The ROI case for propensity-driven collections is not an AI argument — it is a contact economics argument. A collections agent making 60 calls per day on a DPD-ordered queue is spending the same cost per contact regardless of whether each account will pay. A collections agent making 60 calls per day on a propensity-ordered queue is concentrating that cost on the accounts where each call has the highest probability of producing a payment. The difference in recovery rate is not a technology outcome. It is a prioritisation outcome.
What DPD-ordered collections actually produces
In a DPD-ordered collections operation, the agent queue is sorted by how long the account has been overdue. The DPD 90 accounts receive the most attention; the DPD 5 accounts receive the least. This order has an internal logic — the older accounts are at higher risk of permanent non-recovery — but it ignores two critical variables: the borrower's current collectibility, and the cost of agent time spent on accounts that are not collectible today regardless of how much attention they receive.
A typical DPD-ordered queue for a 2,841-account DPD book might look like this: the top 500 accounts are DPD 60–90, of which perhaps 180 (36%) are genuinely collectible today based on their engagement and financial signals. The remaining 320 accounts receive agent contact that produces no payment — at the same per-contact cost as the 180 that do pay. Meanwhile, 200 DPD 15–30 accounts that received salary credits yesterday and have clicked payment links this morning are sitting at the bottom of the queue, receiving contact on day 3 after the salary window has already begun to close.
The before/after comparison: same team, same contacts, different order
uplift (63.4% → 78.2%)
contact (₹412 → ₹265)
DPD (54 → 32)
Where the uplift comes from — the four mechanisms
The 14.8 percentage point improvement in recovery rate is not magic — it comes from four identifiable mechanisms that propensity-ordered collections enables and DPD-ordered collections does not.
Mechanism 1 — Salary window capture. In a DPD-ordered queue, an account with a recent salary credit sits at position 1,200 because its DPD is only 18. In a propensity-ordered queue, the same account with a propensity score of 88 sits at position 3. The contact happens on day 2 after the salary credit — inside the 5-day payment window that closes as discretionary spending absorbs the credit. This mechanism alone is responsible for an estimated 6 percentage points of the recovery rate improvement.
Mechanism 2 — Legal notice urgency capture. The 7 days after a Section 13(2) notice is served represent a high-propensity window — the borrower has received a legal document and is more motivated to engage than at any other point in the delinquency cycle. In a DPD-ordered queue, contacting an account within 7 days of notice service requires knowing the notice was served and prioritising accordingly — which rarely happens systematically. In a propensity-ordered queue, the legal notice event is a signal that moves the account up the queue automatically.
Mechanism 3 — Contact cost reallocation. At a cost of ₹412 per contact in a DPD-ordered queue (including the 82% of contacts that produce no payment), agent capacity is consumed by accounts that are not collectible today. In a propensity-ordered queue, the same capacity is concentrated on the 28% of contacts that produce payments — the non-collectible accounts receive lower-cost automated channels (IVR, SMS) rather than agent time. The ₹147 saved per contact is reallocated to additional contacts on high-propensity accounts.
Mechanism 4 — Earlier resolution reduces write-off provisioning. Accounts that cure at DPD 32 (propensity-ordered average) rather than DPD 54 (DPD-ordered average) spend 22 fewer days in the NPA provisioning bucket. For a portfolio of 2,841 accounts, this difference in provisioning duration has a direct P&L impact in the period — the institution carries lower provisions for a shorter period, improving the reported NPA ratio even holding recovery rate constant.
The ROI by portfolio segment — where propensity scoring adds most
| Segment | DPD-Ordered Recovery Rate | Propensity-Ordered Recovery Rate | Uplift | Primary Mechanism |
|---|---|---|---|---|
| DPD 1–30 · Salaried · Metro | 84% | 94% | +10pp | Salary window capture — contacts arrive in window, not after |
| DPD 31–60 · Mixed · All cities | 62% | 76% | +14pp | Engagement signal capture — partial payment signals acted on same day |
| DPD 31–60 · Post-legal notice | 48% | 67% | +19pp | Legal urgency window — highest uplift segment |
| DPD 60–90 · Secured · Engaged | 41% | 58% | +17pp | Engagement pattern identification — engaged borrowers surfaced from DPD-buried queue |
| DPD 60–90 · Disengaged · Unsecured | 22% | 26% | +4pp | Low uplift — cost reallocation away from this segment is the real gain |
| DPD 90+ · NPA · Hard Bucket | 18% | 21% | +3pp | Marginal uplift — primary value is agent time freed from this segment for higher-propensity accounts |
The financial model: what the improvement means in rupees
For a portfolio of 2,841 accounts with an average outstanding of ₹8.4 lakhs per account and a weighted DPD 0–90 book value of ₹238.6 crores, the 14.8 percentage point improvement in recovery rate represents approximately ₹35.3 crores in additional recoveries compared to the DPD-ordered baseline. Against the cost of the Propensity Scoring Agent AI, the return is realised within the first quarter of deployment — primarily driven by the salary window capture and legal notice urgency mechanisms in the DPD 1–60 segment where the uplift is largest.
The provisioning benefit adds to this. Moving the average resolution DPD from 54 to 32 for the accounts that cure reduces the NPA provisioning requirement during the period — a balance sheet improvement that, for a lending institution running against provisioning targets, may be as significant as the direct recovery improvement.
The ROI is not about better technology — it is about better decisions made faster
The collections manager who builds a propensity-ordered queue manually — reviewing each account's salary schedule, call history, engagement pattern, and financial signals — would make the same decisions the Propensity Scoring Agent AI makes. But they cannot do it in seconds for 2,841 accounts, updated daily, with real-time signal integration. The AI does not make qualitatively different decisions from an expert collections manager. It makes the same decisions that an expert would make for every account in the book, simultaneously, every morning, before the team makes its first call. The ROI is the scale at which expert-quality prioritisation becomes operationally possible.