Find your biggest AI opportunities in under 30 minutes.Book a consultation
Back[ Coventa ]Case Study

[ RELATIONSHIP BANKING ]LIVE

Hub for Real-time Customer Intelligence in Banking

Guidance during the conversation, not after it. Hub transcribes live interactions across 50+ languages, extracts intelligence as they happen, and suggests hyper-personalised next best actions drawn from comparable prior interactions and their outcomes — while capturing notes and summaries automatically.

HUB // LIVE GUIDANCELive call · M. Laurent
Hub Advise/Live call · M. Laurent
Human approval requiredLast sync 12s ago

Queue

00:04:12 · liveChurn signal

"…thinking about moving the business account, the fees have crept up…"

00:02:40Expansion

"…we just signed a lease on the second site…"

00:01:15Renewal

"…the mortgage renewal letter arrived…"

Drafted reply · grounded in ontology

Suggested now: acknowledge the fee concern directly — she raised it in March. You can move her to the Growth tier today (saves ~$140/mo) given the new site.

Then: the second site implies working capital. She's pre-approved for a $250k overdraft; mention it once, don't push. Mortgage renewal can wait for a follow-up.

Mark as usedCapture outcomeDismiss

Customer 360

Client
Marie Laurent · SME owner · 11 yrs
Relationship
Business a/c · mortgage · 2 savings · $1.4M
Prior interaction
Fee complaint Mar 2026 · partial waiver
Detected
Expansion → working capital need
Eligibility
Fee tier upgrade · pre-approved OD $250k

// Problem

The Problem

The most valuable moment in relationship banking is the live conversation, and it is the moment the bank supports least. A relationship manager on a call is working from memory and whatever the CRM screen shows, which is a transaction list rather than an understanding of the client. The relevant precedent — what was recommended to a similar client in a similar position, and whether it worked — exists across thousands of prior interactions and is unreachable in real time. Afterwards, notes are written from recollection, which is when the detail that mattered is lost.

  • Guidance arrives after the interaction, when the opportunity to act on it has passed.
  • What worked for comparable clients is buried in interaction history nobody can query live.
  • Note-taking competes with listening, so the record is thin and the client repeats themselves next time.
  • Multilingual operations fragment the approach, because support quality varies by language.

// Overview

Hub equips relationship managers and service teams during the interaction itself. Audio is transcribed in real time with support for over fifty languages, and intelligence is extracted as the conversation proceeds. Against the ontology — a contextual understanding of the customer and their anticipated behaviour — the system suggests next best actions and data-informed financial guidance drawn from similar previous interactions, the recommendations made in them, and the outcomes that followed. Information capture runs alongside: notes, interaction summaries and customer reactions are recorded automatically for downstream analytics rather than reconstructed afterwards. The business rules underneath are configurable, so the same engine serves retail, corporate and private banking, and transfers to adjacent contexts such as claims management or incident resolution.

// AI System

Why AI

Real time is the constraint that decides the architecture. Transcription, extraction and retrieval all have to complete inside the pause in a conversation, which rules out any workflow with a human in the middle of the loop. What makes the suggestion trustworthy rather than generic is grounding: recommendations are drawn from this institution's own interaction history and the outcomes it recorded, so the guidance reflects what has worked here. The relationship manager remains the one who speaks — the system suggests, never acts — which is the only defensible arrangement when the output is financial guidance.

// Specs

Specifications

TRANSCRIPTION
Real-time audio-to-text, 50+ languages
GUIDANCE
Next best actions grounded in comparable prior interactions and outcomes
CAPTURE
Automated notes, summaries and customer reactions
ONTOLOGY
Evolves per interaction, so guidance sharpens per customer
CONFIGURABILITY
Retail, corporate and private banking; transferable to adjacent contexts
AUTONOMY
Suggests to the operator; never speaks or acts on its own

// Features

Features

  1. 01Live transcription and intelligence extraction across more than fifty languages.
  2. 02Hyper-personalised next best actions surfaced during the interaction, not after it.
  3. 03Recommendations grounded in similar prior interactions and their recorded outcomes.
  4. 04Automated note-taking and interaction summaries, freeing the operator to listen.
  5. 05Customer reactions captured for downstream analytics and intelligence extraction.
  6. 06Configurable business rules spanning retail, corporate and private banking contexts.

// Architecture

Architecture

INTERACTION FLOW

Runtime · one item, left to right


  1. 01Live Audio
  2. 02Transcription + Extraction
  3. 03Contextual Match + Next Best ActionCustomer ContextPrior InteractionsRecorded OutcomesBusiness Rules
  4. 04Operator Guidance
  5. 05Automated Capture
  6. 06Ontology Update

dashed = the inference step, where the system exercises judgment

System stack

Data in · decisions out

01

Sources

The live moment and the record

Telephony / live audioCore bankingaccounts, balances, productsCRM & interaction notesProduct eligibility rulesRecorded outcomes

02

Ingestion

Real time, on-premise where required

Streaming transcription< 300ms · on-prem optionIntent & entity extractionCore banking read APIOutcome capture

03

Ontology

Customer understanding, not just transactions

Client · RelationshipInteraction · IntentProduct · EligibilityOutcomeRule

04AI

Intelligence

Match, rank, guide — in the call

Live intent detectionstreaming LLMContextual matcherprior interactions ↔ currentNext-best-action rankertrained on outcomesSuitability rules engineregulatoryEval suiteguidance acceptance rate

05Human

Human control

The adviser advises

Adviser judgmentGuidance, not scriptsSuitability & consentDismiss with reason

06

Actions

Written back

CRM interaction recordProduct application pre-fillOntology update

Observability

Every model call traced; evals run on real cases, not anecdotes.

Governance

Entitlements enforced at retrieval; rules versioned by the organisation.

Write-back

Systems of record are written only through the approval gate.

The ontology evolves with every interaction, so guidance becomes more specific to each customer over time.

// Impact

Impact

50+
Languages supported for live transcriptionverified
In-conversation
Guidance timing, against post-call reviewdesign intent

Interested in Real-time Customer Intelligence?

Let's talk about what your relationship managers cannot see mid-call.

Get in touch
// End of case studyHub for Real-time Customer Intelligence in Banking