I interviewed Valsan Ponnachath, President, COO and Co-founder of iTuring.ai, an integrated AI platform for regulated industries.
When banks try to retain customers whose deposits are about to mature, they often send the same offers to everyone. But what if some of those customers were going to renew anyway? The bank ends up paying more to retain customers who weren’t planning to leave.
Valsan argues that this is where predictive intelligence becomes important. AI agents can automate the workflow, but predictive models help banks decide which customers to prioritize, what offers to make, and how to reach them.
We also discussed how banks can build a reusable intelligence layer from fragmented data, how predictive models and AI agents work together, and what it takes to integrate and govern these systems within existing banking infrastructure.
Use the links below to jump directly to the topics that interest you.
- What AI needs to know about a customer before taking action
- Why predictive intelligence matters beyond automation
- How predictive models and AI agents work together
- What banks should reuse across AI use cases
- Integrating AI with existing banking systems
- Why personalization requires more than an LLM
What AI needs to know about a customer before taking action
Manu: Let’s say a bank uses a system of AI agents to retain customer deposits.
Two customers have deposits coming up for renewal. What information would the system need about each customer before deciding whether to make them a retention offer?
Valsan: Basically, you need a 360-degree view of the customer.
You already have information through your CRM and transaction data. You’ve built a profile of the customer, and you’re constantly updating that profile.
You’re calculating a number of things about the customer, with lifetime value being one of the biggest.
You’re also trying to understand their intent at various points in time, their different types of preferences, and what life events they’re going through.
When I say 360 degrees, I mean a truly comprehensive view.
What type of interactions do they like? How often do they go to the branch?
When you think about AI, you take data from different parts of the bank and create features. Those are the granular attributes you’re creating.
Those features help the AI make decisions about the things we’re discussing.
Why predictive intelligence matters beyond automation
Manu: You’ve talked about the data and features that help AI make decisions. How do you see the roles of predictive intelligence, LLMs, and AI agents in this process?
Valsan: Everybody is talking about LLMs and agentic AI today.
In my view, LLMs are about interaction. They allow you to interact with a system in English or another language. That makes it easier to get information from the system and give information to it.
Agentic AI is fundamentally about automation.
You’re making a process more efficient. Something that was previously a manual task can now be automated.
But the true value comes from intelligence, from being able to make decisions and judgments.
That is all in the predictive layer.
Agentic AI is something anybody can do. But predictive analytics gives you the intelligence to make an appropriate decision.
That’s where the beauty of AI really is.
Unfortunately, that gets missed in all the noise today.
And to get that intelligence, you need to bring data together from different parts of the bank.
Most banks, if you go to them and talk about embracing AI, will tell you they’re building a data lake or data warehouse first.
They’ll say, “Once I’ve built that, I’ll come back and do AI.”
That seems logical.
But in today’s world, with the kind of platform we have, we can bring data together from different sources without requiring a traditional data warehouse or data lake first.
We have over 250 connectors that allow us to access a wide range of core systems through flat files, databases, or APIs, bringing all the information together in real time.
Bringing fragmented data together is the first order of business.
Once you’ve brought the data together, the next step is data engineering.
Let’s say you’ve collected data from the bank’s loan, credit, and core banking systems, along with its customer service records.
Now you take that data and create features.
For example, you might start with 15 columns of data. Through data engineering and time-series analysis, you can create thousands of features from those original data points.
That’s where you start building intelligence.
When you’re building a predictive model, this data preparation and feature-engineering work typically takes a significant amount of time.
Building a model might take two to three months, and roughly 55 to 60 percent of that time can be spent on this exercise.
At iTuring, we’ve built 25,000 prebuilt features for banking and insurance, so we can reduce much of that effort.
But there’s a second benefit.
By having those features available across banking, we’ve created what I would call an intelligence layer.
In the old days, you might build a reference data model for your data warehouse. Once you’ve mapped the bank’s data, you can reuse that model for different analytics applications.
I believe the same principle applies to enterprise AI.
You need to start by building an intelligence layer that you can reuse.
You can use it for deposits, collections, underwriting, fraud, and other banking activities.
They’re all coming to the same source to build their predictive models.
Manu: Let’s go back to the deposit example. What do those predictive models actually do?
Valsan: They create insights.
If I have all the predictive insights about you, I have a much better understanding of who you are, what your preferences are, your life situation, and your demographics.
The models can identify patterns.
Who is most likely to leave the bank?
Who is most likely to renew a fixed deposit?
Is there a life event that’s causing someone to withdraw their deposit?
Once you have those insights, you can decide who to reach out to.
Let’s take a specific example.
Today, when banks run fixed-deposit renewal campaigns, they typically look at everyone whose deposits are expiring over the next three months.
Then they start sending emails and perhaps offer customers an additional 50 basis points if they renew.
But think about two customers.
I’m the lazy person who doesn’t even pay attention to when my deposit is renewing. I might renew anyway because I don’t particularly care.
You, on the other hand, might be actively shopping around. You know Bank B is offering 7%, and you’re considering moving your money.
Now the bank sends both of us the same offer.
It gets me to renew and gives me an additional 50 basis points for no reason.
The bank has just lost money.
With predictive intelligence, the first thing you do is identify who is most likely to churn.
Now you can prioritize those customers instead of sending the same offer to everybody.
That’s a big deal.
But even within that prioritized list, you need to understand why someone might leave.
Let’s say you have a customer who has been with the bank for seven years. They’re likely to leave, but they’re not particularly sensitive to interest rates.
There’s something else that’s causing them to consider leaving.
You don’t necessarily have to offer them a higher interest rate. You might offer them a loyalty bonus instead.
Another customer might be actively shopping around.
You know that from their previous interactions with the bank. Maybe they’ve been asking questions about when their deposit expires or what rates are available.
Now you need to determine the next-best offer.
If they’re already getting 7% elsewhere, an additional 50 basis points might not convince them to stay.
You could consider matching that 7%, but only if their lifetime value supports it and the bank’s approved limits allow it.
Those are the critical decisions.
That’s the intelligence you need before you do all the agentic work.
And there’s another decision after that.
How do you reach the customer?
Should a human call them? Should you use a voice bot? Would they prefer an SMS or WhatsApp message?
What’s the best channel, and what’s the best day to reach them?
Once you know that, you can put together the offer in the bank’s brand voice, with the right level of personalization.
That’s where the LLM or smaller language model comes in.
It can write the offer or communicate it in a tone appropriate for the customer.
That’s how I think about deposit retention.
And this is true for customer acquisition and other banking activities as well.
How predictive models and AI agents work together
Manu: When you talk about predictive models, are you referring to models that banks have already built, or models that iTuring builds for them?
Valsan: We build custom models for banks based on their historical data.
Before we start, we take their historical information, including CRM data and transaction data.
Typically, we need at least one year of data, preferably two.
We use that to build the models.
Manu: What if a bank already has its own predictive models? Can it integrate those models with your platform?
Valsan: Yes, absolutely.
Our platform is modular.
We have three pillars. There’s the predictive pillar, the agentic pillar, and the generative pillar.
For us, the predictive pillar is the most important.
We’re fine if a bank already has agents or uses its own LLMs. They can continue using those.
But I believe that to improve outcomes, you need predictive analytics.
Otherwise, you’re just automating whatever you were already doing.
Unless you can be more discerning about the actions you’re taking, how do you improve outcomes or do something better than your competition?
That’s the fundamental difference.
These three pillars sit on top of a data foundation that includes the connectors and the feature store.
We also have model governance, explainability, and observability capabilities.
When we create a model, we produce detailed documentation. It can be a 25-page dossier explaining the data, the features used, and how the model makes its decisions.
You can trace which data was used, which features influenced the decisions, and how the model was created.
We also track who created the model, who approved it, and which features were selected.
That helps make the models ready for examination and review.
Then there’s observability.
Once a model is deployed, you need to monitor whether it’s behaving as expected.
Has there been model drift? Has bias been introduced? Is the model still performing as intended?
Those are important model-risk considerations.
For example, if a bank already has predictive models, it could use our feature store to provide additional features to those models.
Or the bank might already have both the features and the models but need better tools for deployment and monitoring.
It could use the MLOps capability for that.
The same applies to model governance and model risk management. Banks can register their existing models and manage them through those capabilities.
Manu: Let’s come back to the actual deposit-renewal workflow.
Suppose Customer A and Customer B both have deposits coming up for renewal. How do the agents, predictive models, and data interact?
Valsan: Think about how a human would perform the same workflow.
Typically, someone would take a list of customers and prioritize it.
Now the human gets a prioritized list and takes whatever action is required.
Replace that human with an agent.
The agent gets the prioritized list and continues the work.
It creates an offer, sends the offer, and determines which channel to use.
Every time it needs to make a decision, it can get that intelligence from predictive analytics.
For example, when it needs to determine which communication channel to use, it gets that information from the predictive layer and continues the workflow.
In any workflow, there are multiple points where decisions or judgments are required.
Today, you might need a human at each of those points to exercise judgment.
The human might have to perform 50 different calculations before the workflow can move forward.
We can either automate those decision points or enable the human to make a faster, more informed decision.
That’s the difference.
Manu: So the agent isn’t just accessing raw customer data. It also needs the intelligence generated by the predictive models.
Does the agent directly access the models, or does it consume the outputs produced by those models?
Valsan: The models are producing scores.
For example, a model produces a churn score for each customer.
Now you have a prioritized list of customers based on their likelihood of leaving.
You might also have scores indicating which customers prefer WhatsApp or another communication channel.
That information is available for agents to consume when performing their tasks.
Predictive analytics is agnostic to whether a human or an agent performs the action.
If there is no agent, you can present the information in a dashboard. A human looks at it and does the work.
If there is an agent, it can consume that information through an API.
The way to think about it is that predictive analytics provides the intelligence required for the decision. The agent uses that intelligence to perform the task.
Manu: How should banks think about the ROI of a system like this, particularly for deposit retention?
Valsan: You need to look at the business outcome.
With deposit retention, for example, the bank wants to reduce churn and improve the amount of deposits it retains.
But you don’t necessarily have to wait until the entire solution is deployed to evaluate whether the predictive models are useful.
We can start with a proof of concept using the bank’s historical data.
We build a predictive model and test it against out-of-time data to see how accurately it predicts customer behavior.
For deposit churn, for example, we aim to demonstrate accuracy above 70 percent in that initial exercise.
You can do that within a few weeks.
The actual business outcome takes longer because you need to run the campaign and wait for customers to respond or renew their deposits.
But you can use the predictive-model results to estimate the potential improvement.
If you’re able to identify customers who are likely to leave, you don’t have to spend money offering incentives to everybody.
You’re conserving resources and focusing your retention efforts on customers who are most likely to need them.
You can calculate the financial value of that improvement.
What banks should reuse across AI use cases
Manu: You mentioned in one of your articles that banks shouldn’t think only about ROI. They should also think about reusability.
As banks scale the number of agents across an organization, what components should they try to reuse across different use cases?
Valsan: You can’t reuse agents.
An agent does a specific task.
An agent that handles deposit retention cannot do collections for you. You need a separate agent for that task.
What you can reuse is the intelligence layer.
Think about how banks operate today.
The deposits team might create its own features and models to understand customer behavior.
The collections team might independently build another set of features and models.
Now you have different definitions and different ways of looking at the same customer.
That’s what we need to avoid.
You want common nomenclature and common definitions.
You need an intelligence layer that’s sufficiently abstracted so everyone can use the same underlying features.
That doesn’t mean every business should use the same predictive model.
Different teams should build models appropriate for their own workflows.
You should have as many models as you need.
The real question is whether you’re governing AI properly across the enterprise.
Are you deploying models systematically?
Are you managing them, testing them for scale, and making sure they’re ready for production?
Today, many banks have different divisions running pilots independently.
Everybody is experimenting with LLMs and agentic frameworks because the tools have become accessible.
It’s relatively easy to build something and automate a task.
But that doesn’t necessarily make it scalable.
You still need engineering to make it enterprise-ready.
And you need a framework to govern and manage those models and systems across the bank.
Integrating AI with existing banking systems
Manu: Banks already have their own systems, models, and workflows.
How does iTuring integrate with those existing systems? Does a bank have to replace what it already has?
Valsan: Depending on where you are in your AI journey, you can pick and choose what you need.
For smaller banks or community banks that don’t have enough resources or tools, they can start with one use case at a time.
And even within that use case, they don’t necessarily have to go all the way into building agents and hyper-personalized communications.
They can start with predictive analytics.
Take deposit retention.
You can continue using your existing workflow. We provide a predictive score that helps you decide which customers to focus on.
You can use that score in a dashboard and improve your outcomes.
Or you can use it within your existing workflow, whether that workflow is agentic or not.
We provide the insights through an API, and the existing system can consume them.
Then, if you find that your existing workflow isn’t able to fully use those insights, you can introduce agents.
And if you want to personalize the communication further, you can use LLMs or smaller language models.
So you can take a very systematic approach.
You don’t have to implement everything at once.
Once you’ve started with predictive analytics for one use case, you’ve established a foundation that can be extended to other use cases.
You can scale both within the original workflow and across the enterprise.
Manu: When you build predictive models for banks, do you rely only on their historical data, or do you also bring in external data sources, such as social intelligence?
Valsan: Absolutely. We bring in external sources as well.
We can bring data from different transaction systems, credit bureaus, and third-party sources.
We encourage banks to bring data from as many sources as possible.
The more relevant data you have, whether internal or third-party, the more intelligence you can potentially build into the model.
For many of the bureaus, we also have precomputed features available.
So the predictive models don’t have to rely only on the bank’s internal historical data.
They can also use features derived from external sources.
Why personalization requires more than an LLM
Manu: What’s the biggest thing banks and financial institutions are overlooking when it comes to agentic AI transformation?
Valsan: I think the biggest misconception is around personalization.
Many organizations believe that if they use an LLM to write an offer, include the customer’s name, and send it, they’ve personalized the communication.
But that’s not really personalization.
To truly personalize an offer, you have to understand the customer’s current situation, their preferences, and the life events they’re going through.
You need that intelligence before you decide what offer to make.
Then you need the right tone of voice and the right communication channel.
It’s not just about automation.
It’s about having the intelligence to personalize the interaction, make the process more efficient, and predict the likely outcome.