Finding out who our small-business customers actually were

TL;DR

• Context: Sole UX owner for a retail bank’s small-business segment (~160K customers).
• Problem: The ~160K-customer base had no real segment map — UX work had nothing to target.
• What I did: Built personas from actual money flow, not interviews — 4 industry segments × 2 income tiers = 8 personas.
• Result: 6 groups = 75% of customers and 80% of inflows; the same method later forecast overdraft sales.

Thesis

Personas usually start in a meeting room: a handful of interviews, some sticky notes, a memorable name. The trouble is they’re easy to ignore — they feel like opinions. We had ~160K small-business customers and no real map of who they were, which meant none of the UX work could be targeted. So I built the personas from the one thing that doesn’t lie: what money actually flowed in and out, by industry.

flowchart TD
    A["~160K small-business customers"] --> B["6 groups<br/>75% of customers / 80% of inflows"]
    B --> C["4 groups stay distinct"]
    B --> D["4 of 6 groups merge into 2"]
    C --> E["4 industry segments"]
    D --> E
    E --> F["x 2 income tiers<br/>lower ~5-6th decile / higher 7th+"]
    F --> G["8 personas"]

    classDef data fill:#87CEEB,stroke:#333,stroke-width:2px,color:darkblue
    classDef result fill:#90EE90,stroke:#333,stroke-width:2px,color:darkgreen
    class A,B,C,D data
    class E,F,G result

What I did

• Drove a quantitative segmentation of ~160K small-business customers by business-activity code (their industry) — two axes per group: share of customers, and share of monthly inflows.
• Consolidated the result into 8 personas — 4 industry segments × 2 income tiers — starting from a 6-group cut.
• Put the segmentation to work on targeting research recruitment and profiling products for high-cash-flow segments.
• Stretched the same method into a sales forecast for overdraft credit — which held up against actual take-up.

Results

• 6 groups = 75% of customers and 80% of inflows — the segment had real, concentrated structure hiding under one blurry label.
• Consolidated into 4 industry segments × 2 income tiers = 8 personas (4 of the 6 groups merged into 2) so product and research could actually act on it.
• Found a wealthy tail: inflows rose roughly in step with size up to the 7th–8th decile, then shot up exponentially — split into 2 income tiers (lower ~5–6th decile, higher 7th and up) so high-value customers could be served differently.
• Bonus outcome: the segmentation method was reused to forecast overdraft sales from each segment’s credit-consumption pattern — and the forecast matched actual take-up.

The problem

What did we actually know about our small-business customers?

Almost nothing structured. The bank served ~160K of them — sole traders and tiny companies — but “small business” was treated as one blob. There was no segmentation, no map of which industries mattered, no sense of where the money really sat. That’s a problem when you own the UX: you can’t target research, you can’t profile products, and every “who is this for?” question gets a shrug.

Why not just run five interviews and call it personas?

Because persona decks built on a handful of conversations are the first thing people stop believing. They read like design opinions. I needed something the product managers, the analysts, and the business couldn’t wave away — and that meant grounding it in data they already trusted: the transactions.

My role / scope

• I drove the analysis and the entire thinking process — which data to pull, how to cut it, what the groups meant. Full intellectual ownership of the method.
• I didn’t have direct database access; an analyst ran the SQL against my specifications. I owned the questions, not the queries.
• I was the sole UX owner for the small-business segment, working with PMs, dev teams, and a separate UX research team.

Process

• Segment by industry code. Cut ~160K customers by their registered business-activity code — the public classification that says what industry a company operates in. Two axes per group: what share of customers they were, and what share of monthly inflows they brought in.
• Find the structure. Six groups accounted for 75% of customers and 80% of inflows. Four of those six turned out to be close enough to merge into two, leaving four real, distinct segments.
• Find the wealthy tail. When I plotted the data, inflows rose roughly in step with company size up to the 7th–8th decile — then took off exponentially. That’s the signature of a wealthy tail hiding inside “small business.” So I crossed each of the four segments with two income tiers — lower (~5–6th decile) and higher (7th decile and up). Same industry, fundamentally different money: eight personas, each industry in a lower- and a higher-income variant.
• Read the behavior. Per group, I looked at what people actually did with their accounts. A few patterns stood out (kept broad on purpose): in one group, mobile payments dominated regardless of company size; in another, roughly every third larger company carried a credit line, paid it down, and drew again; one specific group opened savings deposits far more than the rest.

Final structure — 4 industry segments × 2 income tiers:

Key decisions & trade-offs

• Build personas from data, not interviews. The unconventional call. Trade-off: you lose some of the warm, qualitative texture that ethnographic interviews give you. Payoff: personas that are defensible, reproducible, and impossible to dismiss as “just design’s opinion.” Interviews still happened — but now we knew who to interview.
• Four segments, not six — then doubled by income tier. I could have kept all six groups distinct. I merged four into two because a segmentation people can’t hold in their head isn’t actionable. Four segments was the ceiling; the income split then doubled it to eight personas — a tidy eight (four industries × two tiers), not a sprawl.
• Two income tiers — because the curve said so. The exponential tail wasn’t an assumption; it was in the data. Splitting customers at the 7th–8th decile into a higher tier (7th and up) versus a lower one (~5–6th decile) meant product profiling could treat the wealthy tail differently instead of averaging it away.
• Honest constraint: I owned the questions, not the queries. No direct database access — an analyst pulled every number to my spec. That shaped the work: the rigor lives in the method and the interpretation, not in me writing the SQL.

How it turned out

The personas became the backbone of how the segment was worked. Three concrete uses: research recruitment could finally target the right industries instead of whoever answered; product profiling aimed high-cash-flow offers at the segments that matched — construction and production, transport, wholesale and retail, and office/professional services — each in a lower- and a higher-income variant. And the stretch: the same segmentation method was reused to forecast overdraft sales from each segment’s credit-consumption pattern — and the forecast held up against what customers actually took up. A UX method that predicts revenue is, in my experience, vanishingly rare.

What I took away

• A persona is only worth what it’s grounded in. Sticky notes get ignored; transaction data doesn’t. Grounding personas in real money flow changed how seriously the rest of the business took them.
• The shape of the data tells you how to cut it. The exponential tail wasn’t something I went looking for — it showed up when I plotted the curve, and it demanded its own tier. Read the data before you decide the buckets.
• Segmentation is worthless if it isn’t actionable. Six groups collapse to four segments because a map nobody can memorize is a decoration. Usability applies to analysis too.