Seamlessly Onboarding AI Into Debt Collection

20 million conversations in, the pattern is already clear.

Gail's conversational AI agents have spoken with more than 20 million people across voice, SMS, webchat, and WhatsApp, handling millions of sales and service calls, including the ones that land at 4 a.m., and collecting more than $500 million in late payments along the way.

The pattern across that volume is consistent: AI agents can do the core work of a traditional collections agent for roughly 20% of the cost, and in some workflows, operate up to 3x more efficiently.

None of that is what's holding the industry back.

The hesitation was never about whether AI works.

It's about sequencing. How do you test a new channel, a new vendor, or a new script without pausing an operation that's already recovering revenue today? How do you tell which of the dozen voice and SMS vendors at the conference actually integrates with your servicing system, versus which one asks you to rebuild your stack around it? And how do you keep compliance, IT, and collections operations aligned while you find out?

Two paths show up over and over, and both tend to fail.

Build it yourself. In practice this means solving roughly eighteen hard problems at once: model selection, telephony infrastructure, prompt design, guardrails, compliance logic, integrations into the system of record, orchestration, human handoff, monitoring, evaluations, security, and an ongoing improvement loop. Very few collections organizations are staffed for that, and FDCPA, Regulation F, TCPA, and state-level licensing turn a slow, imperfectly audited internal build into a liability, not just a slow project.

Stitch vendors together. Faster to start, with a hidden cost: fragmented data, an inconsistent customer experience across channels, and a compliance surface that now spans four or five vendor contracts. Every script change has to propagate across systems that don't share context. Every audit means reconciling call, text, and payment logs from different sources.

Buying AI today is a lot like buying a car in pieces. Most vendors will happily sell the engine, the wheels, and the dashboard, and leave the agency to build the car. That's a reasonable ask for a company whose core competency is engineering. It's an unreasonable one for a collections operation whose job is recovering receivables, not integrating machine-learning infrastructure on the side.

A framework for onboarding AI in parallel, not instead.

The agencies that get this right don't treat AI as an all-or-nothing migration. They run it the way they'd test any new collections strategy: on a defined slice of the book, alongside the current process, measured against a clear baseline, expanding only once the results earn it.

1. Map & segment.. Pick one queue, not the whole book: a single population like multi-invoice, high-balance commercial accounts, or one region. Map the current workflow end to end, and capture a baseline first: contact rate, promise-to-pay rate, cost per dollar recovered. Without it, there's no way to know later whether the AI channel actually moved the number.

2. Pilot in parallel, not in place of. Run the AI channel alongside the existing team on that one queue. Compliance signs off on scripts before a single call or text goes out. Every other queue keeps running exactly as before, so there's zero operational risk if the pilot underperforms. If a vendor's onboarding plan opens with a systems migration before you've seen a single result, that's a signal to slow down, not a feature to celebrate.

3. Wire the stack, don't replace it. Once the pilot shows something worth acting on, integrate read/write access into the systems already in place (system of record, CRM, payment processor, dialer) instead of moving data into a new platform. Confirm auditability here: every call recorded and retrievable, every decision logged, Do-Not-Call and opt-out suppression enforced automatically rather than reconciled by hand after the fact.

4. Scale & optimize. Expand from the pilot queue to adjacent segments using the same cost-per-resolution math that proved out the pilot. Share of the book should be earned by performance, not allocated by default. Layer channels in sequence: voice first, then SMS with payment links, then a coordinated omnichannel cadence, rather than launching everything at once.

Every phase is reversible, measured against a baseline, and runs beside the operation, never instead of it.

What separates a real partner from a booth demo.

Not every AI vendor at a conference is built for a regulated, high-stakes recovery workflow. Five things are worth checking before signing anything:

• Compliance by design : mini-Miranda, right-party verification, time-of-day and DNC rules, full call recording, and an audit trail built into the agent itself, not bolted on after the first complaint.

• Integration depth : native read/write access to the CRM, system of record, and payment processor, not just an API the agency has to build against on its own.

• True omnichannel : voice, SMS, email, and WhatsApp coordinated from one platform on one cadence, not separate vendors that don't share data or context.

• No forced headcount change : the pilot should prove ROI before the agency adds or cuts staff, not require a reorg just to be tested at all.

• Transparent, auditable outcomes : live dashboards on outcomes, recovery, and sentiment that can go straight to compliance and finance, not a black box the agency has to take on faith.

What this looked like for two agencies that already ran it.

A national waste-management company was recovering only about 8.5% of overdue balances from multi-invoice accounts using a weekly text-only blast, with no scalable way to add outbound calling without adding headcount. Running an AI voice campaign in parallel with its existing process, across 14,000+ past-due accounts over five weeks, produced $979,570 in revenue recovered, 3,113 live conversations, and 775 payment promises, at a campaign cost equal to roughly 1.5% of collections, an estimated 80% savings against hiring the equivalent human capacity.

One of the largest consumer banks in Mexico runs one of the largest consumer-collections programs in the Americas: roughly one million delinquent accounts in active daily treatment, historically served by nine external call centers staffing 6,000–10,000 collectors between them. An AI voice agent entered as a tenth, unproven challenger, measured against the other nine on the same KPIs. Within a 12-week pilot-to-contract cycle, it grew from under 1% of the active book to roughly 25%, on pure performance with no allocation enforced, reaching a 2.9x effective contact rate against the leading incumbent call centers' average, and became the bank's top-ranked collections agent.

The result that mattered in both cases wasn't "an AI agent was deployed." It was that the agent outperformed the incumbents on the metrics that already governed the program, while everything else kept running exactly as before.

The advantage isn't adopting AI first.

It's adopting it without breaking what already works.

The collections agencies that come out ahead over the next few years won't be the ones that moved fastest on AI in the abstract. They'll be the ones that tested it against their own book, their own compliance bar, and their own systems, in parallel, at a pace their operation could absorb, before making it permanent.

That's a sequencing problem as much as it is a technology problem, and it's solvable with a plan rather than a leap of faith.