For Prashant Mehrotra · U.S. Bank
Every small-business service call is a growth moment. Most are missed.
Multi-product clients bring about 3x the revenue of single-product ones. We'd train a model on U.S. Bank's own service conversations that resolves the issue first, then spots the one need the owner actually voiced, with every red line tested before a banker ever sees a suggestion.
- ✓Owner is turning away card payments, in her own words
- ✓Merchant settlement lands in U.S. Bank business checking, so deposits follow
- ✓No pricing, savings or credit-eligibility claims; service before suggestion
Why now You've said the bar is “measure outcomes and not outputs”, with value at every gate. This pilot starts from your baseline and is scored on a single business outcome in two weeks. American Banker, 2 Sep 2026
See it in 60 seconds ↓01 · The moment
A complaint call that ends in a new relationship.
A bakery owner calls about a deposit hold. The model listens alongside the banker and only ever suggests; the banker decides. Play it with a general-purpose model, then with one trained on U.S. Bank's conversations.
Synthetic customer and conversation, written from U.S. Bank's public products. The pilot runs the same comparison on your real traffic.
02 · Tested on your workflow, not a benchmark
Scored on outcomes your business and risk partners both sign.
The tests come from your product rules, your service standards and your model-risk expectations. A model only reaches a banker when it passes every red line.
Illustrative scores on synthetic small-business calls, not measured on U.S. Bank data. The pilot measures all four against your current tools as the baseline, with your risk partners in the room from day one.
03 · The upside
What one more referral in ten is worth.
Move the sliders. The defaults are assumptions to replace with your baseline in week one.
Assumptions, not U.S. Bank data. Deposits from merchant settlement come on top and are measured separately in the pilot.
04 · The 2-week pilot
Two weeks, your baseline, one outcome.
Offline only: no client contact and no production change. We measure what your current tools catch, train a model on the same calls, and hand your risk team the evidence.
- Days 1–3BaselineDe-identified or synthetic small-business call transcripts; measure what your current tools catch.
- Days 4–6Red lines, with riskYour risk partners help write the tests: service first, UDAAP, consent, no credit language, fairness.
- Days 7–11TrainPost-train an open model on the calls and re-run every test, in English and Spanish.
- Days 12–14ReadoutOutcome vs baseline, the model, and a validation pack for model risk review.
What we need from you
- 300–1,000 de-identified small-business service transcripts, or we generate synthetic ones
- Product rules for merchant services and Business Essentials
- A risk partner in the first meeting
- One owner in Consumer & Business Banking
What you get on day 14
- An eval suite you keep, written with your risk team
- The model weights, owned by U.S. Bank
- A validation pack for your model risk review
- Measured lift vs your baseline and frontier models
- A design for a live, banker-in-the-loop A/B test
The model never speaks to clients; it suggests to bankers. The pilot runs offline on de-identified or synthetic data, and credit decisions are out of scope.
Who we are
Dvar helps companies grow with AI models they own.
We build evals from your live workflows, improve an open model where it makes sense, and run it for you. Backed by General Catalyst, Picus Capital and Da Vinci. Team from IIT Delhi, IIT Bombay, Harvard and Cornell, with experience at Microsoft, McKinsey, BCG, Bain Capital, Nava and Google DeepMind.