Every ARM platform claims AI now. That phrase has become as meaningless as ‘cloud-based’ or ‘data-driven.’ What’s less common is an honest accounting of what compliance-grade AI in collections actually requires, and why most of what’s marketed as AI doesn’t meet that bar.
The gap matters. Regulators and plaintiff attorneys are catching up to how AI gets used in collections, and the questions they’re asking are technical, specific, and hard to answer without documentation.
What regulators actually want to know
The compliance inquiry around AI in collections comes down to 3 areas: explainability, fairness testing, and audit trails.
Explainability means your AI can tell you why it made a decision. If your system scores an account and recommends a contact strategy, can you reconstruct the reasoning? If a regulator or plaintiff attorney asks why a consumer received a particular communication at a particular time, is there an answer on record?
Fairness testing means you’ve actively checked whether your model produces disparate outcomes across demographic groups. Disparate impact claims are increasingly showing up in collections litigation. If you can’t demonstrate you tested for it, you’re exposed.
Audit trails mean every AI-driven decision is logged, timestamped, and retrievable. Not for 30 days. For the full applicable statute of limitations.
Most AI in ARM doesn’t meet all 3. Some meets none.
The scoring problem
A lot of AI in ARM is actually predictive scoring: build a model at placement, score the portfolio, work the accounts. That’s useful, but it’s not what separates compliant operations from exposed ones.
The problem is that accounts change. A consumer who was uncollectable at placement may become collectible 6 weeks later. A consumer who was contactable under Reg F may have since revoked consent. A static scoring model can’t see any of that.
TSI’s CollectX platform scores 200 million consumer records daily. Not at placement. Not quarterly. Daily. The difference matters because collections decisions made on stale data create legal exposure. Contacting someone after they revoked consent because your system didn’t update isn’t a gray area. That’s a TCPA violation.
Real-time monitoring isn’t a feature, it’s a requirement
Quality assurance in most contact centers means sampling. Someone listens to 2% of calls, scores them on a rubric, and reports results monthly. By the time a compliance issue surfaces in that workflow, hundreds of additional problematic interactions have already occurred.
TSI’s Ripple AI monitors 100% of communications in real time. Every call, every channel. Not to grade agents on performance metrics. To catch compliance violations before they generate liability.
That distinction matters. Sampling-based QA protects you from knowing about problems. Real-time monitoring protects you from having problems.
MENSA AI: what predictive contact optimization actually looks like
TSI’s MENSA AI optimizes contact timing and channel selection based on individual consumer behavior patterns, not population averages. Population-average contact strategy is how you end up calling everyone at 9 AM because that’s when aggregate answer rates peak. MENSA AI learns when individual consumers are actually likely to engage, and routes outreach accordingly.
The result is higher contact rates with fewer total attempts, which reduces compliance exposure from over-contact. Every decision MENSA AI makes is logged. Every recommendation is traceable. That’s what makes AI defensible when someone asks about it in discovery.
The questions you should be asking vendors
If you’re evaluating AI-enabled ARM partners, the conversation shouldn’t start with capability claims. It should start with documentation requirements.
Can the vendor show you bias testing results from their models? Can they produce a decision audit trail for a specific account on a specific date? Is their scoring model static at placement or dynamic over the account lifecycle? What’s the refresh cadence?
When a vendor can’t answer those questions concretely, the AI they’re describing isn’t compliance-grade. It’s marketing.
The collections industry is going through the same reckoning every financial services sector faces when it adopts AI: the technology moves faster than the compliance framework. The firms that survive that transition built accountability into the system from the start, not after the first enforcement action.