Prepared for RAKBANK Business Banking
Turn half-finished SME applications into funded accounts.
RAKBANK added 12,000 SMEs in nine months. The next few thousand are sitting in stalled applications. We'd train a model on RAKBANK's own onboarding so every stalled owner gets a clear, compliant nudge in their language, and every relationship manager gets a ready-to-approve file.
- ✓Name mismatch: passport vs trade licence
- ✓Arabic greeting, English chat, Malayalam steps for the partner abroad
- ✓Human offered, reason explained, no pressure in the offer
Why now RAKBANK's new promise is “Digital with a Human Touch”, and the AI team is hiring for LLM-as-judge evals on English and Arabic journeys. That is exactly what this pilot builds. Rebrand, 9 Sep 2026 · RAKBANK careers, Sep 2026
See it in 60 seconds ↓01 · The moment
Day 9 of a stalled application. Two ways it goes.
A two-partner trading company in Deira started Quick Apply and stopped at documents. Play the follow-up with a generic assistant, then with one trained on RAKBANK's onboarding.
Synthetic customer and conversation, written from RAKBANK's public products. The pilot runs the same comparison on your real traffic.
02 · Tested on your workflow, not a benchmark
We grade every reply the way your RMs and compliance team would.
The tests come from Quick Apply's document rules, real stall reasons and the CBUAE's February 2026 AI guidance. A model only goes live when it passes them.
Illustrative scores on synthetic Quick Apply cases written from RAKBANK's public products, not measured on RAKBANK data. The pilot measures all four, and the baseline is your current assistant.
03 · The upside
What a few more finished applications are worth.
Move the sliders. The defaults are assumptions to replace with your funnel numbers in week one.
Assumptions, not RAKBANK data. Revenue per SME would include balances, FX, payroll and cards; the pilot replaces every default with your own figures.
04 · The 2-week pilot
Two weeks to show how many stalled SMEs you can win back.
Week one builds the tests and measures what you run today. Week two trains a model and re-runs every test. You get the numbers, and the model.
- Days 1–3Build the testQuick Apply rules and common stall reasons become an eval suite in English, Arabic, Malayalam and Urdu.
- Days 4–6BaselineScore your current assistant and the best frontier models on the same cases.
- Days 7–11TrainPost-train an open model on the cases and re-run every test, including the CBUAE red lines.
- Days 12–14ReadoutScorecard, the model, and a funnel business case for Business Banking.
What we need from you
- Quick Apply document rules and the top stall reasons
- 50–100 example conversations, synthetic or de-identified
- Two review sessions with an RM or onboarding ops lead
- One owner from Business Banking
What you get on day 14
- An eval suite in four languages that you keep
- A scorecard: trained model vs your assistant vs frontier models
- The model weights, owned by RAKBANK
- A funnel business case on your numbers
- A plan for a live test on stalled applications
The pilot runs on synthetic cases by default, so no customer data moves. Real conversations come in only after your security review, and only inside a UAE environment you control. A human is always one reply away for the customer.
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.
