Which Accounts to Work First: Prioritization in Collections

A collections operation never chooses whether to work accounts, only which ones, because capacity is finite and the pool is not. Most sort by balance, and balance is the weakest of the three terms that matter. A worked example below gets 27% more recovery from the same 400 contacts purely by reordering the list. The harder part is what must be excluded before any optimization runs, because a queue that maximizes recovery without gates will find the customers on whom pressure works and call that a performance improvement.

Capacity is the only real constraint

A collections operation does not choose whether to work accounts. It chooses which ones, because capacity is finite and the pool is not. Every account not worked today is a decision, made either deliberately or by whatever order the work list happened to come out in.

Most operations sort by balance, on the reasoning that the largest debts are worth the most effort. That reasoning is wrong in a specific and measurable way, and correcting it is the cheapest available improvement in recovery.

Per our incentive analysis, what gets worked follows what gets measured, so a team paid on dollars collected will sort by balance whether or not the work list does.

Expected value, not balance

The value of working an account is the balance multiplied by the probability of recovering it, minus the cost of the attempt. Balance alone is one of the three terms.

Take a stylized pool of 1,000 accounts in four segments, with a cost to work of $22 per account:

Segment	Accounts	Balance	Recovery probability	Expected value	Net of cost
High balance, low propensity	250	$4,200	8%	$336	$314
Mid balance, mid propensity	300	$1,850	19%	$352	$330
Low balance, high propensity	300	$850	34%	$289	$267
Recent, very high propensity	150	$1,200	46%	$552	$530

The segment with the largest balances has the lowest expected value. The segment with the second smallest balances has the highest. Now apply a capacity of 400 accounts and compare the two orderings:

• Sorted by balance: all 250 high-balance accounts, then 150 mid-balance. Expected net recovery $127,925.
• Sorted by expected value: all 150 recent high-propensity accounts, then 250 mid-balance. Expected net recovery $161,875.

That gain requires no additional resource. It is a reordering, and it is available to any operation that can score propensity at all. Per our benchmarking analysis, the reason it goes unclaimed is that balance is visible without a model and propensity is not.

What drives propensity

The model does not need to be sophisticated to beat balance ordering. Four factors carry most of the signal:

1. Recency of delinquency. The strongest single factor by a wide margin. Per our recency analysis, the probability of recovery falls steeply with time since last payment, which is why the recent segment above is the most valuable despite a middling balance.
2. Contactability. Per our contact analysis, an account with no working contact route has a recovery probability near zero regardless of everything else, and working it consumes capacity anyway.
3. Prior engagement. Someone who has answered, explained, or made a partial payment behaves very differently from someone who has never responded.
4. Capacity to pay. Per our affordability analysis, the balance is what is owed and capacity is what can be paid, and where they diverge sharply the recovery probability is low no matter how motivated the customer is.

Factor two deserves separate treatment rather than inclusion in a score. An uncontactable account should leave the calling queue entirely and go to a tracing process, because keeping it in the queue costs the capacity that a contactable account would have used. Mixing "cannot reach" with "low propensity" hides a fixable problem inside an unfixable one.

Where expected value ordering goes wrong

Pure expected value maximization produces outcomes that are commercially efficient and, in several respects, indefensible. The problems are real and they are not solved by ignoring them.

Problem	Why it happens	What it produces
Vulnerable customers are deprioritized	Low capacity to pay scores as low propensity	The customers most in need of an arrangement are never contacted to offer one
Or targeted, depending on the model	High compliance with pressure also scores as high propensity	The opposite failure, and the more serious one
Never-worked accounts stay never-worked	No engagement history means a low score, which prevents engagement	A self-confirming loop
Disputes and errors go undetected	An account nobody contacts is an account whose problems nobody hears	Wrong balances pursued for years, per our errors analysis
Uneven treatment across groups	Propensity models use correlated inputs	Systematically different contact rates, and a fair treatment question

The second row is the one that should stop a deployment. A model trained purely on who pays will learn that some customers pay because pressure works on them, and per our vulnerability analysis that is precisely the population where intensified contact is least appropriate. An expected value queue with no vulnerability suppression will find those customers and prioritize them, and it will look like a performance improvement in every metric the operation reports.

The case for a contact floor

Pure optimization leaves a large share of the pool never contacted, which is efficient and produces three costs that do not appear in a recovery metric.

• Unheard disputes. An account with a genuine error stays on the books and keeps being reported, per our compliance analysis, because the only route to discovering the error was a conversation that never happened.
• Missed arrangements. Per our payment plan analysis, some low-propensity accounts are low-propensity because nobody has offered a plan they can afford, and the score cannot distinguish that from unwillingness.
• A model that never learns. Per our testing analysis, a queue that only works high-scoring accounts generates no outcome data for low-scoring ones, so the score's accuracy in that range never improves and cannot be checked.

The third point is the one with a technical fix. Reserve a small random share of capacity, on the order of 3% to 5%, for accounts the score would not have selected. It costs a measurable amount of recovery and it is the only way to know whether the ordering is right. Per our selection analysis, a system evaluated only on the cases it chose cannot tell you about the cases it declined.

Building the queue

1. Apply hard exclusions first. Legal restrictions, vulnerability flags, disputes, bankruptcy, deceased. These are gates, not weights.
2. Route uncontactable accounts out of the queue to tracing rather than leaving them to absorb capacity.
3. Score propensity on recency, prior engagement, and capacity to pay.
4. Compute expected value net of the cost to work, which is what makes low-balance high-propensity accounts comparable to the reverse.
5. Apply customer-level caps, per our customer view analysis, because a customer with four accounts must not receive four times the contact.
6. Reserve 3% to 5% of capacity as a random holdout so the score keeps being tested outside its comfortable range.
7. Add a contact floor guaranteeing every account at least one contact attempt within a stated period, regardless of score.
8. Monitor contact rate by segment and by group, not only recovery, because an optimization that improves recovery while concentrating contact unevenly is a compliance exposure rather than a win.

Item seven is the one most likely to be argued against internally, because it costs recovery and buys something that appears in no revenue metric. The argument for it is that an account nobody has ever contacted is an account whose balance nobody has ever verified, and per our audit trail analysis an operation that cannot show it attempted contact before pursuing a debt for years is in a weak position when eventually asked.

What to measure

Metric	Why it matters
Recovery per contact hour	The actual efficiency measure; total recovery hides capacity waste
Realized against predicted propensity by band	Tells you whether the score works, per our monitoring analysis
Share of pool never contacted	The hidden cost of optimization, and the number nobody reports
Holdout performance	Whether the ordering beats random, measured rather than assumed
Contact rate by segment and group	Where uneven treatment becomes visible
Dispute rate on first contact	A high rate suggests balances are being pursued without verification

The third row is the number to add first if you report none of these. Every operation knows its recovery rate. Very few know what share of their pool has never been spoken to, and that figure determines how much of the recovery rate is a result of good targeting rather than of a large denominator nobody touched.

Frequently asked questions

---

This guide is educational and does not constitute legal, compliance, or model risk advice. All figures are stylized illustrations using assumed segment sizes, balances, recovery probabilities, and contact costs, and do not describe any particular portfolio. Contact frequency limits, communication requirements, dispute and validation obligations, treatment of vulnerable customers, and fair treatment expectations vary by jurisdiction, by product, and by whether collection is first-party or third-party. Consult your compliance and legal functions before changing any contact strategy.