The 4-Lens Scan in Practice: Ninety Seconds Before the Decline Button
Simon has spent fourteen years in SME lending at a bank in Singapore, long enough to have watched three generations of credit models come and go, each one smarter than the last. The newest is genuinely good: it reads cash-flow patterns, supplier networks, even the seasonality of a business’s neighborhood, and its recommendations arrive with a confidence score and a pre-filled letter. On a Tuesday morning it recommended declining Mr. Tan.
Mr. Tan has banked with them for twenty years. He runs a laksa stall his father ran before him, employs two people, and wants a modest loan to renovate before his lease renewal. The model’s reasoning, to the extent the dashboard revealed it: thin digital footprint, cash-heavy revenue, low e-invoice volume, an aging customer demographic. The confidence score was high. The decline button was one click away, and Simon had a queue.
Chapter 4 of the book proposes a discipline for exactly this click: the 4-Lens Scan — ninety seconds, four questions, before accepting any algorithmic recommendation about a human being. Simon ran it with a pen on the back of the printout.
Stakeholders — who becomes invisible when we optimize?
Mr. Tan, obviously. But also: the two employees. The regulars who’ve eaten there for decades. And every other cash-economy business in the queue behind him — hawkers, provision shops, the businesses this bank was originally built on. The model doesn’t see them; it sees the absence of QR codes.
Bias Check — what “normal” is baked in?
The model’s idea of a healthy business is digital-native: e-invoices, delivery-platform revenue, a customer base that pays by phone. That’s a real pattern — and it quietly defines a sixty-eight-year-old hawker with a forty-year track record as an anomaly rather than as a different kind of reliability.
Long-Term Ripples — what future does this click build?
If every thin-file business reads as risk, we exit a whole economy one click at a time — and we do it without ever having decided to. Nobody will have made that choice. It will just have happened, letter by pre-filled letter.
Inner State — what’s driving me?
This is the lens that keeps the scan honest, because it points backward.
I’ve overridden the model twice this quarter already, and I’m aware of how that looks. Some of what I’m feeling is loyalty to a customer I like. Is that judgment, or nostalgia? The first three lenses say it’s judgment. But I needed to ask.
Ninety seconds. Then Simon did something more useful than quietly overriding the recommendation: he attached the four answers to the file and escalated it as a written case — not “I have a feeling about Mr. Tan” but “here is what the model cannot see, lens by lens.” The loan was approved on review. More consequential was what happened next: his write-up traveled upward, and within a quarter the bank had a formal “thin-file review” lane — a human path for long-tenure, cash-economy customers whose reliability lives in decades rather than in data exhaust.
The scan’s promise, the book says, is that with practice it becomes as automatic as checking your mirrors before changing lanes. Simon’s version of that: the model now does what models do well — flag, sort, accelerate — and ninety seconds of lenses stand between its confidence and anyone’s livelihood. The queue still gets cleared. It just gets cleared by someone who can say, for every decline, who he checked for in the mirrors first.
Simon is a composite — drawn from conversations with bankers and lenders navigating algorithmic credit decisions, with details changed. Mr. Tan’s laksa, however, deserves to be real.
Try it yourself
Before you accept the next AI recommendation that touches a real person — a hire, a loan, a flag, a ranking — take ninety seconds and write one line per lens: who becomes invisible, what “normal” is assumed, what future repeated clicks build, and what’s actually driving you. If the recommendation survives all four, click with a clear conscience. If it doesn’t, you’ve just found the paragraph to escalate.
Comments