Recurring-Gift Recovery Agent
Fixed daily retries treat every decline the same. A decline-aware policy, with an agent that triages each failed gift and drafts a warm note to the donor, recovers more monthly gifts with far fewer retries.
40,000 donors · 5 fictional nonprofits · 480,000 monthly charges · 35,992 failed (7.5%) · $1.75M at stake
How the agent works · Detect → Prioritize → Recommend → Quantify → Learn
- 1Detect
Reads each failed gift’s decline code and sorts it: soft, expired, hard or ACH.
- 2Prioritize
Ranks the queue by expected recoverable value, so the biggest wins come first.
- 3Recommend
Proposes one plan per gift. A person approves it before anything is sent.
Console - 4Quantify
Records every decision and outcome in a hash-chained impact ledger.
Impact ledger - 5Learn
Outcomes show which retry timing and outreach work best for each decline.
What it learned
vs $766k baseline · $269k more a year
43.8% → 59.2% of failed dollars recovered
64% fewer authorisation attempts · $8.8k in attempt fees
vs 27,926 under baseline — retries on do-not-retry codes or past the cap
14,304 of 21,835 donors who had a failure stay giving
Forward value of the extra donors kept, at 1.5% monthly churn
Recovered dollars, cumulative over 12 months
Same failures, two policies. The dotted line is every failed dollar — the ceiling.
- Smart policy + agent
- Baseline: fixed daily retries
Where the dollars came from
Recovered dollars by the action that won them back.
- Retries
- Card updater
- Donor outreach
- ▸ Updater first: expired cards are refreshed the same day, not at month end ($196k vs $8.3k).
- ▸ Outreach that works: a three-step, warm sequence with a self-serve link ($184k vs $70k).
- ▸ Fewer, better-timed retries: $655k from 50,404 attempts in total, vs 138,351.
Recovery rate by decline category
Share of failed dollars recovered; the number is the difference in points. Baseline edges ahead on do not honor and ACH R01 funds by retrying more often — for ACH R01 only by re-presenting past the cap.
- Baseline
- Smart
Attempts per recovered gift
Share of recovered gifts by how many authorisation attempts it took. Baseline: 25,393 attempts could never have succeeded.
- Baseline
- Smart
Time to recover
Median 2 days (baseline) vs 2 days (smart). Payday-timed retries and outreach take a little longer and win back far more.
- Baseline
- Smart
What the policy learned from outcomes
From 156,529 simulated retries and 39,998 outreach sequences across both policies. These are synthetic outcomes, so they show what the model assumes; the real version learns the same table from de-identified BBIP retry history.
| Decline | Best retry timing | Outreach that worked best |
|---|---|---|
| Insufficient funds5112,593 failures | ✓ 0–3 days after payday40% of these retries worked, vs 8% at other times (46,165 attempts). | Three-step sequence · 13% repliedvs 4% to one generic notice. Most replies (54%) came from the friendly email. |
| Do not honor055,411 failures | ✓ 0–3 days after payday24% of these retries worked, vs 14% at other times (23,261 attempts). | Three-step sequence · 13% repliedvs 5% to one generic notice. Most replies (53%) came from the friendly email. |
| Issuer unavailable911,804 failures | ✓ Within hours72% of these retries worked, vs 42% when retried later (6,121 attempts). | Three-step sequence · 15% repliedvs 8% to one generic notice. Most replies (53%) came from the friendly email. |
| Suspected fraud591,770 failures | ✓ After the donor confirms80% of these retries worked, vs 2% before the donor confirmed (9,080 attempts). | Three-step sequence · 46% repliedvs 4% to one generic notice. Most replies (55%) came from the friendly email. |
| Expired card547,178 failures | ✓ On the updated card86% of these retries worked, vs 1% on the original card (39,439 attempts). | Three-step sequence · 30% repliedvs 11% to one generic notice. Most replies (54%) came from the friendly email. |
| Lost or stolen411,103 failures | ✓ On the updated card81% of these retries worked, vs 0% on the original card (5,840 attempts). | Three-step sequence · 30% repliedvs 11% to one generic notice. Most replies (57%) came from the friendly email. |
| Account closed621,405 failures | × Never retry0 of 7,025 retries of the same method succeeded. | Three-step sequence · 29% repliedvs 9% to one generic notice. Most replies (57%) came from the friendly email. |
| Invalid card141,105 failures | × Never retry0 of 5,525 retries of the same method succeeded. | Three-step sequence · 27% repliedvs 9% to one generic notice. Most replies (57%) came from the friendly email. |
| ACH R01 fundsR012,152 failures | ✓ 0–3 days after payday50% of these retries worked, vs 12% at other times (6,718 attempts). | Three-step sequence · 13% repliedvs 3% to one generic notice. Most replies (58%) came from the friendly email. |
| ACH R02 closedR02922 failures | × Never retry0 of 4,610 retries of the same method succeeded. | Three-step sequence · 25% repliedvs 10% to one generic notice. Most replies (55%) came from the friendly email. |
| ACH R03 no acctR03549 failures | × Never retry0 of 2,745 retries of the same method succeeded. | Three-step sequence · 25% repliedvs 9% to one generic notice. Most replies (57%) came from the friendly email. |
By organisation
Five clearly fictional nonprofits with different gift sizes and ACH mixes. The uplift holds for each.
| Organisation (fictional) | Donors | Failed | Baseline | Smart | Uplift |
|---|---|---|---|---|---|
| Riverbend Food Collective | 10,310 | $352k | $155k 44% | $212k 60% | +16.2 pts |
| Northgate University Foundation | 8,907 | $524k | $234k 45% | $310k 59% | +14.4 pts |
| Harbor Light Children’s Hospital Foundation | 9,627 | $434k | $187k 43% | $253k 58% | +15.2 pts |
| Cedar Valley Animal Rescue | 6,324 | $210k | $94k 45% | $128k 61% | +16.2 pts |
| Lumen Arts Alliance | 4,832 | $230k | $96k 42% | $133k 58% | +16.1 pts |
Baseline
Retry every day for 5 days whatever the decline said, run the card updater at month end, send one generic notice.
Smart policy + agent
Never retry do-not-retry codes. Updater first for expired cards. Retries timed to payday. Warm three-step outreach with a self-serve link and an AI disclosure. A person approves each plan.
Fair comparison
Both policies replay the same twelve months of failures with the same random draws per gift, so only the policy differs.
Illustrative model. The real number comes from a backtest on de-identified BBIP retry history. Default seed, +15.4 pts uplift. Networks cap reattempts and flag do-not-retry codes; limits here are simplified. Synthetic data — no Blackbaud, donor or card data is used.