you’d like to look at what your own conversations are telling you, and decide where AI should help and where a person should stay in charge, start a conversation with WeKinnect. We’ll help you set the boundaries first, and then read the patterns together.
Photo by Vitaly Gariev on Unsplash
Written By
Maren Hale focuses on brand positioning at WeKinnect Global Branding Agency. Guided by the core value of Service before Self, she writes for entrepreneurs and family-owned legacy businesses. Maren writes on Business Listings, Google Business Profile, SEO, GEO (being found by AI tools like ChatGPT, Gemini, and Grok), and Community. WeKinnect serves businesses across restaurants, arts, travel and tourism, banking, home services, nonprofits, and the pet industry. Maren brings that same spirit of service to every piece she writes.
Darlene Gagnon
16 Oct 2026
