AI Calling Mistakes Insurance Agencies Make in 2026

An agency in your position turns on an AI caller, watches it book three appointments in the first afternoon, and feels good about the decision. Two weeks later, someone finally checks the calendar those appointments landed on and finds a handful of slots nobody confirmed, a caller who asked to stop being called and got called again anyway, and a number that’s started going straight to voicemail because it picked up a spam label nobody was watching for. None of that happened because the AI didn’t work. It happened because the parts around the AI, the CRM sync, the consent check, the number hygiene, weren’t built before the first call went out. This piece is a straight list of the mistakes that cause that, in the order agencies actually hit them, with what fixes each one.

Two-thirds of independent insurance agencies plan to increase their use of AI in the next 12 months, according to the 2026 Big “I” Agents Council for Technology Tech Trends Report, a national survey of Big “I” member agencies released February 19, 2026. Adoption is moving fast. The same report found that 55% of those agencies have no written policy governing how AI gets used, and that data privacy or compliance risk is the single most-cited concern, ahead of cost, ahead of accuracy, ahead of everything else. The gap between “we’re adding this” and “we’ve thought through how” is exactly where the mistakes below live.

Why this is happening now, and why it’s mostly a setup problem

AI calling tools got cheap and easy to turn on faster than most agencies built the internal process to turn one on correctly. That’s the pattern the Big “I” data shows plainly: 38% of surveyed agencies said they’re very likely to increase AI use this year, another 30% somewhat likely, against a current adoption picture where 33% describe themselves as just experimenting, 22% using it in limited areas, and only 8% with AI embedded in daily workflows. Most agencies are early, moving fast, and doing it without a document that says who’s allowed to turn a new tool on or what it’s allowed to say.

Read those four numbers together and the shape of the problem is obvious: nearly a third of agencies haven’t started, another third are experimenting without a settled process yet, and only 8% have reached the point where AI is a routine, governed part of daily operations. Most of the industry is sitting in the two middle stages right now, exactly where the mistakes in this article do the most damage, because the tool is live enough to matter but the process around it hasn’t caught up.

None of the mistakes in this article are about the underlying technology failing to work. A voice pipeline that recognizes speech, decides what to say, and speaks a response is a solved problem at this point; we’ve written separately about how that pipeline actually works if you want the mechanics. What breaks is everything wired around it: whether a human gets told when the AI hands them something, whether the list being dialed has valid consent behind it, whether a transfer actually lands on a person, and whether anyone wrote down what the AI is and isn’t allowed to do before it started doing it. Every mistake below is an integration or a policy gap, not a technology limitation, which is also why every one of them is fixable without switching platforms.

Mistake 1: Turning it on with no written policy

Nobody at most agencies has written down who’s allowed to enable a new AI tool, what it’s permitted to say to a caller, when it has to stop and hand off to a licensed human, and how consent gets checked before a number gets dialed. The Big “I” report puts a number on how common that gap is: 55% of independent agencies have no written AI policy, 23% have one under development, and just 13% have a formal policy in place. Separately, when the same survey asked what worries agencies most about AI, data privacy or compliance risk topped the list at 24%, ahead of inaccurate outputs at 22%, losing human interaction at 17%, and not knowing how to apply it effectively at 16%.

That’s a strange combination to sit with: compliance risk is the top concern, and more than half the industry hasn’t written down the policy that would actually manage it. A written policy doesn’t need to be a legal document. It needs to answer four questions plainly enough that a new hire could read it and know what to do: who can turn on a new automated calling tool, what scripts or topics require a human, what counts as valid consent before a number gets dialed, and what the escalation path is when a call goes sideways. Write that down before the next tool gets enabled, not after something goes wrong with the one that’s already running.

Mistake 2: Automating a list you haven’t checked for consent

An AI caller dials faster and more consistently than a person working the same list by hand, which means a bad list gets worked all the way through before anyone notices the problem, instead of getting caught three calls in when a staffer starts wondering why half the numbers are dead. Automation doesn’t create a consent problem. It removes the slow, uneven pace that used to buy time to catch one.

The relevant standard hasn’t changed: automated or artificial-voice calls to a wireless number require prior express consent under the TCPA, and marketing calls specifically require prior express written consent, per the FCC’s February 2024 ruling confirming that AI-generated voices count as “artificial voice” for this purpose. What has changed is how many agencies are sitting on aged or resold lists where the original consent language is vague, missing, or was captured through a lead source that’s since shut down. Before pointing an AI caller at any list that isn’t fresh, pull the actual consent record for a sample of it. If you can’t produce what was disclosed and what box was checked for a given lead, that lead doesn’t go to automated dialing; route it to a live, manually-dialed call instead, which isn’t subject to the same autodialer and artificial-voice consent standard. We’ve written a full breakdown of what changed with aged lead consent and the Reassigned Numbers Database in our piece on aged insurance leads if your database needs that level of detail.

Mistake 3: A warm transfer with nobody actually staffed to take it

This is the mistake that produces the caller stuck in dead air. An AI caller qualifies a lead, decides they’re ready to talk to a person, and tries to connect them, and if nobody’s actually available on the other end, the caller either sits on hold or the call drops. This is the same failure mode the Telemarketing Sales Rule was written to prevent for predictive dialers, and it applies just as directly to an AI transfer.

Under 16 CFR 310.4(b)(4), a call is considered abandoned if it isn’t connected to a live sales representative within two seconds of the person’s completed greeting, and a telemarketer’s abandonment rate across a campaign, or a rolling 30-day period, can’t exceed 3% of all calls answered by a person. The rule includes a safe harbor: allow the phone to ring at least 15 seconds or four rings before disconnecting an unanswered call, and if a rep isn’t available within that 2-second window, play a required recorded message rather than leaving dead air. None of that changes because a machine, not a person, initiated the qualifying conversation. If your AI caller’s warm-transfer step routes to a queue that’s frequently unstaffed, you’re running the exact scenario this rule targets, whether you built the dialer yourself or bought it from a vendor.

The fix isn’t a smarter AI. It’s making sure the transfer target is genuinely available, whether that’s a live agent, a round-robin group with real people in rotation, or a clearly defined fallback (book a callback slot, don’t leave the caller hanging) for when nobody’s free. We covered the mechanics of building this correctly in GoHighLevel in our workflow piece on trigger-to-warm-transfer builds, and the broader case for warm transfers over other follow-up methods in our comparison of warm transfers against voicemail drops.

Mistake 4: Booking the appointment, telling no one

This is the single most common version of “the handoff dies.” The AI does exactly what it was built to do, qualifies a caller and books them on a calendar or logs a disposition, and the workflow stops there because nothing was configured to tell a human it happened. The lead exists in the system. Nobody’s looking at it.

HighLevel’s own documentation for its Contact Created trigger describes the mechanic plainly: a trigger fires “regardless of how the contact is created, manually, via form submissions, or through integrations,” and from there a workflow can send a notification, assign a user, or update a field, but only if something’s actually chained to that trigger. The gap agencies hit isn’t that the trigger doesn’t fire. It’s that the automation stops at “assign the lead” and never adds the step that makes a specific person’s phone buzz. A contact sitting unassigned, or assigned to someone who was never actually told, produces the exact same outcome as no AI caller at all: a lead that goes cold while everyone assumes someone else has it.

The fix is a five-minute check, not a rebuild: open the workflow that handles a booked appointment or a completed AI call, and confirm there’s an internal notification action, addressed to a specific person or a round-robin group, chained after it. If the workflow ends at “update contact record” and nothing else, that’s the gap. Pair that with a native sync between the AI caller and your CRM, so the transcript, disposition, and outcome land automatically instead of depending on someone remembering to log it by hand.

Mistake 5: Ignoring the number until it’s already flagged

An agency that ramps outbound call volume overnight on a number that’s never made a call before is doing the exact thing that gets numbers flagged. Carriers watch for a sudden jump in outbound volume from a previously quiet number, high rates of unanswered or quickly-hung-up calls, and complaint reports, and none of that cares whether a human or an AI is dialing. Once a number picks up a “Spam Likely” or “Scam Likely” label, pickups drop off a cliff and nobody tells you it happened; the connect rate just quietly falls and stays there until someone investigates.

Number warmup, gradually ramping volume on a new number over days or weeks rather than going from zero to full volume on day one, and ongoing spam monitoring aren’t optional extras on an AI calling setup. They’re the difference between a number that keeps working and one that silently stops. We go deep on how carriers make this determination and the specific warmup routine that keeps a number clean in our piece on number warmup and spam labels; the short version for this article is that if your AI vendor doesn’t mention warmup and ongoing spam defense as a standing part of the service, ask what happens to your connect rate in month two.

Mistake 6: Assuming the AI vendor now owns compliance

This is the mistake with the most expensive downside. An agency turns on an AI caller and, somewhere in that decision, starts treating “the AI handles it” as a substitute for “we’ve confirmed this is compliant.” It isn’t, and no vendor’s marketing copy changes the legal answer.

The FCC settled the core question in its February 2024 declaratory ruling, FCC 24-17: AI-generated and cloned voices fall within the TCPA’s definition of “artificial or prerecorded voice,” which means the same consent, identification, and opt-out requirements that have always applied to automated calling apply in full to an AI caller. A violation, an unconsented call, a missed opt-out, a disclosure that didn’t happen, carries the same exposure it always has: under 47 U.S.C. § 227(b)(3), a consumer can recover actual damages or $500 per violation, whichever is greater, and a court can raise that to as much as $1,500 per violation for a willful or knowing violation. That exposure sits with the licensed agent and agency, not with whoever built the AI platform.

Medicare adds a specific, separate layer on top of general TCPA compliance. Under 42 CFR 422.2274(g)(2)(ii), marketing and sales calls touching Medicare Advantage or Part D have to be recorded in their entirety and retained for a minimum of six years, with the first three years specifically in audio format, not just a transcript. An AI calling platform built generically for restaurants or home services almost certainly isn’t configured to retain recordings that long in that format by default. If your lead flow includes any Medicare-touching contacts, that’s a setting to confirm before the first call, not a gap to discover during an audit.

If you want to hear how a platform actually handles disclosure and opt-out in a live call before you commit to anything, there’s a demo call on the homepage that walks through exactly that. https://theaffordableai.com/

Mistake 7: No plan for the caller the AI shouldn’t handle

A caller in the middle of a claims dispute, someone who’s clearly upset about a rate increase or a denial, or a question that requires actual coverage advice from a licensed producer, all call for a human who can read tone and use judgment, not a system optimized for working through a qualifying script. An AI caller that pushes forward with its script regardless of what it’s hearing turns a bad moment into a worse one, and it’s a completely foreseeable failure if nobody designed for it.

The fix is a defined escalation path, built and tested before launch: what specific signals should trigger an immediate handoff (explicit distress, a request to speak to a person, a question outside the AI’s scope), what happens when a human is available, and what happens when one isn’t. That last part matters as much as the first two. If a vendor you’re evaluating can’t describe concretely what their system does when a distressed caller can’t reach a person right away, that’s a question worth pushing on before your agency’s name is the one on the call.

What getting this wrong actually costs

The numbers involved aren’t abstract. A single TCPA violation carries $500 to $1,500 of statutory exposure per call under 47 U.S.C. § 227(b)(3), and a list dialed at AI-caller volume without a consent check can generate that exposure many times over before anyone notices the underlying problem. A flagged number that silently stops connecting doesn’t show up as a line item anywhere; it shows up as a connect rate that quietly drops and a pipeline that quietly thins out. A booked appointment nobody was told about costs exactly what the lead itself cost to acquire, plus whatever that policy would have been worth, for nothing in return.

This single Big “I” survey is the clearest current data on how independent agencies are actually approaching AI adoption, and it’s worth being direct that it’s one organization’s self-reported national survey without a second, independently-run dataset measuring the identical question. Treat the specific percentages as a snapshot of one trade association’s membership, not a universal industry constant, while the underlying pattern, fast adoption outpacing written policy, is consistent with what shows up anecdotally across insurance forums and vendor case studies more broadly.

How we build around these seven mistakes

TheAffordableAI is built so most of this list isn’t something an agency has to remember to configure correctly; it’s the default. Warm transfers check real-time calendar and agent availability before attempting the handoff, rather than blind-transferring into a queue that might be empty. Every call, transcript, and disposition syncs natively to HighLevel the moment it happens, so the notification gap in mistake four doesn’t depend on a workflow someone built correctly six months ago and never checked again. Number warmup and spam defense run as a standing part of the service, not a one-time setup step, and multi-calendar intent routing sends a booked appointment to whoever’s actually available rather than one person’s calendar by default.

None of that removes the two mistakes that are genuinely the agency’s job: writing the policy, and verifying consent on the list before it gets dialed. No platform can do those for you, and any vendor implying otherwise is worth being skeptical of. Full detail on what’s included on every plan is on the features page.

Where AI calling is the wrong tool, plainly

It’s worth saying where none of this advice matters because the tool itself is the wrong fit. An agency with a handful of long-tenured client relationships and low call volume doesn’t need a managed AI caller to solve a coverage gap it doesn’t have; the setup and monitoring overhead of doing this correctly costs more than the problem is worth at that scale. A book of business built on clients who expect to reach someone who already knows their history loses something real if the first voice they hear on every call is automated, even briefly, even when it’s configured perfectly.

And if your actual constraint is lead quality rather than follow-up speed or coverage, no amount of AI calling fixes that. A faster, better-covered path to a bad lead is still a bad lead. Diagnose which problem you actually have before buying a tool built to solve a different one.

The seven mistakes and the fix for each, side by side

The audit to run before you change anything

Before adding, replacing, or expanding an AI caller, walk through the seven mistakes above against your current setup, in order: is there a written policy, has the list being dialed been checked for consent, is the warm-transfer target actually staffed, does a booked appointment trigger a notification to a specific person, is the outbound number on a warmup and monitoring routine, does everyone understand that compliance liability stays with the agency regardless of what tool is calling, and is there a tested plan for the caller the AI shouldn’t handle alone. Most agencies that run through this list find they’ve solved four or five of the seven already and have a genuine, specific gap on the rest. That’s a far more useful starting point than a general sense that “the AI thing” is or isn’t working.