We build and run AI texting and AI voice agents for BHPH collections. We also audit those conversations in bulk — hundreds at a time — which means we have an unusually clear view of how they fail. This is that view, including the parts that are inconvenient for us.

If you're being pitched AI collections right now, the useful question isn't "does it work." It's "which specific jobs does it do better than the alternative, and what happens when it doesn't."

Where AI genuinely earns its place

Volume coverage on the delinquency list

This is the real one. A human collector working a full day makes somewhere around 50–70 meaningful call attempts. A four-person desk therefore cannot touch a 500-account past-due list more than about once a week, and that's assuming they do nothing else — no inbound, no paperwork, no repo coordination.

AI doesn't get tired at 3pm and doesn't stop at 5. Every account on the list gets touched on schedule. That's not a marginal improvement in efficiency; it's a different coverage model entirely.

After-hours and weekend availability

BHPH customers call when they're not at work. If nobody is there at 7pm on a Thursday, that call is either abandoned or converted into a callback that may never happen. AI answering at 7pm and taking a payment is straightforwardly better than voicemail.

Payment reminders and promise follow-up

Structured, low-judgment, repetitive, high-volume. Reminding someone of a payment due Friday and following up on a promise made Tuesday are exactly the tasks that burn out human collectors and that AI handles without degradation.

Consistency of disclosures

A human collector on their sixtieth call of the day skips things. Not maliciously — from fatigue. A properly configured bot says the required disclosure the same way on call 600 as on call 1. For anything with a compliance dimension, that consistency has real value.

Where AI is genuinely bad

Hardship conversations

A customer who just lost a job, had a medical event, or is going through a divorce needs a human. Not because a bot can't parse the words — because what that call requires is judgment about what this specific customer, with this specific history, can realistically do. Handing that to a bot is how you lose a recoverable account and earn a complaint.

Negotiation

Anything involving a real trade — restructuring, a deferral, deciding whether to take $200 now against $400 next week — is a judgment call with portfolio consequences. Bots are bad at this and, more importantly, shouldn't be trusted with it. That's why ours only takes a partial or sets a promise inside limits the dealer sets in advance — anything outside them goes to a person.

Anything a customer is emotional about

Anger, confusion, fear about the vehicle. A bot handling an angry customer reliably makes it worse. The correct behavior is a fast, clean handoff to a person — and whether that handoff actually works is one of the most important things to test in any demo.

The failure modes nobody demos

This is the section we'd want to read if we were buying. Every item below is something we have found auditing real production AI calls across the industry.

Name mispronunciation

It sounds trivial. It isn't. A bot that mangles a customer's name in the opening line destroys credibility in about three seconds, and the rest of the call is spent recovering from it. In a BHPH book with substantial Hispanic and other non-Anglo naming, this is a first-class problem, not a rounding error. Ask any vendor how their system handles it and listen for whether the answer is specific.

Phantom promises

The worst one. A bot logs a promise to pay that the customer never actually made — misreading "I'll try" or "maybe Friday" as a commitment. Your pipeline now contains promises that were never given. Your promise-keep rate collapses and you'll blame the customers.

This failure is invisible unless someone listens to calls against the logged outcomes. Most dealers running AI collections have never done that check.

Silent failure on payment posting

The call goes perfectly, the customer pays, and the payment doesn't post to the DMS because an API call failed and nothing surfaced the error. The customer believes they paid. Your system says they didn't. The next call accuses a paying customer of being delinquent.

Ask specifically: what is the posting failure rate, how is a failure detected, and who gets alerted. If a vendor doesn't have a number for that, they aren't monitoring it.

Sounding like an IVR

Customers hang up on things that sound like phone trees. If the opening two seconds read as automated, the call is over regardless of how good the underlying system is. Conversational quality isn't cosmetic — it's the gate on everything else.

Language handling

A large share of BHPH portfolios need genuine Spanish capability. Not a translated script — actual conversational handling, with a Spanish-speaking human available on handoff. Half-measures here are worse than English-only, because they set an expectation the system can't meet.

Questions to ask any AI collections vendor

  1. Can I hear twenty unedited, randomly selected production calls? Not a demo. Not a reel. Random real ones. The answer to this question tells you most of what you need to know.
  2. What's your payment posting failure rate, and how do you detect failures?
  3. How do you verify a logged promise-to-pay actually matches what the customer said?
  4. What triggers a handoff to a human, and what happens if no human is available?
  5. How does the system handle a name it can't pronounce?
  6. Who reviews these calls for quality, how often, and can I see the rubric?
  7. What's your actual right-party contact rate? Be skeptical of anything dramatically above low-teens; real BHPH contact rates are frequently single digits.
The tell: a vendor who answers question 1 with an immediate yes is confident in their production quality. A vendor who offers a curated highlight reel instead has told you something important.

The framing we'd argue for

AI is a capacity tool, not a replacement for judgment. It should absorb the repetitive volume so your people have room for the conversations that need a person — and it should be measured continuously, because the failure modes above are silent by nature.

Any pitch built on "the bot handles collections" is either overselling or hasn't listened to enough of its own calls. The honest version is narrower and more useful: the bot handles the calls that don't need a human, so the humans can take the ones that do.

Here's how we structure that split in practice.

See what this looks like on your portfolio

Fifteen minutes on Teams. We'll look at your delinquency, show you real AI text conversations, and tell you honestly whether we can move your numbers.

Book a walkthrough

← All resources