Map
Build continuously-updating customer behaviour profiles from live transaction data
The situation
Cosmos Bank is a mid-scale national cooperative bank serving retail consumers and small business owners across multiple states in India, combining a branch network with a growing digital presence and a large daily volume of transactions across savings, current, lending, and insurance products.
Before the engagement, the bank was carrying:
- First-call resolution stuck around 61% — roughly four in ten callers left without their issue resolved
- 8 to 14 minute wait times during peak periods, with escalations adding further delay
- Agents constantly navigating multiple internal systems mid-call
- Millions of transactions per month generating statements, but no spend insight for anyone
- A prior IVR project scaled back after low containment, and an 18-month data warehouse proposal that never moved forward
Customer volumes had grown meaningfully over the preceding two years — a good problem, but one that exposed a structural weakness in how support was staffed. Leadership was also increasingly aware that competitors were using AI to personalise offers; the bank had the data, just not the infrastructure to use it.
The challenge
The challenge was to relieve a contact centre running at capacity and turn a large, unused transaction dataset into something relationship managers and customers could actually act on — without another long-horizon infrastructure programme.
Our approach
We structured the engagement in three phases — discovery and architecture design, phased deployment of the three AI systems, and a stabilisation period with iterative refinement — with the bank's IT and operations teams involved throughout as active collaborators, not passive recipients.
Our objective was to:
Train
Train call agents on real product documentation, policy materials, and resolution logs
Categorise
Build a transaction categorisation model accurate enough for relationship managers to trust
Hand off
Keep human agents in the loop for complex, sensitive, or ambiguous cases
What we did
1
Built a continuously-updating consumer behaviour mapping engine
The engine ingests transaction data and builds live customer profiles — spend frequency, category mix, channel preference, life-stage signals, and change over time — useful to relationship managers and downstream systems alike.
Relationship managers had told us during discovery that stale data was almost worse than no data — they'd recommended products that no longer fit a customer's situation because the insight was months old. The live-update architecture addressed that directly.
2
Deployed AI call agents across voice and chat
Conversational AI agents were trained on the bank's product documentation, policy materials, and a structured dataset derived from historical query resolution logs, covering the full range of routine query types the support team handled.
Designed around a clean handoff:
- End-to-end resolution for balance enquiries, transaction clarifications, eligibility checks, and account service requests
- Clean handoff to human agents for complex, sensitive, or ambiguous cases
- Human agents see the AI-handled context before a call is transferred to them
3
Built a spend categorisation model relationship managers could trust
Every transaction is classified against a category taxonomy developed with the bank's retail team — everyday spending, discretionary categories, financial outflows, and income and transfers.
Where the categorisation went:
- Customer-facing spend summaries in the bank's mobile app
- A relationship manager view surfacing category shifts and financial events
- A macro analytics dashboard for the product and strategy teams
The bank's previous manual effort achieved roughly 68% accuracy — functional, but unreliable enough that relationship managers had stopped trusting it. The AI model was trained and refined to 91% before scale deployment.
4
Kept the bank's own teams in the loop throughout
The bank's IT and operations teams validated the approach at every phase, and the handoff logic between AI and human agents received particular attention — a bad handoff experience would have undermined the entire proposition.
What this collaborative structure delivered:
- A phased rollout the operations team could stress-test before full deployment
- A stabilisation period for iterative refinement post-launch
- Buy-in from the human agents whose day-to-day work the system was changing
How it works
One data layer feeding three AI systems — behaviour mapping, call agents, and spend categorisation — into a single set of customer and relationship manager outputs.
1
Voice, chat & transaction data
Inbound calls, digital chat, and the bank's full transaction stream feed the system.
Voice · chat · transactions2
Behaviour mapping engine
Continuously-updating customer profiles are built from live transaction data.
Continuously updating profiles3
AI call agents
Agents trained on product docs and resolution logs handle routine queries end-to-end.
Trained on real resolution logs4
Spend categorisation model
Every transaction is classified against a shared taxonomy built with the retail team.
91% categorisation accuracy5
RM & customer dashboards
Spend summaries, relationship manager views, and a macro analytics dashboard consume the outputs.
App · RM view · macro dashboard“
The AI agents have genuinely changed what our contact centre feels like to work in. Our people were spending most of their day answering the same ten questions — now they're handling the cases where a human really matters.
Cosmos Bank
“
What Atomic Loops delivered in a few months made the data we already had work harder than it ever had. The spend categorisation alone changed how our relationship managers think about customer conversations.
Cosmos Bank
61%→83%
First-call resolution
AI and human agents combined
~67%
Queries contained by AI
Resolved without human involvement
34%
Faster handling time
Reduction across all channels
91%
Spend categorisation accuracy
Up from ~68% manual tagging
The outcome
What changed within the first three months:
- First-call resolution improved from 61% to 83% across AI-handled and human-handled queries combined
- AI agents contained approximately 67% of inbound queries without human involvement
- Average handling time fell by 34% across all channels
- Spend categorisation accuracy reached 91%, up from ~68% achieved manually, covering 100% of transacting customers from day one
- Relationship managers reported materially better customer conversations within the first month
- Human agent satisfaction improved as routine queries shifted to the AI layer, leaving agents more time on cases where judgment mattered
Frequently asked questions (FAQs)
1) What is AI-powered banking intelligence?
It's the combination of behaviour analysis, conversational AI, and transaction categorisation that turns a bank's existing data and service channels into faster resolution and more useful customer insight.
2) How do AI call agents improve first-call resolution?
By handling routine queries — balance checks, transaction clarifications, eligibility questions — end-to-end, and handing off cleanly to a human agent with full context for anything complex or sensitive.
3) What is consumer behaviour mapping in banking?
It's a continuously-updating profile of a customer's financial behaviour — spend patterns, channel preference, life-stage signals — built from transaction data so relationship managers work from current information rather than static reports.
4) How accurate is AI-based spend categorisation?
In Cosmos Bank's deployment, the model reached approximately 91% categorisation accuracy at scale, compared with roughly 68% from the bank's prior manual tagging effort.
5) How do AI call agents hand off to human agents?
The AI agent transfers the conversation along with the context it has already gathered, so the customer doesn't have to repeat themselves and the human agent can see what's already been discussed.
6) Can AI call agents work across voice and chat channels?
Yes. The same underlying agent capability can be integrated with both a voice channel and a digital chat interface, giving customers consistent service regardless of how they reach out.
7) Is customer transaction data secure in an AI banking system?
Systems like this are built to operate within the bank's existing data governance and security requirements, with categorisation and behaviour models trained and run under the bank's own controls.
8) How does Atomic Loops help banks act on data they already have?
By building applied AI systems — categorisation models, behaviour engines, conversational agents — that work with data the bank is already collecting, rather than proposing a multi-year infrastructure rebuild first.
Industry Services
Education
Focus areas - Education
© 2026 London School of DIGITAL BUSINESS.
All Rights Reserved