AI collections software guide from Intelligent Contacts

A working paper analyzing 22 million debt-collection cases at a Chinese consumer finance company found that AI callers collected 9% less repayment value than human agents during the first 30 days after delinquency and remained 5% behind one year later, according to Yale School of Management Insights. That result reframes the buying question for Intelligent Contacts readers and every other regulated operator. Dialing speed isn't the test. Long-horizon repayment, broken promises, escalation quality, and the audit trail are.

What AI collections software does today

AI collections software is the system that decides who gets contacted, when, and through which channel, places and connects calls, takes payments through self-service, and records every interaction so a reviewer can rebuild it later. Intelligent Contacts (IC) is a contact center and payments platform built for collections, healthcare revenue cycle, and financial services teams. IC serves healthcare, financial services, and accounts receivable organizations, and Agentic Collections is a product offered by IC.

Scale explains why small workflow changes matter. The Consumer Financial Protection Bureau (CFPB) 2024 Fair Debt Collection Practices Act (FDCPA) report reported that the third-party debt-collection market was worth approximately $20.2 billion and employed about 140,000 people, and that nearly one in five people with a credit report had at least one collections tradeline as of the first quarter of 2023. A one-point change in contact policy applied across that volume moves real money in either direction, which is why the record each interaction leaves matters as much as the contact itself.

Core capabilities and where people still decide

Grace AI, IC's collection agent, offers up to 60% faster collection cycles within configured compliance policy. Read that figure against the boundary in the table. A faster cycle that produces promises breaking thirty days later is a loss, so the number worth requesting alongside any speed claim is the kept-promise rate on accounts the agent handled end to end, plus the escalation reason on the ones it passed to staff.

Practical rule: List the capabilities an operation needs, mark which decisions must stay with staff, and finish that list before requesting a single demo.

A vendor that can't map its product to that list is selling activity. A vendor that can map each capability to an owner, a policy limit, and a record a reviewer can rebuild is selling an operating system.

How AI collection agents hand off to people

Consider an illustrative portfolio of 1,000 accounts. One account carries a hypothetical $240 medical balance. The account receives a text first, then an agent call within configured contact limits and during permitted hours. The order inside the call is where quality shows up.

One account through the call

Identity verification comes before any balance is read. Once identity is confirmed, the agent states the balance only from verified account data. It doesn't estimate, round from memory, or fill a gap with plausible language. That same discipline governs which records the platform writes, and it's what makes a handoff reviewable later.

The agent then offers a payment option within configured policy. If the person agrees, the system records a promise with its amount and due date. If the person can't afford that option, the agent presents a clear alternative in plain words and confirms the amount before capture. Pressing for a larger commitment the person doesn't understand creates a missed payment and, often, a complaint.

That payment journey matters as much as persuasion. A debtor who can't understand the options, verify the balance safely, or afford the plan isn't a conversion problem. That account needs a clearer explanation, a workable plan, or a person.

Handoff triggers that should fire without debate

• Dispute raised: The person disputes the debt, the amount, or ownership of the account.

• Live agent requested: The person asks for a person, regardless of how the request is phrased.

• Hardship mentioned: Job loss, illness, or inability to pay requires a trained conversation, not a script loop.

• Validation requested: The person asks for validation or supporting documentation.

• Repeated broken commitments: Two missed commitments in a row signal that another automated promise won't fix the problem.

The last trigger is the one vendors argue about most. Automating a third promise attempt on a debtor who already missed two doesn't raise recovery; it stretches the timeline and writes a worse record. Long-horizon repayment is the honest test here. A workable plan kept for six months beats three fast promises that each collapse within a week.

The up to 60% faster cycle figure IC reports for Grace only holds if the handoff boundary keeps disputes and hardship cases out of the automated lane rather than absorbing them. Teams configuring that boundary can review how to create an AI agent with defined escalation paths before they set triggers in production.

At handoff, the live agent should receive the full transcript, the policy trigger that fired, the account state, and the next eligible contact date. The debtor shouldn't have to repeat the story. A cold transfer tells the person the operation values automation more than resolution, and it tells quality reviewers that the record is incomplete.

The demo test is simple. Ask the vendor to show the handoff log for a disputed account. A credible log shows what was said, which trigger fired, when the transfer happened, what the receiving agent saw, and how the account state changed afterward. A dashboard screenshot without that chain isn't evidence. A promise made and broken three transfers ago is still part of the account's story, and the record should show it.

Predictive dialing and speech analytics within call limits

Predictive dialing places more calls than there are available agents because some calls won't connect. When a person answers, the system connects that call to a free agent. The trade-off is direct. Aggressive pacing raises agent talk time, but it also raises the risk of abandoned calls and poor contact patterns. Pacing thresholds are therefore an operations control and a compliance control, not a tuning detail.

The IC platform includes predictive dialers, self-service portals, AI tools, and reporting dashboards. Those parts only work together if the dialer respects the contact ledger before it places the next call.

The Regulation F test the dialer must pass

Under the CFPB Regulation F debt collection rule FAQs, a debt collector is presumed to violate the FDCPA prohibition on repeated or continuous calls when it places more than seven telephone calls to a person about a particular debt within seven consecutive calendar days, or places another call during the seven-day period after speaking with that person about the debt. The seven-day period begins on the date of the completed telephone conversation.

That rule isn't a universal daily cap. It creates rebuttable presumptions, includes exclusions, and applies to a particular debt and person. Calls made with prior consent, calls that don't connect, and calls to certain professionals may be excluded from the calculation.

A compliant platform must therefore count per debt and per person, distinguish connected calls from unconnected calls, detect a completed conversation, and suppress further calls after live contact. A single counter at campaign level can't do that work. Neither can a dialer that treats disposition codes as an afterthought.

A count alone isn't protection. Even seven calls on the same day can create risk when the pattern and purpose look harassing.

What speech analytics should catch

Speech analytics earns its place after connection. It should detect promises to pay, objections, hardship language, disputes, and required disclosures on connected calls. Flagged calls then move to quality review with the relevant transcript segment attached, so a supervisor doesn't have to listen to an entire recording to find the problem.

Three vendor questions belong in writing:

• Pacing adjustment: How does pacing change when connect rates, agent availability, or abandon risk shifts during a campaign?

• Connect detection: How is a connected call, and a completed conversation, measured and recorded?

• Suppression testing: How is suppression logic tested before a campaign goes live, and who signs off on the test?

Vague answers here become audit findings later.

Compliance controls and the records to demand

An auditor doesn't start with a policy binder. An auditor asks for the record behind one account. The control matters only if it produces that record on demand.

Controls and the records that prove them

The call-frequency presumptions cover telephone calls. The broader ban on harassing, oppressive, or abusive conduct reaches every channel. Voice, SMS, and email therefore need one shared contact counter and one opt-out list. Separate queues that can't see each other will eventually produce a contact pattern nobody intended and nobody can reconstruct quickly.

Verified-data messaging is the control behind every automated statement. The record should keep the script version and the account fields each message drew from, so no agent or model invents a balance, fee, or legal status. Operators reviewing how AI agents keep collections work compliant at scale should still test that claim against a sample message and its source fields.

Retention also needs a wider definition than a final case note. Prompts, transcripts, recordings, scripts, and model outputs are the evidence that explains why an automated contact said what it said. Without them, a complaint investigation becomes a reconstruction exercise built on memory.

For payments and healthcare balances, buyers shouldn't accept labels without mechanics. Built to meet PCI-DSS standards should mean encryption, access logs, role permissions, and audit trails at each payment step. HIPAA-ready infrastructure should show who accessed patient balance information, when, and under which role.

Human review isn't a sufficient control when the reviewer can't see the model's trigger, confidence, policy basis, and complete interaction history.

Governance needs the same discipline. Document intended and prohibited uses for each model, test accuracy and disparate impact by borrower segment, monitor drift, and define rollback and escalation controls under the NIST AI Risk Management Framework. Canadian operations fall under a different rulebook, so federal and provincial requirements should be confirmed with counsel against primary sources.

The closing request for any vendor is a sample audit export for one disputed debt. If that export can't be produced during evaluation, it won't appear under examination pressure.

Outcome measures that show whether it pays off

Contact rate is the number vendors quote first because automation makes it move. Calls dialed and messages sent move for the same reason. They prove activity, not repayment.

Adoption doesn't prove results either. TransUnion's 2024 Debt Collection Industry Survey found that the share of debt collection companies investing in AI or machine learning rose from 11% in 2023 to 18% in 2024. More buyers in the market says nothing about whether a specific workflow collects more over the full repayment period.

Outcome metrics and how each one can mislead

Metrics need partners. A high promise rate beside a high broken-promise rate means commitments are easy to win and easy to break. Strong early recovery beside weak plan completion means cash arrived fast, then the account failed. Low complaints beside slow opt-out latency doesn't reassure a compliance reviewer. It raises a question about whether problems are being captured.

The baseline method should be set before automation starts. For example, pull 90 days of pre-automation data, then compare accounts at the same age after launch. Comparing fresh placements under automation with older accounts under the prior process creates a false lift or a false decline.

Leaders refining their dashboard can compare these measures with contact center KPIs that leaders track beyond activity counts, then keep only the measures tied to repayment and consumer outcomes.

Every vendor demo should end with one demand. Show recovery, broken promises, and complaints on the same screen, over the full repayment period, for the same account cohort. A vendor that separates those views is choosing which story gets told.

Integrating and deploying without breaking your stack

Deployment starts with the systems already running the business. That usually means a customer relationship management (CRM) system, a billing or claims system, an electronic health record (EHR) system for healthcare balances, and in-house software that nobody wants to replace during a collections project. The IC platform connects to CRM, EHR, billing systems, and custom in-house software.

Map the data before mapping the workflow

Account status and balances move outward to the contact platform. Payments post back to billing. Suppression lists sync in both directions so an opt-out captured in one system reaches every other system. If any one of those flows runs on a delayed file or a manual upload, staff will work from stale balances and consumers will get contacted after they should've been suppressed.

Integration usually fails on duplicate or stale records. Bad account data causes wrong-party contacts, incorrect balances, and broken suppression. Data cleanup isn't a pre-launch courtesy task. It's the control that keeps the first dial from creating the first complaint.

Buyers should ask each vendor whether connectors are native or custom work. The follow-up matters more. Who owns field mapping after launch, and who approves a mapping change when billing adds a new field?

Run the pilot in parallel

Multiple locations, branches, or remote teams can be managed from a single administrative portal, including permissions, reporting, and routing. Multi-site operations should still get the role model in writing. A permission that works at headquarters can expose the wrong queue, portfolio, or payment function at a branch.

For rollout, request a written plan with milestones rather than a generic timeline. Pilot one portfolio segment alongside the existing dialer. Run the same accounts through both systems, reconcile contact counts, payment postings, and suppression results, and resolve every mismatch before cutover.

Script changes need the same discipline as code changes. Each change should have an approver, a version, an effective date, and a log entry. Without that trail, a team can't prove which language a consumer heard on a specific date.

Buyer checklist and how to shortlist vendors

Shortlists should be built on fit. A collections agency running high-volume outbound needs pacing, suppression, and ledger controls that hold up at scale. A healthcare revenue cycle team needs patient payment plans, clear balance explanations, and careful access control. A financial services contact center that mixes inbound service with collections needs routing and payment flows that work together. Feature count doesn't capture those differences.

Buyer checklist by decision group

Pricing for Agentic Collections is not published and is available by contacting the company. Every vendor, including IC, should provide a written quote tied to the same volume, channels, users, and term assumptions. Quotes built on different assumptions can't be compared fairly.

A hands-on trial of any vendor’s AI agent is useful, but it doesn't replace a scored evaluation. Score two or three shortlisted platforms against the same sample of 50 accounts, with one column per checklist group. Include routine balances, a dispute, a hardship case, a broken promise, and an opt-out in that sample.

Rejection rule: Remove any vendor that can't produce the handoff log, the call-frequency ledger export, or the cohort recovery report for the test sample.

That test can run this week. It doesn't require a full procurement cycle, and it exposes the difference between a polished demonstration and a platform that leaves evidence.

Intelligent Contacts offers AI collections software that connects compliant outreach, AI agent and live agent handoffs, self-service payments, and audit-ready reporting in one platform for regulated collections, healthcare, and financial services operations. Buyers who want to test Grace against real dispute, hardship, suppression, and payment scenarios can start at Intelligent Contacts.