A homework sketch · AI underwriting and claims under exam · prepared for Corgi
I'd love a short call to hear what Corgi actually needs. You're swamped between rounds and new lines, though, so I did a little homework first: a few small prototypes, built only from your public footprint and synthetic stand-in data, on a problem the whole industry is wrestling with this year.
They're rough sketches, not your data and not a finished product; a faster way to show what working together could look like than a blank-page call. They all circle one question an examiner asks first: when an AI declines a risk or denies a claim, can you run the rule that made the call? If it lives only inside a model, no.
All three run in the browser on synthetic or public data, labeled on screen; nothing you type leaves the page.
A real share of your model-made calls should be a rule.
A non-trivial slice of model-routed decisions are compliance-bearing: an examiner can ask you to run the rule, and there is none. The audit counts them; if it finds zero, the hunch is wrong and the work is priced to that.
The exposure compounds faster than you can architect for it.
Each new state and vertical multiplies the decision-by-jurisdiction surface, faster than a 100-person team can hand-build the governance layer before the next launch.
The fairness gap is testable, and fixable without a rebuild.
A protected-class denial is examined on its outcome gap, not the model's internals. A calibration layer compresses it while holding accuracy; the engagement runs the test on your outcomes.
Governance platforms (Credo AI and the like) give you a monitoring dashboard and a policy template, and do that well. What a dashboard structurally can't tell you is which specific decision should not be a model in the first place. A Big Four project will, at a long price, from people who mostly haven't shipped ML. I ship ML, and the argument here is where not to use it. The real value isn't capability your team lacks; it's independence: an examiner discounts the model team grading its own homework.
A fixed-scope diagnostic, six to eight weeks. It produces the decision inventory an AIS-Program exam asks for first, the classification, a rules-in-front-of-model re-architecture spec, and one regulator-readiness walkthrough.
Fixed-scope; indicative, final scope set after a call. If the audit surfaces no compliance-bearing decision an examiner would challenge, you hear that in the call. Honest bound: it doesn't replace your model team and never makes a decision without a human in the chain.
One 30 to 45 minute call. Bring one underwriting or claims decision your model makes today and the state it ships in. I'll map it against the AIS-Program request list live and tell you, in the call, whether an examiner could ask you to run a rule you don't have. If there's nothing there, I'll say so.
Book it: jeffpinto.com/engage · Method: the rules-before-model note
Jeff Pinto runs a small, independent data and AI advisory practice (jeffpinto.com). Thirty years across AI data and privacy, health tech, marketing analytics, renewables, logistics, and broadcasting; the last seven in ML and AI. Hands-on at Meta, Uber, and IBM, plus six startups (one turnaround, three acquisitions). Two MScs: CS (Toronto) and engineering (Loughborough). Engagements are fixed-scope, four to twelve weeks, no platform and no subscription; whatever gets built, the IP transfers to you. The edge for this one: my UofT/CAMH master's thesis took a 35-point subgroup parity gap on a tiny clinical corpus down to roughly one point, no accuracy hit. A protected-class denial in underwriting or claims is the same shape, which is what the disparate-impact sketch does.
Sources · Corgi disclaimers (producer entity, 45 jurisdictions) · NAIC AI Model Bulletin · Colorado SB21-169 · 35-to-1 parity is Jeff's published UofT/CAMH thesis. Synthetic and illustrative figures labeled as such in each prototype; no Corgi-internal number appears.