Second session. Thirty-some owners on, an hour and a half by the time the last person signed off. No names in here, and nothing anybody said about a specific vendor made it in. What is below is the substance.
The through-line: it is a blend, not a replacement
The framing we opened with was three eras of agency operations.
The sticky note era. Somebody called in, and the best system was a note written down and laid on the right person’s desk. Worth saying plainly: nearly everybody on the call still lives in that world at times. We are human and it is easy.
The if-this-then-that era. Deterministic automation. A lead lands, a text goes out, a task gets created. This has worked well and a lot of agencies grasped it in the last few years.
The judgment era. Goal-based work, where you describe the outcome instead of every step.
The counter-intuitive part, and the thing that generated the most honest conversation of the day: the agencies that did the absolute best at if-this-then-that are often the ones who struggle hardest with the judgment piece. Being excellent at defining every step makes it uncomfortable to hand off a process where you cannot see every step.
But the correction that matters is that era three does not replace era two. Take a lead that lands from a paid ad. A call task fires, a text goes out with a link to a short form, and if there is no reply inside a set window a call gets triggered. That is the same every single time. There is no judgment call anywhere in it, and there is no reason to burn AI tokens on a process that never varies.
What happens after the form comes back is where judgment starts to earn its keep.
The rule underneath it: if the step is identical every time, a rule handles it. The moment it requires reading something and deciding, that is where AI belongs. Both halves run on a schedule. Neither one needs a person babysitting it.
Where people are actually getting stuck
The most useful moment on the call was an owner describing, out loud and without hedging, why this is hard for him. Paraphrased with permission of the ideas, not the person:
That is not a technical objection. It is a trust problem, and it is probably the single most common blocker in the room.
The way out that came up: stop trying to design the path and start by defining the end result, then work backward from there. What has to be true when this is done. What would a human have done. That reframe is what unlocked it for him.
A second owner added the counterweight from the technical side. Insurance is rule driven, but the agent side lives in gray areas, and gray areas need reasoning rather than branching. His point: that is also where it gets scary, because you are taking judgment you exercise daily and assigning it to something else. So you build safeties into your skills, and you always give it a fallback: either a reasonable, explicitly defined assumption, or an actionable task that hands the decision back to a person.
A third and unresolved version of the same problem: several people have automation that already works. Tearing up functioning Zaps to rebuild them with judgment in the loop is genuinely expensive, and nobody on the call had a clean answer for when it is worth it. Open question.
The prerequisite nobody wants to hear
Before any of this, forget the technology entirely.
Drill down to what actually needs to happen. If humans were the only thing involved in the process, what would the workflow look like? Who touches it, what triggers the next step, what does done look like?
Two reasons it matters. First, rules break down because every agency has a niche, and niches force you to layer rules on top of rules until the stack is the problem. Second, and more practical: if the AI cannot read something, it cannot help you with it. Tribal knowledge that lives only in your best CSR’s head does not transfer.
Related, and it came up more than once: an agency running two or three tools is in far better shape than one running fifteen. Fifteen tools is fifteen logins and fifteen expectations you are putting on your team.
Patterns worth stealing
Anonymized. All of these were described by agencies on the call.
A scheduled task that prepares, then hands off. One owner runs two scheduled jobs against his phone system, which exposes a connector. The first watches for new leads, runs a prep skill that looks at what data came in and researches what is missing, drafts a text, and sends it. The second checks later for a response and drafts a follow-up if there was none. The boundary is explicit and worth copying: if somebody actually replies, it stops and a human takes it. No back and forth.
That example got singled out in the wrap as the cleanest illustration of the whole point. It is if-this-then-that on the front end and an intelligent, situation-specific message on the back end.
Gap analysis before rating. Another agency is building a flow where the system holds a list of the minimum data required before a quote can proceed. A lead arrives, it checks against that list, and branches. Complete, it goes to the rater. Incomplete, it reads the customer’s stated contact preference. Phone preference routes to a human immediately. Text or email gets a specific, personal request naming exactly what is missing rather than a generic “we need more info.” No response, it drips or asks again.
Document ingestion with a human in the loop. One agency ingests driver’s licenses and dec pages, extracts the data, lands it in a form, and uses automation to fill the management system screens. Critically, a person reviews and corrects before anything becomes a bindable quote. His own honest read: not fully accurate, not fully automated, but far better than before.
Comparison tables customers actually read. Same agency built a premium comparison output. Customers respond to seeing carriers, premiums, and lines laid out side by side rather than reading three PDFs.
Bulk quote summarization. A commercial account with eight to ten quotes across admitted and E&S carriers, summarized into a single readable table in about two minutes by uploading the documents and prompting for it. Customer name and contact details were scrubbed by hand first.
Auditing the pipeline instead of eyeballing it. We ran an audit against our own paid lead pipeline this week. It read lead notes, phone transcriptions, and texts sent across every stage, and produced who needs a second follow-up, who needs a third, and which quotes are half finished. It found roughly fifty leads that were not being worked properly. We paused the ads until the process could keep up. Not a flattering finding, but it is exactly the kind of thing a person scrolling a board does not catch.
Building skills that hold up
The most practically useful stretch of the call.
Expect two to four iterations. Nothing complex works on the first pass. One rater skill took four rounds before it was dialed in. That is normal, not failure.
Build a self-repair mechanism into the skill. At the end of a test run, have it report back on how the skill itself should be improved. Then feed that report back in and update it.
Worked example: a rater asks whether the insured smokes in the house. The agency never collects that. The skill’s guardrail says if you do not know an answer, you cannot guess it, so you stop. So it hung on that question every single run. It reported the problem itself: this is an issue I hit every time, do we need to fix this. That is the mechanism working.
Start with agency knowledge skills, not portal skills. The advised order:
- A brand voice skill
- An SOP skill
- Some form of operational skill
All independent of each other, then referenced from your organizational prompting. Those three are the foundation everything else stands on.
One portal, one skill. A skill can do several things. But if it has to drive two different portals, that is two skills. Then a coordinating skill sequences them. One agency described five separate skills feeding a quote orchestrator that knows the order to run them in.
Skills are not portable. Nobody can hand you theirs and have it work. They encode your carriers, your stages, your rules. What transfers is the pattern, not the file.
Brand voice, and sounding like yourself
This ran longer than expected and generated the most cross-talk of any topic.
Build it from many real examples of how you actually write, and put it at the organizational level so everything inherits it.
Bake your SOPs into it, not just tone. If your standard is that an endorsement text always includes the last four of the VIN of the vehicle that changed, the voice skill should know to include it. The voice and the process live together.
Do not connect it straight to the client yet. At least one agency is having it draft and a human paste, deliberately, while they verify it stays relevant and sounds right. That is the correct order of operations.
Two smaller notes with more teeth than they look:
- Over-correct polish is now the tell. Sounding very proper reads as AI. That was true before AI and it is worse now.
- Nobody should be forced into stiff writing to look professional. The way your team actually talks is the asset.
If you want to build one, the shortcut that got shared: find a good writeup of what a brand voice profile should contain, open a fresh chat, and say the goal is to understand this content and ultimately build a brand voice skill, walk me through it. Then paste the whole thing underneath.
Rolling it out to a team
From the operator side of the room, and the part most technical conversations skip.
Ask your team what is most aggravating. Not what should be automated. What is the most annoying, most repetitive thing they do all day. That list is your roadmap and it is more accurate than yours.
Know how each person learns. One agency runs a personality assessment on each team member so they know who needs a diagram and who needs written detail. Same process, two different trainings.
Check in weekly. A standing pain-point conversation, owned by somebody specific.
The line that framed it: it does no good to automate and map a process if the team does not understand it.
Also worth stealing: one of the earliest and highest-return automations any agency described was simply recording and transcribing phone calls into the management system, so nobody has to rekey a call.
Security, and where the line sits
One participant raised it directly. He scrubs customer names and contact details by hand before uploading anything to a model, and named the discomfort plainly: the disclaimers say your data will not be used, and you are still trusting a company.
The answer given, and it was not a comfortable one:
You have to make your own conscious decision about what works for your agency. Read the actual privacy terms of what you are using. Commercial and business plans generally carry contractual language that consumer and free tiers do not, and that distinction is worth understanding rather than assuming.
A workable personal threshold, offered as one person’s line and not as advice: if the information is on your website, or on a public business directory, or two searches away, it is low risk. Names, business phones, business emails. Driver’s license numbers and social security numbers are a hard line. That is where real exposure starts.
And a forward-looking note: this is moving fast enough that contractual terms are going to change and AI clauses are going to start appearing in agreements where they did not exist before. Worth watching.
One gap we should own. In an hour and a half, nobody raised credentials, multi-factor, or how a browser-driving skill authenticates into a carrier portal or a rater. Several of the workflows described in this session necessarily run inside logged-in systems. That is a real topic and it did not get discussed. Putting it on the list.
The honest list
Things raised and not solved.
- Tearing up automation that already works, in order to rebuild it with judgment in the loop. Nobody has a clean rule for when that is worth it.
- Consistency. More than one person reported building something that worked and then not getting the same result every time.
- The trigger layer. Knowing what you want a skill to do is easier than deciding what should wake it up.
- Credentials and authentication for skills that drive browsers. Not discussed at all, and it should have been.
- Two-way conversational AI with customers. Actively avoided by most of the room rather than solved.
- Getting the quiet majority to talk. Two direct invitations for beginners to speak up got silence both times, and the two people who eventually asked the best beginner questions did it near the end, after the room had thinned out. That is on the format, not on them.
What is next
Next session is Thursday at 11:00 Central, same link.
Topics are coming from what people wrote on the signup form and sent by email, so if there is something you want in the room, say so.
Two things already on the list: more concrete worked examples of a process moving from if-this-then-that to an intelligent response, because the abstract version of that argument is clearly not landing on its own. And a walkthrough of a full lead-to-proposal flow, if it is proven by then. Not before. Anything that is not proven is a lie, and there is no sense talking through something until we know it works.
One standing offer: if something in here sounds like a conversation you want to have with the person who described it, say the word and introductions get made.
Nobody is selling anything. Keep sharing it.