The problem
The firm already had named client types, but they were built for messaging only. Nobody could say which persona was worth more, which chapter each one actually filed, or where the money came from. Marketing spend, intake scripts and content priorities were being set on instinct.
The first instinct, letting a clustering algorithm find the segments, did not work. With chapter and payment included, the model simply defined clients by their own outcomes. With them removed, the segments fell apart into age and income slices with no commercial meaning.
What I did
A trigger-first hierarchy. I rebuilt the segmentation by hand around the event that makes someone call: garnishment, repossession, foreclosure, a prior discharge, then income source, then general debt pressure. Each case lands in exactly one persona, so nothing is double counted and every segment has a trigger a marketer can write to.
Value at equal maturity. Because every case was filed the same year, raw totals flatter older cases. I compared collected value by case age (0–60 to 181–240 days) and by chapter, so personas are judged like for like.
A playbook per persona. Each persona carries what to say, what never to say, the channel that reaches it and the intake risk, linked back to the firm’s earlier named client types.
What it showed
- 34.4%Of collections from one persona: the “slow squeeze” household with no dramatic trigger
- 47 / 53Chapter 7 vs 13 split of cases, while Chapter 7 brings in about 60% of collections
- 4 newIncome-source and business segments the earlier persona work had missed
The full interactive report is below, anonymised: the firm, locations, volumes and revenue are removed, and money is shown as an index.
Ten trigger-first personas, priced by case age.
A segmentation built from an eight-month 2026 case export for two offices of a US consumer bankruptcy firm. Persona definition is kept separate from chapter and collected value, so both can be measured as outcomes.
How to read the numbers. Client name, locations, volumes and revenue are removed. Money is shown as an index where the portfolio’s average paid per case to date = 100, and volumes as shares. Chapter and amount paid are outcomes here, not inputs. Monthly expenses and income were left out of the segmentation. All cases were filed in 2026, so this is an early-lifecycle cohort, not mature lifetime value.
Why the segmentation is trigger-first
Two automated clustering runs were tested before this segmentation was hand-built, and neither was kept. A 10-cluster KMeans that excluded chapter and amount paid scored about 0.10 on silhouette, against 0.18 for the earlier model that included them.
So the report uses a mutually exclusive trigger-first hierarchy. Urgent enforcement triggers come first, then asset-protection and repeat-filing signals, then income-source and life-stage patterns, and finally broader car and debt pressure. Each case belongs to exactly one persona. That prevents double counting and keeps the trigger readable for marketing and intake.
Earlier model: circular, superseded
Included chapter and amount paid, so the segments defined themselves by their own outcomes.
New KMeans: tested and rejected
With the outcomes removed, separation collapsed. The result was demographic slices, not commercially distinct client types.
Trigger-first hierarchy: in use
Rule-based, so it has no silhouette score. Judge it on whether the personas differ commercially. The table below shows they do.
- 01Being garnished
- 02Repo, car booted, licence suspended or parking tickets
- 03Foreclosure or equity
- 04Prior discharge
- 05High income
- 06Business income
- 07SSI, retirement or disability income
- 08Food or government assistance, unemployment, household support
- 09Car as a stated reason
- 10Remaining debt-pressure cases
Chapter 13 carries the case count. Chapter 7 carries the collections.
Collected value rises sharply with case maturity
Amount paid to date, by days since filing. Compare chapters at similar age, never raw totals.
Chapter 7 average, 0–60 days to 181–240 days.
Chapter 13 average over the same bands.
No mature Chapter 13 cohort and no payment-level history in the export.
Median describes the typical client. Average sizes the revenue.
When the average sits well above the median, collections are concentrated in a minority of cases. Use the median to describe a typical client and the average to size revenue, and never mix the two in one claim.
| Persona | Share of cases | Chapter mix 7 / 13 | Median paid | Avg paid | Share of collections ▾ |
|---|---|---|---|---|---|
| All personas | 100% | 47 / 53 | — | 100 | 100% |
Select a persona
How the ten personas map to the named client types
Earlier persona work named client types by trigger event for messaging: The Slow Squeeze, Keep-My-Car, The Short Paycheck, The House Defender, The Booted Driver and The Second-Timer. This segmentation splits the same population more finely so it can carry chapter mix and collected value. Four of the ten are new: the income-source and business segments the earlier work did not separate.
Compare at equal maturity, not raw totals
Use the personas to compare acquisition quality at the aggregate level: persona × chapter × channel × case age. For near-term cash economics, compare clients at similar maturity rather than lifetime-to-date totals.
Anonymised portfolio version of a client report by Anna Mari Pitta. Firm name, office locations, case volumes, dates and revenue removed; money shown as an index. Internal analysis, not legal advice.