AI Dunning in 2026: What Works vs Marketing Hype

The Vendor Demo That Changed Everything

For CFOs, the only defensible outcome measure is cash recovered net of fees, not engagement proxies. A polished interface and frequent references to “AI” can mask a familiar reality: schedule‑driven dunning dressed in modern UX. Treat vendor demos as hypothesis generation, not proof. Set a pre‑commitment plan that includes a clean control group, clear recovery KPIs, and a time‑boxed pilot gated to renewal. Require line‑of‑sight from model output to money in the bank and insist on reporting that reconciles to your ledger. If uplift cannot be demonstrated within the first two quarter‑ends, the technology is an automation tool—not an intelligence advantage—and should be resourced, priced, and governed accordingly.

The AI Label Problem

“AI” has become a marketing veneer in collections. In most platforms, intelligence resolves to deterministic rules: if invoice age exceeds a threshold, send the next template; if a keyword appears, route to a scripted path; if no response, repeat. That is workflow, not learning. The financial risk is clear: premium pricing for capabilities that cannot generalize, adapt, or improve without reconfiguration. To protect ROI, mandate model transparency: enumerate each model, objective function, training data scale and refresh cadence, and drift monitoring. Demand outcome‑level A/B tests with holdout controls and recovery‑rate confidence intervals, not vanity metrics. If a vendor cannot provide feature importance, confusion matrices, and documented uplift versus a rules baseline, you are buying automation with an “AI” sticker—priced like strategy, functioning like a flowchart.

The Chatbot Illusion

Collection chatbots are positioned as autonomous negotiators; in reality, they classify a small set of intents and traverse a shallow decision tree. Experienced debtors rapidly identify delay paths and exploit non‑committal fallbacks. The result is activity without leverage: more transcripts, no acceleration in cash realization. For finance leaders, the governance posture is straightforward. Limit bots to low‑risk triage—document retrieval, payment‑link generation, balance confirmations—and set strict timers for rapid human takeover. Instrument every conversation with outcome tags and measure by dollars collected, not messages exchanged. Most importantly, align escalation to jurisdictional realities: bots cannot weigh enforceability, credit exposure to supply continuity, or reputational signals. Human collectors can—and they change the timeline from discussion to decision.

What Genuine AI in Collections Would Look Like

Real AI learns and adapts. In collections, that means predictive payment modeling continuously recalibrated on new outcomes, optimal‑timing engines that choose channel, day, and hour per debtor cohort, and policy learners that shift tactics as behaviors change. Delivering this requires rigorous data foundations (entity resolution, outcome labeling, legal event encoding), MLOps (drift detection, shadow testing, rollback paths), and governance (model cards, fairness checks, auditable feature lineage). Validation must be causal, not correlational: randomized holdouts, period‑over‑period stability, and uplift attributable to model decisions. Ask vendors to ship evidence packages: feature importances aligned to business intuition, performance by jurisdiction, and recovery lift versus your current automation. If they cannot, you are not buying intelligence—you are renting templates.

The Premium You Are Paying

Premium pricing without premium outcomes is leakage. Tie spend to recovery by demanding performance clauses, control‑group testing, and compensation indexed to net cash collected. Standardize KPIs across vendors: recovery rate by aging bucket, time‑to‑cash, and variance versus prior cohorts. De‑emphasize surrogate metrics—opens, clicks, chatbot sessions—that do not move working capital. Reallocate budget from “AI” licenses to jurisdiction‑specific expertise early in delinquency, when leverage is highest and settlements are least costly. The pattern is consistent: when skilled collectors enter before day 60, recovery curves steepen and legal exposure narrows. Pay for results, not rhetoric.

Why Debt Collection Resists Full Automation

Collections is not a neutral workflow—it is an adversarial negotiation constrained by jurisdiction and timing. Decisions that determine outcomes, such as whether and when to escalate legally, cannot be reduced to static thresholds. They require judgment about enforceability, liquidity, counterparty incentives, and portfolio signaling. Local presence amplifies credibility: a call from a recognized number, fluency with court processes, and the ability to activate counsel materially alter debtor calculus. Automation excels at consistency and scale; it struggles with ambiguity, stakes, and strategy. For CFOs, the operating model that endures blends automation for hygiene tasks with human escalation that carries real consequence. The aim is not to eliminate people; it is to deploy them precisely where leverage, law, and psychology intersect.

Intelligence That Cannot Be Automated

INTERCOL fields seasoned collectors across the UK, EU, USA, and UAE—professionals who convert ambiguity into enforceable outcomes. Their value is operational intelligence: selecting the right venue, reading intent from conversation dynamics, and timing escalation to compress cash cycles. Technology supports the work—data access, document trails, and monitoring—but it does not substitute for jurisdictional expertise and presence. If your current dunning stack already meets target recovery, keep it. If not, stop funding if‑then trees masquerading as intelligence and redirect budget to accountable results. See what happens when strategy is executed by experts with real leverage. Talk to INTERCOL at intrcl.com and align your collections spend to the only KPI that matters—cash in.

Related Intelligence

Sources & References