Leaping AI
AI agents that call and text in multi-day campaigns
396 followers
AI agents that call and text in multi-day campaigns
396 followers
We allow companies that operate in the physical world (home remodeling, roofing, trades) to automate inbound and outbound calling & texting and run multi-threaded campaigns spanning several weeks. Our AI agents can hold 100+ parallel conversations, speak in multiple languages and integrate into any CRM. We help put appointment scheduling, customer service & confirmation calls on autopilot and eliminate missed calls, slow speed to lead and insufficient lead follow ups for our customers.
ClinicFrame
Kevin, the multi-day campaign is the part I would stress test, and it is not the dialing. You covered opt in and STOP over SMS, which is more than most voice products say out loud. What I would check is the spoken revocation.
If someone tells the agent on day 7 to stop calling, that counts even though it never touched your SMS keyword path, and it has to end the remaining calls and the texts. I work in a regulated space where consent is tracked per channel, and the failure always has the same shape: the person revokes on the channel in front of them and the sequence keeps running on the other one.
Does a spoken stop write to the same DNC record as a texted STOP, and does it close the whole campaign or only the call leg?
Leaping AI
@clemente_lopez1 Yes: one DNC pool, channel-agnostic. A spoken stop writes to the same record as a texted STOP, and once the number is on the list it suppresses every further attempt across all channels for that customer, so it closes the whole campaign, not just the call leg. It also works from any direction: if they call into an unrelated inbound agent and ask to stop, that agent can add them and the outbound sequence ends too. That's how we make sure nothing falls through the cracks.
Refocus
The horizontal to home remodeling niche-down is a smart call, the physical-world trades are so underserved by voice AI. I work on the consumer side (daily check-in calls to aging parents), and the thing that quietly kills outbound for us is carrier spam labeling: even a wanted, friendly call gets flagged "Scam Likely" and never picked up. With 100+ parallel outbound calls, how are you handling number reputation and STIR/SHAKEN attestation so calls actually connect? Would love to hear what held up at scale.
Leaping AI
@igorgurovich We run our numbers through a third-party reputation service that tracks number health and handles remediation, so we see a number degrading before connect rates drop and can rest it instead of burning it. You'd want to keep the number of dials per day per number low (usually below 100 dials/day is optimal). 100+ parallel calls is fine as long as no single number carries it.
The "similar to HubSpot sequencing but for calls/texts" framing makes total sense, and the channel-agnostic DNC record (spoken stop = texted STOP) is the kind of detail most teams skip until it bites them. We deal with a version of this on the WhatsApp/email side, once someone's mid-conversation with our AI and a human needs to step in, we keep the same thread going so there's no "starting over" moment.
Curious how the handoff feels on your end when a homeowner insists on a real person mid-call, does the human rep see the full call history instantly, or is there a summary step first?
Leaping AI
@mittalpatel There's a summary step. We condense the conversation down to what matters: enough that the human never has to re-ask something the homeowner already said, but not so much that they're hunting for the important pieces while someone's waiting on the line. Full transcript is there if they want it, but the summary is what they land on.
That's interesting. When a homeowner insists on a real person, how does the handoff actually work?
Leaping AI
@dhiraj_patel5 The agent picks up on the request and transfers the call to an available rep, along with a short summary of what's been discussed – so the rep comes in knowing who they're talking to and what the homeowner already said, instead of starting from zero. If nobody's available, it doesn't dead-end: the request gets logged and a callback scheduled rather than leaving the person stuck with a bot.
multi-day is the part that sounds small and is not. a campaign running for weeks means the agent's picture of a contact has to survive between sessions, and every stale field is a chance to say something that was true last tuesday.
the thing i would want measured is the gap between "we placed the call" and "a human heard it". we work on an adjacent problem, confirming a form submission actually registered on someone else's system, and the honest version needed independent confirmation rather than our own send log. voicemail, a carrier drop, and someone who hung up in two seconds all look like a completed attempt from the sender's side.
how are you counting those?
the reactivation angle is the one that would worry me most honestly. an old lead who didn't buy could mean "not ready yet, try again later" or it could mean someone who told a rep months ago they weren't interested and that just never got logged properly. both look identical in a CRM as "no purchase, no explicit opt-out on file." if the campaign resurfaces someone who already said no once, even politely, that's a much worse experience than a missed call. how do you handle that gap between what's actually in the DNC list versus what a homeowner remembers telling someone on the phone six months ago?
The multi-day part is the interesting bit, and I think it is where this gets genuinely harder than single-call voice AI, because "did it work" stops being a per-call question and starts compounding.
We run an AI that answers and places phone calls, including recurring scheduled ones, which is a much simpler version of what you are describing. Even there, the thing that bit us was state carried between attempts. A call that half-succeeded is worse than one that failed outright: the next attempt has to know what was already said, or the agent reintroduces itself to someone it spoke to yesterday. That single failure did more damage with our users than any outright error we shipped, because an outright failure reads as software being software, and re-introducing yourself reads as nobody being home.
What held for us was making the terminal state explicit and enumerable instead of inferred. The agent emits a structured outcome, never prose, and something dumber and more literal decides whether the campaign advances. "Stopped because the person said no" and "stopped because we ran out of attempts" have to be different states, because one should never be retried and the other should. If those collapse into a single "unsuccessful," a multi-week campaign will eventually call someone who already declined, and in the trades that is the call that gets you a bad review rather than a missed lead.
Given you span calls and texts, I would guess the same question shows up across channels: does a text that went unanswered count as a touch the next call should acknowledge, or as nothing happened? How do you represent what a campaign already knows about a contact between touches?