Pipelines, Not Prompts

Why we did this

The goal of the first session was not a demo. It was to put a handful of agency owners who are actually running AI inside their agencies in the same room and let them talk to each other. Not a webinar, not a pitch, not a product tour. Whether something runs through a connector or not was beside the point.

What came out of it was better than expected, so here it is. Everything below came from the call itself. Nothing has been added.

The five things worth remembering

1. Do not hand your staff a chat box. The point that opened the call: if you hand your team a connection and say “ask it anything,” almost nobody uses it near its potential. The stronger move is building automations around what the connection can do. Staff end up interacting with it constantly and mostly not realizing it.

2. One portal, one skill. Then roll up. A skill that goes looking for something, carrier notifications every morning for example, can cover every carrier at once. But a defined, repeated data entry workflow into a specific portal gets its own skill. That is not a home quote skill, that is a skill for one carrier. Once you have four, roll them up into a single quoting skill that knows to reference the others, and put your safeguards at the roll-up level.

3. Give the machine its own stages. Pipelines are going to split into human stages and machine stages. A machine stage is one the bot watches on a schedule. Work lands there, it does its job, then it funnels the file into the next human-in-the-loop stage. Two agencies on the call had built this independently. One calls it “staged,” the other calls it “triage.”

4. The machine should ask, not guess. The single most important instruction in a data entry skill is to never fill a field it does not know. If it does not have the roof age, it does not put a roof age. It does not reason its way to one from the year built. It asks.

5. This is robot work, not agency work. Everything discussed was administrative. Data entry, downloading reports, moving documents, updating stages. None of it was the part of the job that requires a licensed human talking to a client. That distinction is the whole point.

What people are running

One agency, the deepest build in the room:

• A non-pay and late-pay pipeline running fully automated. It runs headless every morning, connects to the carrier, downloads the reports the management system does not provide, posts cancellations and reinstatements, enters payment amounts, notifies customers.
• Renewals with a real premium comparison that decides whether a file needs agent or CSR attention, and moves it accordingly.
• An e-signature pipeline that pulls documents out of the customer file, sends them for signature, attaches the signed copy back, uploads to the carrier portal where the carrier requires it, notes that it was uploaded, then pulls trailing documents and sends those too. The first run surfaced 88 signed documents that had never made it onto a customer file.
• Proposals posted to their own website as web pages instead of attached as PDFs. The system builds the proposal, drafts the email, and after approval sends it from the owner’s own account. Then it attaches the carrier quote PDFs to the file, moves the opportunity to Quoted, and fills in carriers and premiums. That is 10 to 12 manual steps gone.
• Email to file matching on a schedule. The management system only auto-files email when the sender address matches the customer on file, so underwriting and excess and surplus mail falls through. Their build searches for a matching policy number or subject line and posts the email into the customer file as a note.
• A browser extension, written by the AI rather than by a developer, that shows a customer’s active policies inside the management system and opens the carrier portal on that customer’s file in one click. Saves eight or nine steps every time.

Another agency: a machine stage for post-renewal cleanup that runs once a day toward the end of the day, around 5:15. If staff filled the custom fields out properly, it creates the cross sale lead, sells the cross sale lead for tracking, creates the onboarding ticket, and closes the renewal ticket. No manual input.

An owner six weeks in: most of what arrives from carriers lands in one main inbox, and a daily sweep classifies it and creates the right task. Plus a fifteen page month-end report covering open tasks, sold business, what is sitting in binding and why it has been there so long, and how many signature requests are outstanding. Her words: information she has never had before.

Another owner, running inside Google Workspace: email, chat, and Drive folders all connected, and more useful than expected. Candid about where he is, which is still working out how to be sure the work actually got done rather than just getting a recommendation about what should be done. Uses a triage stage for anything the bot cannot finish.

Patterns worth stealing

Pattern	What it looks like in practice
A holding stage for machine work	A stage named “staged” or “triage” where the bot parks anything it cannot finish. One carrier does not put the payment amount on a cancellation notice, so those items sit there until a human fills it in. Half-finished work stays visible instead of being silently dropped.
Sanity checks on volume	A carrier report came back with 220 non-pay entries instead of a normal day’s worth, because the carrier had run it against the whole book. The bot flagged it and asked rather than processing it. Build the “does this number look right” check in from day one.
Emoji as an approval interface	When a document cannot be matched to a file, the bot posts to chat. A check mark means proceed, an X means it is already handled. Staff can also type a correction, which kicks off a fresh search and a completion message back. No new software for anyone to learn.
Hardcode the pipeline and stage IDs	The connection caches organizational context daily. Tell your skill exactly which pipeline and stage IDs it owns. They do not change, and it saves the lookups.
Compare the terms, not just the dates	A renewal comparison that actually reads the terms caught a carrier quietly changing roof settlement from replacement cost to actual cash value, including on newer homes. Twenty minutes produced a complete list of affected policies. The carrier itself did not have that list.
Carrier reports beat downloads	Pulling the carrier’s own report daily catches things the management system download misses, particularly non-renewals and underwriting requests.
Never bind, never pay	Rating and quoting work stays supervised and visible. The bot fills things out. It does not bind coverage and it does not take payments.

Logins, MFA and credentials

The biggest practical blocker on the call, and two answers came out of it.

The simple version. Most carrier portals hold a session for about eight hours. Make it part of the morning routine: open the browser your automations use, open the tabs, log in and clear two-factor once. When a scheduled task hits that URL two hours later, it is already authenticated. Nothing to build, and it is the right starting point for most people.

The hands-off version. For anyone who wants it running while they are on vacation:

• Credentials and API keys live in a password manager, accessed outside the browser, so nothing passes through a chat window or gets stored anywhere it should not be. Not every password manager can do this. At least one on the call could not.
• For carriers offering an authenticator code, the password manager generates the six digit code and it goes straight in.
• For carriers that email a code, the code forwards to a separate, locked-down inbox the bot can read. Read only, no send ability. It grabs the code, enters it, and continues.

Worth saying plainly: if you hand an AI a username and password and tell it to go log in somewhere, it will refuse. That is by design, and it is why the credential path has to be built properly.

Tools and costs named on the call

Tool	Context
E-signature platform	No native connector, so API access runs about $600 a year on top of the account. Verdict from the person paying it: a bargain, given what the first run alone turned up.
Comparative rater	API access has not been forthcoming without fees. Two workarounds discussed: drive the browser interface under supervision, or build a skill against the front-end endpoints.
Password manager	The credential and MFA answer. Check whether yours can hand credentials to an automation securely, because not all of them can.
Team chat	The notification and approval layer for two agencies. Also searchable by the AI when the platform’s own search falls short.
Microsoft Teams	The known gap. The AI can read Teams but cannot send outbound messages to it, where it can in Slack. If notifications matter to your build, plan around this.
Separate database	Two people keep commission reconciliation data outside the management system, because the management system is not built for that comparison.

The honest list: what is still broken

Every one of these came up and none of them got solved.

• Teams cannot send outbound messages, which breaks the notification loop that makes chat-based builds work.
• Rater API access is still gated behind fees, and the browser path is clumsy.
• Some management systems do not hold enough policy detail, so the real record lives elsewhere.
• Uploading files into some systems is not exposed through the API, which is why people park cloud storage links in activities instead of real documents.
• Verifying that work actually got done, rather than just getting a recommendation about what should be done, is an open problem for anyone early in this.
• Skills that drive a browser will break if a vendor changes their interface. Endpoints have been stable for close to a year, but that is a when, not an if.
• Carrier notice blind spots are real. Mail sent directly to an insured is invisible to every system any of us run.

One last thing

There is a fair amount of gatekeeping around this topic in our industry right now, with people putting a paywall in front of conversations that ought to be free. That is not what this is. Nobody on that call was selling anything to anybody else on it, and the standing offer from one of them was some version of: take whatever I have built, I am not going to support it, but it is yours.

Keep sharing it. That is the entire point.