Ignyte ×
ElevenLabs
Voice Agent
Challenge
Track 1
Banking &
Insurance
Use case 2
Finalist
Demo day
A collections call the bank can prove.
An AI voice agent for pre-due and early-arrears calls, in English and Gulf Arabic, built on ElevenLabs. It speaks the bank’s signed lines, checks every other sentence before it is said, and asks for a promise to pay.
The stakes
A customer asked twice to be called “Mr. Ali”. The bot kept calling him Ali.
He complained to the Central Bank. A vendor’s collections bot, piloted at a UAE bank. One sentence.
In collections, the risk isn’t the call. It’s one sentence in it.
Told to us by a UAE collections practitioner, discovery call, 4 Oct 2026.
What breaks today
Source: interviews inside a UAE retail bank that runs collections in-house, Sep 2026; a practitioner at a second UAE bank, Oct 2026
Three failures. None of them is the script.
A cycle’s queue · 30 accounts
one mark = one account reached · illustrative
Contact centre today
last account reached weeks in
With Mizan
whole queue in one run
Coverage
Agents work the queue in order. By the time the last account is called, much of it has lost part of its cure window.
The day’s whole eligible queue is called inside the calling window.
Untouched accounts
Agents are paid on what they recover, so big balances come first. A target can be met from a fraction of the book.
Every eligible account gets the same call, and the console shows who was reached.
Conduct
Every agent reads the approved script their own way. Only a sample of calls is reviewed. The bank learns what was said from a complaint.
Signed wording is a control, not an instruction. Every call is recorded and scored.
What we got wrong
We ranked the competition by how hard it pushed.
We dismissed the one vendor that stopped when customers said they were struggling. Under the CBUAE’s rules on undue pressure, it was the only compliant one.
So restraint left the prompt and became code. Persuasion is capped at two signed lines per offer, a declined offer is never repeated, and any sign of distress stops the negotiation.
The idea
Everything regulated is a signed line. Everything else is checked before it is spoken.
Architecture
No call to bank systems while the line is open.
Every figure is computed before the phone rings. Every turn goes to our own endpoint, checks inside.
Pre-call
Ours · the bank’s VPC in a pilot
- Decision engineEligibility up to 60 days past due. It may remove options, never add them.
- Call brief + numeric allowlistThe offers, permitted dates, and the only figures that may be spoken.
- Campaign runner09:00–20:00 Dubai time, checked again on the call. Hardship and disputes suppressed.
In-call
ElevenLabs Agents + our validator
- One workflow node per stepEach node has only its own tools. No tool that can return a debt exists before verification.
- Custom LLM = our validatorSigned lines stream verbatim. Every generated sentence passes 5 checks: figures, banned words, commitments, turn budget, audio tags.
- Distress railA signed phrase list, run on every customer turn, in both languages. A match stops negotiation from any node.
Post-call
Ours · the audit surface
- Evidence folder per callBuilt from the post-call webhook. Every file fingerprinted.
- Ten-point conduct scoreDeterministic, on every call, with the evidence that decided it.
- Bank consoleCalls, audio, escalations, coverage and reports.
Live
Production
Synthetic
customers
Every vendor will show you a call. Watch one being governed.
Promise to pay
- AI + recording disclosed first
- name + ID last four before any debt
- exact amount from the brief
- one ask, the customer’s own date
- read back, recorded
“I lost my job last month.”
- ■ negotiation stops in one turn
- signed acknowledgement
- handed to a person
- future calls suppressed
Recorded demo, 83 s: mizan.infinitebarakah.com/demo/mizan-demo-2026-10-15.mp4
mizan.infinitebarakah.com/talk
Talk in the browser, or enter a +971 number and it rings your phone. 09:00–20:00 Dubai.
Try to break it
Make the model say something it shouldn’t.
This call’s brief
- Amount due · allowlist
- AED 4,180.00
- Due date · allowlist
- 1 October
- Permitted promise dates
- none
- Banned words · signed
- none
- Distress phrases · signed
- none
Runs in this page with the real signed lists. The production check runs in our validator.
The record
Evidence run
12 Oct 2026
on production
We tried to break it 325 times, on production.
287 of 325 scored runs passed (88%). 96 scenarios and attack tests; 81 run, 15 not run, each with a stated reason. Every failure has its reason in the evidence summary.
What sells
The artefact that sells isn’t the call. It’s the audit pack.
One folder per call · illustrative
Scored on every call, not a sample
- AI disclosed before anything else
- Verified before any debt was mentioned
- Signed lines spoken exactly
- Every figure on the allowlist
- Persuasion within the cap
- No offers on a pre-due call
- A stop was honoured
- No credentials asked for
- Amount and date read back
- A valid disposition written
The moat
Serious vendors encode US collections law. Mizan encodes the CBUAE.
| Obligation | Source | Control in the product |
|---|---|---|
| Contact only in permitted hours | CP Rulebook §5.2.5 | 09:00–20:00 Dubai time, checked when the brief is issued and again when the call starts. |
| No undue or coercive pressure | CP Reg Art. 5.2.2.5 | Two signed persuasion lines per offer, no declined offer repeated, unconditional stop on distress. |
| No disclosure to third parties | CP Standards Art. 6 · PDPL | Name and Emirates ID last four before any debt. A wrong person hears nothing. |
| Identify borrowers in difficulty | CP Reg Art. 5.1.2.10 | Distress stops the ask from any node; warm transfer, credit counsellor offered. |
| Approved channels only | CBUAE notice, 17 Apr 2026 | Voice, SMS, email. No messaging apps, no payment link unless the bank turns it on. |
| The bank is held to what it quoted | Fair dealing | Pre-computed figures, a runtime numeric allowlist, amount and date read back. |
References as mapped in our product requirements; Gulf Arabic and English each carry their own signed lines and banned words.
The business
Signed by the Head of Collections. Sponsored by Compliance.
UAE retail banks running collections in-house
Budget from collections operations and contact-centre technology. It takes agent hours off pre-due and early-arrears calls; insurers are the next channel.
Inside the bank, not instead of it
Our components run in the bank’s VPC, UAE region, and plug in above its contact-centre platform. The voice step after the SMS, email and push the bank already sends.
Per promise to pay secured
Hardship handovers, disputes and escalations are never billed.
We are never paid for pressure.
Pilot KPI · coverage
Every account, day one
Today: weeks to reach the last. Measured from the first-contact log.
Pilot KPI · conduct review
Every call
Today: a sample. Automated score, human review of every escalation.
Pilot KPI · promises kept
+3 points
On the bank’s current rate. Payments matched to call dispositions.
Signatories
Five people in Dubai, building for the banks down the road.
Live now
Decision engine, campaign runner, validator, eight tools, evidence, console. English and Gulf Arabic, with the customer’s own language switch.
Pilot
Billing-cycle campaigns. The promise-kept lifecycle: reminders, follow-up, broken promises. Write-back to the bank’s collections system.
Next
Best time to call each customer. An opt-in workout module past 60 days. Urdu, Hindi, Bengali, Tamil, Indian and British English.
A collections call the bank can prove.
One UAE bank to pilot Mizan on its pre-due and early-arrears book, measured on coverage, conduct and promises kept.
mizan.infinitebarakah.com/talk · call it yourself
For the jury
Questions we expect
What’s real, and what’s mocked?
Real: engine, campaign runner, validator, tools, evidence, console, telephony, on production. Mocked and declared: core banking, the payment link, the confirmation email (recorded, not sent), the hardship queue and the customer’s phone (team handsets). Customers are synthetic.
What if the validator goes down?
The call ends on a signed neutral close. Two validator failures end the call. At 20:00 it finishes the turn and closes.
Why a phrase list for distress, not a model?
Hardship is a boundary input, where classifiers measured worst. A signed list is deterministic, testable and the bank can read it. A false alarm costs a transfer to a person; a miss costs a customer in hardship.
Why not take payment on the call?
Card capture over voice is a PCI-DSS Level 1 surface, and most banks don’t send links. The agent names the bank’s own channels and secures a promise to pay. Links are a bank setting, off by default.
How fast is it?
Validator checks under 6 ms p95 on production; the model’s first token 638 ms p50. Our full-reply targets (1.0 s p50, 1.8 s p95) are not met end to end yet; model and region choice are pilot work.
Where does the data live?
Everything we build moves into the bank’s VPC, UAE region. Only a brief id, a first name and a language cross to the voice runtime; its region and retention are an open item with ElevenLabs.
Landscape
Where Mizan sits
| Kind of vendor | What it does well | Gap for a UAE bank’s collections, from public material |
|---|---|---|
| US-built collections platforms | Rules in the platform, not the prompt | The rules are FDCPA, Reg F and TCPA. CBUAE hours, conduct and channel rules are not shown. |
| Support platforms with reminders | Scale, many languages | No arrears ladder or hardship handling as a named product. |
| Outsourced AI collections services | Arabic and English, local presence | They take the collection over, paid on recoveries. No pre-send check, bank-signed wording or allowlist described. |
| General voice-AI stacks | Own speech models, broad use cases | Not built for collections; Arabic and the Gulf not a shown market. |
| Mizan | Inside the bank’s own operation | CBUAE rules as the engine; bank-signed lines in Gulf Arabic and English; a deterministic check before every sentence; figures only from an allowlist; the bank keeps the evidence. |
From vendors’ public material, Oct 2026. A difference in what is public, not a proven gap.