Eligibility, underwriting & benefits
On this page
"Does this applicant qualify?" is a checklist of criteria, whether for a loan, a policy, a benefit, or a plan. Encoded as rules, that checklist runs in an instant and gives every applicant the same answer for the same facts.
The stakes🔗
Automating qualification decisions transforms both speed and cost. Properly designed automated underwriting lets insurers process up to ~90% of applications straight-through (and lenders even more), turning decisions that took days or weeks into minutes or seconds, with reported ~70% reductions in processing time and ~40% in cost (ScienceSoft). Beyond speed, rule-based decisions are consistent and auditable: the same policy for everyone, written down explicitly. That matters when a regulator asks why an application was declined.
How it works🔗
Qualification criteria become rules, with exclusions stated explicitly. For example, "applicants aged 18–75, earning at least $30,000, in credit bands 1–3, and not in the excluded region 9, qualify." Winnow can also check applications as they arrive: required fields present, categories allowed, numbers within permitted ranges, so malformed data is caught before it ever reaches a decision.
🤖 With AI in the loop
AI underwriting is a model wrapped in rules. Upstart, which fully automated 91% of its loans in 2024, applies each bank partner's own underwriting policy (minimum credit scores, loan-amount limits, maximum debt-to-income) as hard knockout rules before an application reaches its models. Zest AI auto-decisions about 80% of applications by putting lender credit policies and cut-offs around its model scores. That is the same layer that produces the adverse-action reasons U.S. lending law requires for every automated decline. In Winnow the model's score is one column among many; the knockouts, caps, and exclusions around it are explicit text, auditable line by line, and adjustable the day policy changes. More patterns: Deep Dive: AI + Rules Engines.
Why Winnow fits🔗
- Criteria map directly: age, income, and tenure are
number checks; credit band and product are
category checks; exclusions are
NOT; everything combines withAND/OR. - Relative checks too: a rule can compare one field against another
(
requested_amount <= approved_limit) directly. - Consistent and explicit: every applicant is judged by the same written rules, so decisions are reproducible and explainable.
- Validated input, instant output: bad data is rejected on arrival; the decision itself takes microseconds and stays consistent even while the criteria are being updated.
🧪 In simulation — 150,000 qualification rules, 1 million applications
To put numbers on the above, we built a qualification book across 2,000 products — 150,000 rules: per-product criteria like the example above, income-to-amount ratio tiers that compare one field of the application against another, regional programs with explicit exclusions, hard knock-out rules, and a small set of validation rules that catch malformed data on arrival — then decided 1 million applications against all of it, money held as exact amounts. 78% of applications qualified for at least one product, 10% were stopped by an explicit knock-out, and 1.9% were malformed and never reached a decision at all. Every verdict names the exact rules that produced it — which is the paper trail an adverse-action notice needs.
What this is: a test run on a generated qualification book and generated applications, not a customer deployment — the criteria and applicants are invented, the speeds and costs are measured. It ran on one ordinary 16-core desktop computer, priced at a standard cloud rate of $0.043 per processor-hour running around the clock, the same basis as the Deep Dive: Cost Savings. Every decision was double-checked against a slow, exhaustive pass over all 150,000 rules before any stopwatch started — that slow way takes about 6 times longer — so no qualification (and no decline) comes from a rule quietly skipped. Treat the speeds as approximate, moving by about a fifth with machine load, and the counts as exact; the test ships with Winnow for your team to re-run on your own criteria.