The Recovery Frontier: How Hard to Work Each Account

Every collections operation makes an allocation decision thousands of times a day — how much effort to spend on this account — and almost none of them measure whether the answer is right. The economics are not complicated: work an account until the expected recovery from the next attempt falls below the cost of making it, then stop. What makes this hard in practice is that the point differs by an order of magnitude across accounts, while most operations apply a uniform sequence. The result is a specific and expensive failure: being wrong in both directions at once — pouring effort into accounts that will never pay while leaving recoverable balances under-worked because capacity was consumed elsewhere.

The frontier

The rule, stated formally:

Continue while: Δ(probability of payment) × Balance > Cost of the attempt

Three implications follow immediately, and each contradicts common practice.

The stopping point depends on balance. A one-percentage-point improvement in payment probability is worth $1.20 on a $120 account and $92 on a $9,200 account. The same attempt has a seventy-fold difference in value. Any operation applying the same sequence to both is necessarily wrong on at least one.

It depends on where you are in the sequence. The lift from attempt two is large; from attempt fourteen it is typically near zero. So the frontier isn't a property of the account alone — it's a property of the account and how much you've already done.

It depends on your cost structure, which means it isn't a fixed truth. An operation with a $9 cost per attempt and one with a $0.40 cost per attempt should behave completely differently on identical portfolios. Best-practice attempt counts borrowed from another operation are borrowed from another cost structure.

Pricing an attempt

Effort can't be optimized until it has a price, and most operations have never computed one.

Cost per attempt = Total collections operating cost ÷ Total attempts

Include everything, not just the collector's time:

• Collector compensation and benefits
• Supervision, training, and quality assurance
• Telephony, messaging, and postage
• Data, skip tracing, and contact refresh — per our contact data guide
• Systems, licensing, and integration
• Compliance, monitoring, and complaint handling
• An allocation of management time

Worked example. A ten-person team costing $780,000 fully loaded, making 195,000 attempts a year: $4.00 per attempt.

Now apply the frontier. At $4.00 per attempt, an account only justifies another attempt when the expected lift times the balance exceeds $4.00. On a $200 balance, that requires the attempt to raise payment probability by two full percentage points — which is plausible on attempt two and implausible on attempt twelve. On a $6,000 balance, the same $4.00 is covered by a lift of 0.07 points, which almost any attempt clears.

Most operations that run this calculation for the first time discover their cost per attempt is higher than assumed, which moves the frontier inward and reveals that a meaningful share of current activity is destroying value rather than creating it.

The lift curve

The second input, and the one that requires measurement rather than assumption.

Recovery probability rises with attempts, but with sharply diminishing returns. A representative shape — yours will differ, and the point is to measure yours:

Attempt	Cumulative payment probability	Marginal lift	Value at $200	Value at $6,000
1	8%	8.0 pts	$16.00	$480
2	14%	6.0 pts	$12.00	$360
4	21%	2.5 pts	$5.00	$150
6	25%	1.2 pts	$2.40	$72
9	28%	0.5 pts	$1.00	$30
14	30%	0.15 pts	$0.30	$9.00
20	31%	0.05 pts	$0.10	$3.00

Read the two right-hand columns against the $4.00 cost line.

The $200 account crosses below $4.00 somewhere around attempt five. Every attempt after that loses money. An operation running twelve attempts on this account is spending roughly $28 to chase about $2 of expected value.

The $6,000 account is still generating $9 of expected value at attempt fourteen and $3 at attempt twenty. It doesn't cross the cost line until well past twenty attempts. An operation that stops at nine because that's the standard sequence is abandoning positive-value work on its most valuable accounts.

Same portfolio, same team, same policy — and the policy is badly wrong for both accounts in opposite directions.

Worked allocation

What the reallocation is worth. Take a portfolio of 4,000 accounts: 3,200 small (average $250) and 800 large (average $4,500). Team capacity: 195,000 attempts a year at $4.00.

Under uniform treatment at ten attempts each: 40,000 attempts, of which the small accounts consume 32,000. Using the curve above, roughly half of those small-account attempts fall below the cost line — approximately 16,000 attempts, or $64,000, spent below the frontier.

Under frontier-based allocation: small accounts capped at five attempts (16,000 attempts), freeing 16,000 attempts. Redirect them to the large accounts, taking those from ten attempts to roughly thirty. Using the curve, extending from attempt ten to twenty on a $4,500 balance adds roughly one to one and a half percentage points of cumulative recovery probability — call it 1.2 points across 800 accounts at $4,500, or about $43,000 of additional expected recovery, at a cost of $64,000 that was already being spent.

Net effect: roughly $43,000 of recovery created and $64,000 of waste eliminated, from the same team and the same budget. No additional headcount, no additional spend — purely a reallocation.

The numbers are stylized and yours will differ. What won't differ is the sign. Any portfolio with meaningful balance dispersion and uniform treatment has this gap, and its size scales with the dispersion.

Wrong in both directions

Worth isolating, because it's the specific insight that changes behavior.

Collections managers under pressure generally reach for one of two levers: work harder or cut cost. Both are wrong when the problem is allocation.

• "Work harder" adds attempts uniformly, which adds most of them below the frontier on small accounts, because that's where the volume is.
• "Cut cost" removes attempts uniformly, which removes high-value attempts on large accounts along with the waste.

Neither addresses the actual problem, which is that the effort is in the wrong place. This is why collections operations can add headcount and see recovery barely move, then cut headcount and see recovery barely move — both changes are being applied to a distribution that is misallocated, so the marginal unit added or removed is close to worthless in expectation.

The diagnostic that reveals it in an afternoon: plot attempts per account against balance. If the relationship is flat, you have this problem. If it slopes upward but gently, you have a smaller version of it. The chart takes one query and it is the most informative thing most collections operations have never looked at.

Segmenting on the right variable

The frontier depends on expected recovery, not balance — and balance is a poor proxy for it. What actually predicts:

• Contactability. The binding constraint. An account you can't reach has near-zero lift from any attempt, regardless of balance, and the highest-return action is a contact refresh rather than another dial.
• Prior engagement. An account that responded once is fundamentally different from one that has never responded — and this is the strongest single behavioral signal available.
• Payment history before delinquency, which distinguishes a temporary disruption from a structural inability.
• Age. The decay in our collections framework means age dominates most other variables.
• Dispute status, since a disputed account requires resolution rather than escalation — the deduction dynamics in our short-pay guide.
• Capacity indicators where available, which distinguish won't-pay from can't-pay and determine whether a plan or a demand is the right instrument — the design question in our plan guide.
• Settlement potential, since an account whose expected recovery is well below face is a candidate for the discount arithmetic in our settlement guide rather than for more attempts at full balance.
• Balance, which matters for the frontier arithmetic but is a weak predictor of probability on its own.

The practical construction: score expected recovery, then multiply by balance, then rank on the product. Ranking on balance alone over-serves large accounts that will never pay. Ranking on probability alone over-serves small accounts that will. The product is the only ranking that maps to the frontier, and it is not what most operations use.

Setting stopping rules

1. Compute cost per attempt, fully loaded, and publish it internally. Everything downstream depends on this number existing.
2. Measure the lift curve by segment from your own historical data — recovery rate by attempt number, held constant on age and segment.
3. Find the crossing point per segment, where marginal lift times balance falls below cost.
4. Write the stopping rules down as explicit policy per segment, not as a global attempt count.
5. Reallocate freed capacity to segments still above the frontier — this is where the return is.
6. Set a low-cost maintenance track for accounts past the frontier rather than zero contact: periodic automated messaging costs almost nothing and captures the accounts whose circumstances change.
7. Re-measure quarterly, since both cost and curve shift.

On the sixth point — past the frontier does not mean abandoned. It means the treatment should cost less, not that it should stop. An account not worth a $4.00 dial may be well worth a $0.02 message, and the distinction between "stop working" and "work differently" is where most of the remaining value sits.

What automation does to the frontier

The most important structural point, and the one usually framed incorrectly.

Automation is typically justified as cost reduction. The frontier says its more valuable effect is expanding what's worth working. If the frontier is where marginal cost equals marginal recovery, lowering cost per attempt moves that point outward — accounts and attempt counts that were uneconomic become economic.

Return to the arithmetic. At $4.00 per attempt, the $200 account stops being worth working around attempt five. At $0.20 per attempt, the same account is worth working to roughly attempt fourteen — because 0.15 points of lift on $200 is $0.30, still above $0.20. The account didn't change. The frontier moved.

Which reframes the question from "how many collectors do we need" to "how much of our portfolio is currently below the frontier only because our cost per attempt is high?" For most operations that share is large, and it is concentrated in exactly the small-balance, high-volume accounts that consume the most human capacity for the least return.

The corollary worth stating: the optimal attempt count is not an operational truth, it's a function of your cost structure. Any operation that hasn't recalculated its stopping rules since changing its channel mix is running rules calibrated to a cost structure it no longer has.

Where the model stops applying

An analysis that claimed to determine every collections decision would be overreaching. Four genuine limits:

Compliance is not a cost to be optimized against. Contact frequency limits, time restrictions, and dispute obligations in our compliance guide are constraints on the feasible set, not variables in the objective. The frontier operates inside them.

Relationship value isn't in the balance. For a first-party creditor, an account belongs to a customer who may buy again. The retention value of a well-handled collection doesn't appear in the recovery figure, and it should shift treatment toward the accommodating end for customers worth keeping.

Complaint risk is convex. The expected cost of over-contacting isn't linear in attempts — it's small until it isn't, and then it's regulatory. That asymmetry argues for stopping somewhat earlier than the pure arithmetic suggests on any account where intensity would be high.

Measurement is noisy at the segment level. Lift curves estimated on thin segments are unreliable, and confidently reallocating on a curve fitted to two hundred accounts is a mistake. Use broad segments until you have volume to justify narrower ones — and track the aggregate against the roll rates and recovery decay in our metrics guide so a reallocation that looks good in one segment isn't quietly costing you elsewhere.

One further boundary: the frontier tells you how much effort an account justifies, not whether pursuing it is the right instrument at all. An account with real collectibility but a genuine dispute belongs in resolution, and one where the economics have run out entirely belongs in the disposition options our portfolio sale guide covers rather than in another attempt cycle.

Within those limits, though, the core claim holds and is unusually actionable: most collections operations could increase recovery meaningfully without spending anything, purely by moving effort from below the frontier to above it.

Frequently asked questions

---

This guide is educational and does not constitute legal or compliance advice. Contact frequency limits, communication restrictions, and dispute handling obligations are set by federal and state law and constrain any allocation strategy; worked figures are stylized illustrations rather than benchmarks.