The reframe
Somebody put it plainly a couple of weeks ago: I have three systems, they do not talk to each other, and I really do not want to duct tape this thing together.
The instinct is to look at the systems and ask how to connect them. That is the second question.
Start with what your agency actually knows and does, more than with what your systems know and do. An AI that has no idea how you operate will do generic work quickly. The valuable part of your agency is not in the software. It is the tribal knowledge stuck in somebody’s head.
The three things to build first
In this order, and none of them require an API, a connector, or an integration.
1. A brand voice skill
If it writes on your behalf and you have not given it any sense of how you communicate, it immediately reads as AI.
What actually goes in one:
- Real examples, not a description. Connect your email first so it can pull from what you have actually sent rather than from a handful of things you pasted in.
- The words you do not use. One agency opens with the first name every single time and never uses “Hello.” Left alone, AI defaults to “Hi.” Teach greetings and salutations explicitly, because that is where the tell shows up first.
- Restraint on what goes in a message. If you are sending a VIN, are you sending the whole thing or the last four? Bake the answer in.
2. An SOP skill
Roughly half the agencies we talk to have written SOPs. If you do not, the Big “I” template for independent agencies is a reasonable place to start, then refine it into something that matches how you really work.
It needs to cover quote flow, claims, service, and your E&O guardrails. That last one is not optional. How does a lead flow in and out of your system, and where are you protecting yourself.
3. One operational skill
Pick your biggest pain point, or your close process. One. Not a suite.
What building one actually feels like
The most useful stretch of the session, and the part people underestimate.
Expect four or five iterations. Nothing complex works on the first pass. That is not a sign you are doing it wrong and it is not a sign the tool is broken.
Run it in dry mode first. Have it report what it would do without doing anything. You find out what it misunderstands before it acts on the misunderstanding.
Make the skill grade itself. Whether it is a voice skill or an operational one, have it judge its own performance at the end of a run and tell you how it should be improved. Then feed that back in.
Have it email you when it hits a gap. If it cannot finish, or it finds something it does not have an answer for, it should tell you rather than guess or stall silently.
The framing that ties it together: treat this like a team member. You would not hand a new hire a process once and never check in. Same thing.
How to split skills up
One skill can serve many carriers. There is no reason to build a separate commission-finder skill per carrier.
The moment a skill has to update information in two different places, split it. Then add a third skill whose only job is knowing how to use the other two. One agency’s lead flow is four or five skills plus an orchestrator, running as a scheduled task.
Where the humans stay
Worth stating plainly, because it came up several times.
Nobody on this call lets AI send communications on their behalf unattended. It drafts, a person reviews, a person sends. The reason is simple: roughly once every ten days, something comes back that should not go out. That rate is low enough to be useful and high enough that you do not skip the review.
And the boundary that matters most: it is not an agent. It does not bind coverage. It does not make coverage decisions. It expedites the part that was always going to suck.
The best conversation: reconciling commissions
Two agencies described completely different approaches to the same problem, live, in the same twenty minutes. Both work.
Approach one, from the sold list. Upload the month’s sold list from the CRM along with every carrier statement. The skill cross-references policy numbers and produces a per-producer breakdown. It surfaces three categories that are otherwise invisible: paid but the carrier name does not match what your system says, in force but unpaid, and future-effective and unpaid. It also catches endorsements you never billed against.
Time to build it: about five minutes. What used to take four or five hours of cross-referencing takes five minutes.
Approach two, from the bank statement. Start with a bank statement filtered to deposits only and hand it over with “I need to reconcile this.” The skill identifies which carriers those deposits came from, then asks for the matching commission statements. Output is every renewal, agency income, and what each producer is owed.
The architectural correction that came from a third agency: reconcile in both directions. Your system against the statements, and the statements against your system. Statements are the closer thing to truth, but they are missing producer data, so you have to pull that from the AMS regardless.
What nobody has automated yet: actually retrieving the statements from carrier portals. Everyone is still downloading them by hand. The recipe that came out of the discussion, for anyone who wants to try it:
- Start with email, since it is already connected. Ask it to dig through your email and find which carriers send you statements. Expect it to take several minutes.
- For each carrier that does not email them, write a separate browser prompt whose job is to find the statement and report back how it found it.
- Paste those reports back into the main conversation, carrier by carrier.
- Then turn that into a skill and schedule it for the third of every month.
Split the hunting and gathering from the actual reconciliation. Two skills, then an orchestrator.
Problems nobody has solved
Customers that look lost and are not. Systems count renewals in flight and rewrites as losses. One owner said outright that he has stopped looking at the number because it gives you a heart attack. What is missing is a routine that runs on a schedule and finds everyone who cancelled with one carrier and has new business with another. Nobody has built it.
Drafts with broken links. If a skill writes an email containing a link back to your own site and saves it as a draft, the mail provider can rewrite that link into a redirect. Recipients get a browser warning, and it is a deliverability risk. One workaround exists: skip the drafts folder, review the message and its content together, then send through the provider’s API instead. Worth knowing that this trades away the drafts-folder review step, which is the thing keeping a human in the loop.
Data quality, everywhere. The quote of the session: nobody’s data in this industry is accurate, not one hundred percent. Every skill anyone described eventually ran into a field the system should have had and did not. Plan for it rather than being surprised by it.
Testing text automations. You can test email. There is no clean way to test SMS before it goes live. The workaround people use is to turn them on after hours and run yourself through the pipeline as a fake lead.
Security, and one rule worth repeating
Never let people working on your behalf run agency business through a personal account. Not a contractor, not a virtual assistant, not a part-time producer. Company work runs in a company account, period.
The reason is not just policy. When everyone authenticates individually, every action carries the name of the person who actually did it. Share one set of credentials and everything in your audit trail comes back as whoever set it up. If something goes wrong six months from now, you want to know which human it was.
Where documents live. On the customer file in your system of record. Not in a folder beside somebody’s email inbox.
Practical notes from the room
- A PDF page counts as an image, and a chat has a ceiling of around a hundred of them. Two commercial dec pages can exhaust a conversation mid-task. Move to a workspace-style session when you hit that, not before, since it costs more.
- Most of what your team does belongs in a normal chat. Reserve the heavier surface for the jobs that need it.
- Plan levels matter more than people expect. On an entry seat you will hit limits doing real work. Most agencies running builds are on a higher tier, and a regular staff member on a team plan should never be running out. If they are, something is wrong upstream.
- A recorded call becomes a new lead in the CRM in about ninety seconds. Roughly six in ten people on the call already record and transcribe their phone calls, which means most of the raw material is already sitting there.
Rolling it out
One thing worth saying to anybody about to hand this to a team: it is going to feel weird to people. That is normal and it is not a reason to slow down, but it is a reason to explain what is changing and why before it shows up in their day.
And the honest version from the front of the room: our own flow is not perfect yet either.
What is next
Next session is Thursday at 11:00 Central, same link.
By request from the room: twenty minutes on building AgencyZoom automations from the start, the blend of static if-this-then-that automations with AI where judgment is actually needed.
That one is going to be interactive on purpose, because everybody sees pipeline structure differently. Some agencies run a single pipeline. Others run several. Both camps were represented in the room and neither one is wrong.
Anything that does not fit gets picked up in the back half of the call.
Nobody is selling anything. Keep sharing it.