Nbyula × UniCreds — Shared Loan Book Review & Automation Proposal

355students on the shared dashboardthe book being worked

165never dialled once46.5% of the book

44given a real conversation12.4% — 5 minutes or longer

21reached a lender stage5.9% — 8 of them from this book

The whole book in one diagram

Where the 355 stop moving

One dot = one student

How far UniCreds got with each of the 355

165never dialled53no connect93under 5 min445-min+

The book is not cold. The follow-up is.

165 of 355 students have no dial recorded against them. Another 53 were dialled and never answered. Only 44 ever got a conversation longer than five minutes. And 11 of the 19 students our AI classified “Loan — Doable / Fast-Track” were never once connected with — none of the 19 converted.

Outcome evidence

Those 21 students, traced back to ours

Nineteen of the twenty-one resolve to a Nbyula CRM record. Eight of them came through the shared dashboard; the rest were handed over by other means. The lag column is the gap between the day the student entered our CRM and the day they appear on yours.

4 daysmedian handover lag through the shared pipe8 students · max 48 days

142 daysmedian lag when handed over by hand11 students · max 702 days

9disbursed across the whole relationship4 of them came via the shared pipe

2names with no Nbyula record at allAkash C · Priyanka P

Student	Country	Status	Agent	On their dash	In our CRM	Lag	Route	UC dials	AI dials

Call counts are only what the shared Nbyula dialer logged. Calls placed from UniCreds’ own systems never reach this log — four disbursed students show zero UniCreds dials here. Read every “never dialled” as “no attempt visible in the system we share”. The reverse blindness is just as real: our stage changes, AI call summaries and document status are invisible to UniCreds. That mutual blindness is what this proposal fixes.

Two leaks, in the order they cost money

Leak 1 — first contact. 46% of the shared book has never been dialled, and 21 students have had no attempt from either side, human or AI. Leak 2 — the follow-up. Contact arrives in bursts and then stops: the median student has had no logged UniCreds touch for 37 days, and 65 of them for more than sixty.

Leak 1 & 2 · the flags

Red flags across the 355

Counts overlap — one student can trip several. 328 of the 355 trip at least one.

Leak 2, in one chart

Contact arrives in bursts, then stops

April: one dial across 355 students. The AI agent went live in June and has not stopped since.

The pre-assessed queue nobody is working

Every AI loan verdict, against whether UniCreds ever connected

Each of these students arrives with co-applicant relationship and employment type, income, ITR availability, CIBIL band, collateral type and value, amount, university, intake and prior-rejection history already captured. Today none of it crosses the wire. 287 of the 355 carry no verdict at all — the queue has nothing to sort by.

They were not cold leads. They were pre-qualified files.

Our agent has completed a full loan assessment on 68 of the 355 — co-applicant, collateral, CIBIL band, amount, university, intake, prior rejections — and classified each one. 51 came back fundable or workable: 19 Doable / Fast-Track, 23 Borderline, 9 needing a human. Those 51 sat on the dashboard with the underwriting conversation already done.

Eighteen of the 51 were never dialled. Ten got a UniCreds conversation longer than five minutes. One reached a lender stage. None has converted.

The qualified cohorts, step by step

What happened to each verdict

Each row is scored against its own cohort, so the rates compare directly. Read left to right: classified → dialled → connected → a real conversation → a lender stage.

Who did the work on those 51

The discovery was already done — by the machine

Every one of the 51 reached a loan verdict because the agent held a real conversation with them: 48 of the 51 talked for five minutes or more, ten and a half hours of loan discovery in total. On the same 51 students the UniCreds team logged 1 hour 43 minutes. The profile, the co-applicant, the collateral and the CIBIL band were already captured and sitting in our CRM. Nobody had to ask for them again — and nobody did.

51students classified fundable or workable19 doable · 23 borderline · 9 human-intervention

18of those 51 never dialled once25 had a 5-min+ AI call and no UniCreds connect

13 daysmedian age of a qualified studentthe freshest leads in the book

1of the 51 has reached a lender stagenone has converted

These students are new — the agent only began classifying in June, so the cohort is a median of 13 days old and has not had long to convert. That is the point: 18 of them have not been dialled once in those 13 days, and a September-intake loan file cannot wait for the next calling burst.

The verdict is credit policy, not sentiment.

Every classification on the previous tab comes from a fixed rule set the agent applies silently before it speaks, then updates live as facts surface on the call. It is bucketing on one question: is there a viable lender path right now? Below is the whole gate, exactly as it runs in production.

The decision tree, as it runs

How a student becomes Doable, Borderline or Not doable

The intelligence layer

What tips a borderline case either way

Learned from the real lending pool, these sit on top of the base gates and decide which side a marginal file lands on.

pushes toward Borderline

Co-applicant CIBIL is the number-one killer. Impaired scores in the 490s–540s, bounced EMIs, guarantor on someone else’s loan, or a cluster of recent enquiries drops a case out of Doable until it is addressed.
FOIR breach — ratios above 160% have been seen — plus no ITR or an ITR filed purely for the loan, or a co-applicant near 60.
The student’s own strength is a genuine offset. A 730+ personal CIBIL with relevant work experience and good academics keeps a weak-co-applicant file in Borderline, not Not-doable.

pushes toward Not doable

Willingness to fix decides it. An active default the student will clear, with an NOC, stays Borderline. Unwilling to clear it → Not doable.
Backlogs, low CGPA and no work experience together. Any one alone is usually survivable; all three are not.
Location and prior rejections do not change the bucket — they change routing. Avanse Global will not process a Kerala case; a past Prodigy rejection closes Prodigy. The file is sent to lenders it has not already burned.

Before anything is written off

The rescue ladder

How the verdict leaves the call

One tool call, before the closing line

TOOL

classify_loan_lead

Fires once per call, before the agent speaks its closing line, with bucket set to doable, borderline or not_doable. This is the field we want to push onto your shared record.

OVERRIDE

alert_loan_lead_not_interested

A clear opt-out — not interested, arranged elsewhere, asks us to stop — overrides classification and carries a one-sentence reason. These should stop being shared with you at all.

OVERRIDE

request_loan_human_handoff

Distress, a legal or complaint thread, a value beyond the agent’s remit, or the student simply asking for a person. Nine students on the shared book are sitting in this state.

Classification runs silently before the first message, from the enquiry the student already submitted, and is revised in real time as the call surfaces new facts. By the time a verdict exists, the co-applicant, the collateral, the CIBIL band and the amount are already on file — which is exactly the payload the eligibility endpoint on tab 06 would consume.

The full shared book

All 355 students, most-flagged first

“UC” is UniCreds activity in the shared dialer; “AI” is the Nbyula loan agent.

Same 355 students. Same two months. 12× the dials.

1 June – 31 July 2026, like for like. Humans connect 34.9% per dial against the agent’s 15.8% — the people are better on the phone. They are simply not on it. The agent’s advantage is persistence, not persuasion, and persistence is exactly what a loan file needs between the eligibility check and a complete document set.

One dot = one student · same cohort, same buckets

Coverage, side by side

UniCreds team

932 dials · lifetime

165never dialled53no connect93under 5 min445-min+

Nbyula AI loan agent

5,025 dials · 2 months

73never dialled54no connect160under 5 min685-min+

Like for like · 1 Jun – 31 Jul 2026 · the same 355

Head to head

“Human team” is every human caller who touched these students in the window — Nbyula counsellors and UniCreds agents combined. UniCreds agents placed 411 of those 413 dials. Human talk time is connected-call length in the Nbyula log; AI talk time is completed-call duration in the voice log. Both exclude ringing.

What the persistence bought

Coverage the humans never got to

141students the agent reached that no UniCreds caller connected with

36of those held a full 5-minute+ loan conversation

27of those 36 had no UniCreds dial attempted at all

68students carry a structured loan verdict — every one produced by the agent

Evidence · unedited agent call summaries

The depth the agent is already reaching

All five had zero connected UniCreds calls at this snapshot. Names are from the CRM; the summaries are machine transcriptions, so spoken names inside them are sometimes garbled.

Md Sultan AnsariDoableAI 9 dials · 11 minUC 0 dials

“15–20 lakhs for a BSc Computer Science at Kielce University of Technology, Poland; payment deadline 31 July. Self-employed father now has ITRs and a Gumasta certificate and will be co-applicant; grandmother’s land worth 20 lakhs as collateral. Previously rejected for lack of ITRs — now viable.”

Ali FarisDoableAI 4 dials · 10 minUC 0 dials

“Unsecured 35–40 lakhs for an MPH at Birmingham City University, September 2026. Previously rejected elsewhere over a co-applicant brother’s past credit-card settlement; assessed viable given the cleared settlement, strong CIBIL and stable income.”

Regina Ramngaihzuali SailoDoableAI 4 dials · 10 minUC 0 dials

“MSc International Development, University of Bath, September 2026. Offer received, IELTS 6.5, needs 38–43 lakhs. No collateral; paternal aunt, government-salaried, CIBIL 760, proposed as co-applicant.”

Kammampally ShivakumarDoableAI 2 dials · 6 minUC 0 dials

“45 lakhs, MS Civil Engineering, University of Southampton, August 2026. No collateral or co-applicant; already holds an Empower sanction for the US but needs the UK. Prodigy Finance recommended; he agreed to proceed.”

Afroz MohammedNot doableAI 4 dials · 26 minUC 2 dials, no connect

“35 lakhs, MSc International Business, University of Surrey. Offer letter in hand, but low academic marks, parents’ income insufficient without ITRs, no collateral, and the potential co-applicant uncle carries an active overdue loan and high debt.” A saved call: 26 minutes establishing the file cannot be funded, so nobody on the UniCreds team has to.

Today vs proposed

The student journey, before and after

The loop we want to build

Landing page to disbursal, with two human touchpoints

Three blocks need an API from UniCreds. Everything else is ours to build.

What we need from you

Three endpoints

Shapes are a starting proposal. If your engine speaks a different contract, we build to yours.

POST/v1/eligibility/checkAPI 1

Student profile in → every lender that would fund it out, with indicative terms. Synchronous — a student is watching a spinner.

// request
{ "country": "UK",
  "university": "University of Bath",
  "course": "MSc International Development",
  "intake": "2026-09",
  "amount_required_inr": 4300000,
  "offer_status": "unconditional",
  "academics": { "ug_percentage": 72, "backlogs": 0 },
  "tests": { "ielts": 6.5 },
  "collateral": { "type": "none" },
  "co_applicant": { "relation": "aunt",
    "employment": "government_salaried",
    "monthly_income_inr": 65000,
    "itr_available": true, "cibil_band": "750-800" },
  "prior_rejections": [] }

// response
{ "verdict": "doable",
  "matches": [
    { "lender": "…", "product": "unsecured",
      "max_amount_inr": 4500000,
      "roi_range": [10.5, 11.75],
      "margin_pct": 0, "moratorium_months": 12,
      "processing_fee_pct": 1.0,
      "expected_tat_days": 12,
      "conditions": ["co-applicant ITR last 2 years"] } ],
  "declined": [ { "lender": "…",
      "reason_code": "NO_COLLATERAL_UK" } ] }

POST/v1/applicationsAPI 2

Creates the case and returns that bank’s document checklist — which becomes the student’s task list and what our agent chases against.

{ "application_id": "UC-2026-08841",
  "lender": "…", "status": "created",
  "checklist": [
    { "key": "kyc_student", "label": "Aadhaar + PAN",
      "required": true, "formats": ["pdf","jpg"] },
    { "key": "offer_letter", "label": "Offer letter",
      "required": true },
    { "key": "coapp_itr", "label": "Co-applicant ITR, 2 yrs",
      "required": true },
    { "key": "coapp_bank", "label": "Bank statement, 6 mo",
      "required": true } ] }

GET/v1/applications/{id}+ WEBHOOKAPI 3

The push matters more than the poll — it is what stops a student ever going silent again.

// POST https://api.nbyula.com/hooks/unicreds
{ "application_id": "UC-2026-08841",
  "status": "sanctioned",   // received | logged_in
                            // | pf_confirmed | sanctioned
                            // | disbursed | rejected
  "sanctioned_amount_inr": 4300000,
  "roi": 11.25,
  "rejection_reason_code": null,
  "updated_at": "2026-08-14T11:02:00+05:30" }

Plus the non-code parts: a sandbox, a named engineering owner, and an agreed rejection reason-code vocabulary — a rejection we can read is a case we re-route, not a case we lose.

Our half of the bargain

What Nbyula ships

This week, regardless of the API

A first-contact SLA on the shared book. No student sits on the dashboard undialled — today 165 of 355 do. Anything untouched at 72 hours returns to the agent automatically.
Push the AI loan verdict onto every shared record, so the queue sorts by fundability instead of arrival order.
Push the agent’s call summary and captured profile so nobody re-collects what we already have.
Suppress the dead ones. Not-doable and not-interested stop being shared — 37 fewer wasted calls today.

Built on your three endpoints

The public eligibility page. 773 students already arrive through loan-eligibility-check with nowhere to land.
Lender selection → Space → checklist, with upload, verification and a live completion percentage per file.
The agent on document follow-up until the checklist is complete, escalating only on the exceptions it is told to escalate.
A 72-hour cold-case rule. Any Doable student with no movement for three days returns to the agent automatically — on today’s book that single rule catches 10 of the 19 fast-track cases that went cold.
Status rendering in the Space and on the shared dashboard, from your webhook.

Sequencing

Three weeks, each shipping something usable

Week 1

Eligibility

Sandbox and contract agreed. We ship the landing page and run all 355 students through the endpoint in batch — you get a ranked, pre-scored book on day one, before any student-facing code is live.

Week 2

Application & checklist

Lender selection, application create, checklist rendered in Spaces, upload and verification. The agent’s document chase goes live on the fast-track cohort first.

Week 3

Status loop

Webhook consumed, statuses render for student, counsellor and dashboard. Rejection codes route failed cases back into eligibility against a different lender. The loop closes.

The arithmetic we stand behind

UniCreds has 21 students at PF Confirmed or better and 9 disbursed across the whole relationship. Only 8 of the 21 came through this shared book of 355 — and just 4 of the 9 disbursed.
165 of 355 were never dialled and another 53 were never reached. Nearly two-thirds of the book has never had a conversation.
11 of 19 fast-track cases were never connected with, and none of the 19 converted. The nearest thing to free revenue here is the students you already have.
The agent already produces the eligibility payload for 68 of the 355. Today that work is discarded at the handover.

What we are not claiming

Not that the AI closes loans. It qualifies, counsels and chases documents. Your team and your lenders close.
Not that the team is underperforming. Per dial they connect twice as well as the machine. The failure is structural.
Not that our call log is complete. Work done from UniCreds’ own systems is invisible here — which is exactly why the status API matters as much as the eligibility one.
Not a revenue forecast. We will not model a conversion rate we have not earned.

Leaving the call with

Four decisions

Does the eligibility engine exist as a callable service today? If yes — sandbox and contract. If not — what would it take?
Who owns the integration on your side, and can they start next week?
Will you take structured, document-complete files in place of raw leads — and does that change what a file is worth?
Will you commit to a 72-hour first-contact SLA on anything the agent classifies Doable — and let us return untouched cases to the agent automatically?