Last updated: June 2, 2026
Most B2B SaaS leaders watch the voluntary-churn line in their cohort dashboard. They forget that the line right next to it — involuntary churn from failed payments — is bleeding just as much MRR, and almost all of it is recoverable.
The 2025 Recurly Churn Report puts involuntary churn at 20–40% of total churn across subscription businesses, and the broader subscription economy loses an estimated $129 billion a year to failed payments alone. Rules-based dunning recovers a chunk of that. AI-driven dunning recovers most of the rest.
This post is the playbook the top-quartile RevOps and finance teams are running in 2026. It covers the decline taxonomy, the retry timing math, the customer-facing communication sequence, and where AI-decline models actually move the recovery rate.
Dunning is the process of recovering revenue when a recurring charge fails. In a subscription business, that means: a card declines on renewal day, your billing system flags it, and a sequence of retries and customer messages kicks off until either the payment clears or the subscription cancels.
That definition has not changed in twenty years. What has changed is the depth of signal a billing stack now has access to:
"AI dunning management" is just the discipline of using all four signals together instead of running a static retry calendar.
If your finance team reports a 7% gross churn rate and a 4% voluntary churn rate, you have a 3-point involuntary churn problem hiding in plain sight. Three points of MRR, recovered, compounds aggressively at any reasonable scale.
The leverage is real because Recurly's own benchmark data shows up to 70% of involuntary churn comes from failed transactions where the customer never intended to leave. Expired cards alone account for around 42% of failures. A subscriber whose card number rolled from a 4-digit prefix change at their bank does not need a save-team call — they need their card updated.
| Decline category | Share of failures | Right move |
|---|---|---|
| Expired / replaced card | ~42% | Account updater + soft email |
| Insufficient funds | ~20% | Delay retry 24–72h, time it to payday |
| Do not honor / issuer block | ~18% | Vary acquirer, retry on different network |
| Lost / stolen / fraud | ~8% | Stop retries, ask for new method |
| Other / soft declines | ~12% | Smart retry on ML schedule |
If your dunning system treats those five categories the same way, you are leaving recovery on the table. The teams using AI voice agents for debt collection-style workflows on the B2C side have already learned this lesson; B2B SaaS is two years behind and catching up.
A modern dunning system is not one product. It is four layers stitched together. If any one of them is missing, you cap out at ~50% recovery and call it normal.
Visa, Mastercard, and Amex all run account-updater networks that share new card numbers and expirations with merchants before the old credential dies. Plug this into your payment processor first. It silently solves the 42% of "expired card" failures with zero customer friction.
Static retry calendars (day 1, day 3, day 7) ignore everything we know about how issuers behave. Smart retry engines use historical outcomes to pick a retry window per BIN, country, and decline code. Benchmark data from 2025 shows ML-scheduled retries outperform rules-based dunning by 2–4x on soft-decline recovery.
The dunning emails of 2018 — "your payment failed, click here" — feel like spam. Modern comms layers vary copy by customer segment, health score, and channel (in-app banner vs. email vs. SMS vs. for high-ACV, a call from CS). LLM-driven copy variants A/B-test their way to a 10–15 percentage-point lift on open and click-through.
The last 10–15% of failed payments are the messiest: an AP person on the customer side needs a new invoice, a PO has to be updated, or finance needs the COO to approve a re-up. This is where conversational AI workers earn their keep — chasing the long tail of high-ACV failed payments the way a SDR would chase the long tail of stalled deals. Darwin AI's collections worker, Rio, is built for exactly this layer: persistent multi-channel follow-up on the small number of failed renewals that drive most of the recoverable revenue.
The single biggest lever inside a dunning system is when you retry. The wrong answer is "every other day until 14 days." The right answer is "depends entirely on the decline code and the customer."
Insufficient funds is the cleanest example. Retrying 4 hours later is wasted; the balance has not changed. Retrying 48–72 hours later, ideally on a Tuesday or a payday-adjacent date, sees recovery rates 3x higher than naive retries.
"Do not honor" responses are often issuer fraud heuristics tripping on a recurring charge. The fix is to vary the acquirer or the network — sometimes literally re-running the same transaction through a different processor — and the recovery rate jumps without any customer involvement.
Stolen cards, closed accounts, expired cards with no updater hit. Don't retry. Ask the customer for a new payment method, fast. Every wasted retry on a hard decline hurts your authorization rate with the issuer.
This is where AI-driven dunning starts to look less like a billing problem and more like a CX problem. Treat it like one.
If you already run an AI-led churn-reduction engagement program, this sequence should plug into the same orchestration layer. Dunning and voluntary churn are two faces of the same retention problem.
Most dashboards track recovery rate and stop. That is the equivalent of running sales on win rate alone.
Skip vanity metrics like "emails sent" and "retry attempts per failure." They reward activity, not outcomes.
If you fixed only those five, you would likely move your recovery rate 15–20 percentage points without touching the underlying billing stack. Pair them with the right B2B onboarding-and-retention motion, and involuntary churn stops being a quietly accepted line item.
Rio, Darwin's AI collections worker, runs your dunning sequence across email, SMS, and voice — handing every long-tail failed renewal back to your finance team only when it actually needs a human.
See how Rio handles dunning →AI dunning management is the use of machine-learning models for retry scheduling, account-updater networks for card hygiene, and LLM-driven comms to recover failed subscription payments. It improves on rules-based dunning by adapting to the specific decline code, customer history, and account value rather than running a fixed retry calendar.
The 2025 Recurly Churn Report puts involuntary churn at 20–40% of total churn across subscription businesses, with roughly 70% of those failures recoverable when the right combination of retries, account updater, and comms is in place.
Three to four automated retries spread over 7–14 days is the sweet spot. More than that creates issuer fatigue and downgrades your authorization rate. After the final automated retry, pause rather than hard-cancel — paused subs reactivate at 15–25% within 90 days.
No. AI dunning handles the high-volume, low-touch failures (expired cards, soft declines) at scale. CS still owns the high-ACV exceptions where a relationship-based human conversation is the only thing that clears the failure.
Account-updater and smart-retry changes typically show results inside one billing cycle (30 days). Comms-layer changes take 60–90 days to compound through a cohort. Plan for a quarter of measurement before declaring a winner.