Over the last few years, what it means to be a Machine Learning (ML) practitioner has evolved beyond recognition. Early on, when I was leading my first ML teams, many companies were unsure where ML could lend value, particularly when their data was spread across various systems. A huge part of my role involved identifying where ML could even be helpful, and then making the case for our work. Even then, each problem often required a unique approach — training models that were specific for the context they were designed to operate in.
Today, that dynamic has largely reversed. Teams are scrambling to implement AI into their products (often using general-purpose large language models) and are investing with even more urgency and ambition than in previous innovation waves like data, cloud, and mobile.
But, across this change, I believe the main tenets remain the same: you still need to prove that AI works, by reasoning about where it is being put to use and how it will deliver measurable value.
Customer support in finance is unique
As much as customer support appears to be the same, it is not a “one size fits all” flavour of problem. What might work well for retail (perhaps returning a shirt that is too small) won’t apply to the financial services, where customers might be getting in touch with an issue that requires reasoning over their data while continuing to comply with regulatory obligations.
I saw this first hand while leading machine learning at Monzo, which spent majority of its time partnering with colleagues in Operations. Getting the right outcome for each customer was not a function of delivering what they wanted—it was about ensuring that support was compliant, safe, cost effective, and empathetic. If a human agent had made casual, off-hand comments in the way that some AI agents do today, it could have lead to complaints, investigations, and could even break the law.
Imagine the following scenario:
A customer calls and says they’re having trouble accessing their card.
An AI agent looks through the knowledge base to identify possible reasons
One reason that pops up: the customer could be under investigation for fraud.
Unless something explicitly tells the AI agent otherwise, the next best step might be for the agent to tell the customer: you can’t access your card right now because you might be under investigation for suspicious activity. That would potentially qualify as tipping-off, which is illegal in the UK.
Take another example, which seems simple at face value: “where’s my money?” In the context of an international money transfer, this question could have several different meanings. Maybe the customer has transferred money out, is expecting an inbound transfer, is moving money between their accounts, is looking for a refund, has seen an unexpected direct debit, or something else. The generic AI agents we explored were striving for “best effort” resolution, meaning they’ll try to resolve the problem swiftly — oftentimes, without actually understanding it.
Architecting an AI agent for vertical expertise
This challenge was at our core when we started Gradient Labs. Any AI agent that isn’t built to account for the nuances of finance could easily fall prey to all of the mistakes that an untrained human would also make. We knew that to safely automate customer operations in financial services, our agent would need deep domain expertise, from day one.
In our first working session as co-founders, we stood in front of a white board, thinking about what the initial design of our agent might look like. But we quickly flipped our system design problem back into a framework to understand why human agents were so successful compared to AI agents in the financial services. We would ask, for each scenario we examined: “what would a human do?” The result of this design thinking is an agent that takes a nuanced, more human approach to problem-solving.
This question continues to inform every engineering and design decision we make. When ahuman agent is asked: “Where’s my money?”, they do not jump straight to an answer. They ask follow up questions to clearly understand the issue at hand and inspect the customer’s account and investigate what data is there (and, notably, what data might not be there).
Our AI agent behaves the same way. It works to understand the customer’s intent before offering solutions, it knows what guidance to follow, and clarifies further if it does not. It can reason about customer data that might or might not be present, and can pull more information and take action when it needs to.
Reshaping Operations
This kind of automation does not make human expertise is obsolete; quite the opposite. Our safe, compliant agent is only possible because of the people who instruct and empower it to handle an increasing range of processes. Our agent is a new kind of virtual colleague: it can pick up some parts automatically, like learning tone of voice and the company’s implicit knowledge, expressed in the thousands of conversations that have come before. But ultimately its approach to solving problems is driven by the specific instructions that it is given. Even here, sometimes the most useful instruction is to ask for help—for these cases, our partners simply tell the AI agent to use the Ask a Human tool.
The result is a win-win all around, with our agent automating large portions of support operations. Over the last two years, we have partnered with a range of financial services that transform their customer operations: they boost resolution, increase customer satisfaction, and keeps their humans out of the grunt work. They have become the architects that design for the nuanced and unpredictable conversations which are commonplace in a regulated environment. Their efforts often translate into an AI agent that achieves better CSAT than their best agents.
If you want to explore how AI agents can safely automate your support operation and improve customer experience in the process, book time with us here.
Over the last few years, what it means to be a Machine Learning (ML) practitioner has evolved beyond recognition. Early on, when I was leading my first ML teams, many companies were unsure where ML could lend value, particularly when their data was spread across various systems. A huge part of my role involved identifying where ML could even be helpful, and then making the case for our work. Even then, each problem often required a unique approach — training models that were specific for the context they were designed to operate in.
Today, that dynamic has largely reversed. Teams are scrambling to implement AI into their products (often using general-purpose large language models) and are investing with even more urgency and ambition than in previous innovation waves like data, cloud, and mobile.
But, across this change, I believe the main tenets remain the same: you still need to prove that AI works, by reasoning about where it is being put to use and how it will deliver measurable value.
Customer support in finance is unique
As much as customer support appears to be the same, it is not a “one size fits all” flavour of problem. What might work well for retail (perhaps returning a shirt that is too small) won’t apply to the financial services, where customers might be getting in touch with an issue that requires reasoning over their data while continuing to comply with regulatory obligations.
I saw this first hand while leading machine learning at Monzo, which spent majority of its time partnering with colleagues in Operations. Getting the right outcome for each customer was not a function of delivering what they wanted—it was about ensuring that support was compliant, safe, cost effective, and empathetic. If a human agent had made casual, off-hand comments in the way that some AI agents do today, it could have lead to complaints, investigations, and could even break the law.
Imagine the following scenario:
A customer calls and says they’re having trouble accessing their card.
An AI agent looks through the knowledge base to identify possible reasons
One reason that pops up: the customer could be under investigation for fraud.
Unless something explicitly tells the AI agent otherwise, the next best step might be for the agent to tell the customer: you can’t access your card right now because you might be under investigation for suspicious activity. That would potentially qualify as tipping-off, which is illegal in the UK.
Take another example, which seems simple at face value: “where’s my money?” In the context of an international money transfer, this question could have several different meanings. Maybe the customer has transferred money out, is expecting an inbound transfer, is moving money between their accounts, is looking for a refund, has seen an unexpected direct debit, or something else. The generic AI agents we explored were striving for “best effort” resolution, meaning they’ll try to resolve the problem swiftly — oftentimes, without actually understanding it.
Architecting an AI agent for vertical expertise
This challenge was at our core when we started Gradient Labs. Any AI agent that isn’t built to account for the nuances of finance could easily fall prey to all of the mistakes that an untrained human would also make. We knew that to safely automate customer operations in financial services, our agent would need deep domain expertise, from day one.
In our first working session as co-founders, we stood in front of a white board, thinking about what the initial design of our agent might look like. But we quickly flipped our system design problem back into a framework to understand why human agents were so successful compared to AI agents in the financial services. We would ask, for each scenario we examined: “what would a human do?” The result of this design thinking is an agent that takes a nuanced, more human approach to problem-solving.
This question continues to inform every engineering and design decision we make. When ahuman agent is asked: “Where’s my money?”, they do not jump straight to an answer. They ask follow up questions to clearly understand the issue at hand and inspect the customer’s account and investigate what data is there (and, notably, what data might not be there).
Our AI agent behaves the same way. It works to understand the customer’s intent before offering solutions, it knows what guidance to follow, and clarifies further if it does not. It can reason about customer data that might or might not be present, and can pull more information and take action when it needs to.
Reshaping Operations
This kind of automation does not make human expertise is obsolete; quite the opposite. Our safe, compliant agent is only possible because of the people who instruct and empower it to handle an increasing range of processes. Our agent is a new kind of virtual colleague: it can pick up some parts automatically, like learning tone of voice and the company’s implicit knowledge, expressed in the thousands of conversations that have come before. But ultimately its approach to solving problems is driven by the specific instructions that it is given. Even here, sometimes the most useful instruction is to ask for help—for these cases, our partners simply tell the AI agent to use the Ask a Human tool.
The result is a win-win all around, with our agent automating large portions of support operations. Over the last two years, we have partnered with a range of financial services that transform their customer operations: they boost resolution, increase customer satisfaction, and keeps their humans out of the grunt work. They have become the architects that design for the nuanced and unpredictable conversations which are commonplace in a regulated environment. Their efforts often translate into an AI agent that achieves better CSAT than their best agents.
If you want to explore how AI agents can safely automate your support operation and improve customer experience in the process, book time with us here.
Share post
Copy post link
Link copied
Copy post link
Link copied
Copy post link
Link copied