I spent a couple of evenings looking at your office and at how reasoning AI — the Claude-and-GPT kind, not the movie kind — is actually showing up in three-person insurance agencies right now. Four short stories below, plus one simple tool for your own desks. Five minutes to read. If anything speaks to you, tell me; if not, tell me too.
The AI that makes the news — Lemonade's claim bots, State Farm's OpenAI partnership — is enterprise-scale, built over years, costing millions. That's not what this note is about.
What's actually useful for a three-person captive agency is less glamorous. A reasoning model from Anthropic (Claude) takes the repetitive writing, translating, drafting, and routing off your desks. Not a scripted bot with if-then rules — a model that reads each incoming message, drafts a reply in your tone, and hands it to you or Brenda for approval when there's any doubt. No new software to learn. No new hires. Most of the work either happens on its own in the background, or it shows up as drafts you approve in twenty minutes a week.
Where it runs: directly on Anthropic's own hosted runtime — not on a third-party vendor platform sitting between you and the model. No Birdeye middleman, no AgencyZoom renting you back your own data. Your Anthropic account, your tools, your bill.
How much does it actually help? In similar three-person service operations, the reported outcome is roughly 20–30% more time spent on real conversations with customers, less on typing and translating. Not a formula I'd stake a promise on — a benchmark to aim at.
One non-negotiable upfront: before anything below runs live, each piece needs a five-minute OK from your State Farm field consultant. Carrier compliance isn't an afterthought; it's the first checkbox. If something can't clear it, it doesn't ship — and I'd rather discover that up front than after a build.
Below are four places in your day where that AI would be visibly at work, plus one bonus tool for your own laptops.
María and her husband are shopping auto insurance on Saturday night. They fill your quote form at 9:30 PM. The message sits in Brenda's inbox. By Tuesday morning when she opens it, they've already filled forms at a handful of other agencies on Sunrise Highway. Whoever replied first tends to win.
This is not a scripted auto-responder. A Claude agent reads the incoming message, figures out what it actually is — quote request, claim question, policy change, simple curiosity — detects María's language, and drafts a reply in your office's tone. Within ninety seconds she gets a bilingual SMS back: "Got your quote request — Luis or Brenda will follow up Monday 9 AM. If anything urgent, reply here and we'll pick it up." If she replies with "¿pero cuánto más o menos?", the agent reads the follow-up and drafts an informed next reply for you to approve. Same flow for after-hours calls that go to voicemail: bilingual text-back in ninety seconds. Everything threads into one inbox so Brenda sees a real conversation on Monday, not a cold form.
Your weekend stops being a leak. Spanish-speaking families hear back in Spanish, in a channel where messages get opened almost every time, in under two minutes. Monday mornings start with a short list of already-warm conversations instead of an unsorted form queue. Setup is a dedicated agency SMS number, bilingual templates the AI uses as starting material, and about a week of tuning on your tone.
Six hundred-plus Instagram posts, about two hundred twenty followers. The feed is reposted State Farm graphics and generic PSAs, English only. Meanwhile, every week your agency actually does things: a Quotes for Good donation, a Chamber meeting, a customer who walks out saying "home and auto, best agent I've worked with." None of it reaches the feed. Your best content is happening in the office and never leaving it.
Once a week you or Brenda drops three or four short notes into a shared form: "Donated this week to Lindenhurst Food Pantry via Ricardo's quote", "Closed home-and-auto with the Aponte family", "Chamber breakfast Thursday morning." A Claude agent — trained on your 233 reviews and your office's voice — drafts two or three bilingual posts from them: short carousel, quote card, reel script. The Spanish version isn't a translation of the English; it's a parallel draft written directly in Spanish with the right rhythm. You review for twenty minutes, hit approve, it publishes.
The feed starts sounding like your office, not State Farm corporate. Half your reviewer base is Hispanic; half your posts finally reach them in Spanish. Quotes for Good turns into a monthly beat instead of a line in your website footer.
María bought auto from you last September. Her house is with Liberty Mutual. Her son turns seventeen in six months. Her auto renews August 12. She hasn't heard from your office since January. Your reviews are full of "got home and auto with Luis" — customers who noticed the bundle on their own. How many Marías didn't, and are sitting quietly in your book?
Once a month, you or Brenda pulls the usual CSV from your agent portal. A Claude agent reads it end to end — not a filter, an actual read — and picks out the situations that matter: auto-only customers with a home elsewhere, renewals inside sixty days, quotes that didn't close. For each one, it drafts a short bilingual outreach in your voice, grounded in that customer's actual history: "Hey María, Brenda here. Your auto renews Aug 12. Noticed your house is with Liberty — a bundle would save around $X/year, want me to pull a quote?" Brenda reviews, approves, sends. Every message goes out from your office, nothing from any outside system.
One honesty note: this one leans hardest on the compliance check up top. It assumes your agent portal lets you pull that data, and your State Farm agreement permits it going to an outside AI even for drafting. Both are the kind of thing worth a five-minute check with your field consultant before we build any of it.
Cross-sell opportunities stop quietly slipping. Renewals stop being last-minute scrambles. Revenue per customer grows without spending another dollar on new leads. Your book deepens instead of just widening. The monthly CSV becomes a ten-minute habit, not a project.
233 Google reviews at five stars, which is remarkable. Your Yelp page has been empty for years. Your Facebook has three. Customers who shop on Yelp or Facebook before calling don't find you at all — even though 233 five-star reviews would crush whatever's there today. Your review signal is trapped on one platform.
When a new policy gets bound, a short workflow fires. A Claude agent writes a thank-you note from scratch — not a template fill-in — referencing that customer's actual policy and the specifics of what you did for them ("Gracias María por confiar en nosotros con tu auto y con tu casa. Gracias a tu póliza, esta semana donamos al Lindenhurst Food Pantry"). The note lands on their phone with a one-click review link that routes to whichever platform is thinnest right now — Yelp first, Facebook next, Google as fallback for your longest-tenure customers. One gentle reminder at five days, then stop. No one gets asked twice.
Your Yelp page starts accumulating. Facebook follows. Your 233-review signal stops being trapped on one platform. It's a one-time setup, not a monthly subscription.
Separate from the four agents above. This one is for your own laptops — not a workflow in the background but an assistant sitting next to you while you work.
Anthropic's Claude Max plan ($100/month, shared across your team) gives the three of you a desktop app that acts like an on-call fourth team member you can interrupt whenever. Five small use cases I'd expect to matter in your office on any given day:
$100/month total for the whole team, not per seat. Doesn't require the four-agent build above — you can start this next week, standalone.
So you know what I already walked past before this note landed in your inbox:
Claude Desktop on your laptops (Section 06) plus the weekend SMS acknowledgment (Story one). Neither touches your State Farm customer data. Neither changes the corporate website. Neither sends anything customer-facing without your approval. It's the safest possible starting line — and if it proves out, the other three agents become easier conversations.
If one of the four agents — or the desktop tool — sounds like something you'd actually want, or like something you've already solved, or like it's just not the pain, I'd love to hear it. I build these kinds of agents for a living; what I'd want from a first conversation is to understand where you actually sit, not to sell you a stack you don't need.
One-word answers are entirely fine for any of them: resonates / already doing / not my pain / didn't get it.
No deck, no slides. Thirty minutes, neighbor to neighbor.