.png)
Detecting and providing adequate support for vulnerable customers is not only a business-critical consideration for every organisation but an ethical one too. Generative AI is revolutionising this nuanced area of customer support with powerful fine-tuned LLMs. Here’s how we built our proprietary model, Olive 2.0.
Key takeaways:
- In recent years, the number of vulnerable customers has been increasing across many industries.
- Vulnerability is not fixed and intersectional, making it hard to detect.
- NLP models can be used to detect and categorise vulnerabilities, like our model Olive 1.0.
- But fine-tuned LLMs present an innovative development in this field, with the ability to go a leap further.
- Our model Olive 2.0 can now detect, summarise and flag vulnerable customers, even as their circumstances change.
What makes a customer vulnerable?
Within financial services and utilities providers, this question is fundamental to the duty of care companies have for their customer base.
Adequately detecting and providing the right support for vulnerable customers, beyond the ethics of customer wellbeing, is a critical consideration for brand reputation and ultimately the organisational bottom line.
Yet, more often than not, vulnerability is understood in relatively ambiguous umbrella terms, relying upon incredibly broad categorical definitions like ‘Severe or long-term illness’.
Rarely do these categories allow room for nuance and intersectionality nor do they adequately reflect the fluid nature of certain socio-economic factors that can lead to vulnerabilities.
Within the financial regulation landscape, policy tends to focus on mitigating the potential risk of harm rather than identifying the factors driving vulnerability in the first place. Depending on the regulator and location, this can also vary considerably.
The result is often a fundamental lack of understanding of the causes of vulnerability and poor organisational policy rather than embedded proactive solutions.
Within the evolving sector of customer care and due diligence, how can we, as organisations, improve our understanding and detection of customer vulnerability?
In turn, how can we uphold a duty of care that extends beyond poorly defined regulatory tickboxes?
What is vulnerability?
The Financial Conduct Authority defines a vulnerable customer as “someone who, due to their personal circumstances, is especially susceptible to harm - particularly when a firm is not acting with appropriate levels of care. [Personal circumstances] could be poor health, such as cognitive impairment, life events such as new caring responsibilities, low resilience to cope with financial or emotional shocks and low capabilities, such as poor literacy or numeracy skills”.
In recent years, on top of existing hardships, the pandemic, high inflation, the cost of living crisis, wage deficits, geopolitical conflicts, chronic illness and climate change have all put increased strain on many people’s livelihoods.
In 2023 around 21.4% of the total population of the EU were at risk of poverty or social exclusion. New research from NICE has also found that while only 17% of customers in the UK self-identify as vulnerable, as many as 67% could be classified as vulnerable according to the FCA’s criteria.
Vulnerability, however, is not fixed — which is often what makes vulnerability so difficult to define, identify and ultimately mitigate.
Any customer can often move in and out of different vulnerabilities depending on circumstance and their degree of social mobility.
For example, a loss of income for an individual with caring responsibilities or existing financial hardship may be far more serious than a similar loss of income for an individual who doesn’t have pre-existing vulnerabilities.
Utilising Natural Language Processing for vulnerability detection
Rapid advancement in AI technology has created a unique and ever-increasing opportunity for vulnerability measures.
Natural Language Processing (NLP), a type of AI that lies at the intersection of linguistics and automation, enables computers to ‘understand’ natural human language.
When applied to customer communications and interactions, NLP can assist in identifying vulnerable customers and, importantly, determine what’s causing the vulnerability, by analysing language and detecting certain behavioural and linguistic cues.
In 2021, we launched the first version of Olive, our proprietary NLP model, to help us identify vulnerable customers and the possible causes of their vulnerability.
Olive was trained using thousands of sample messages, similar to those that we could expect to receive from customers. Each message was labelled with relevant industry regulation categories of vulnerability. For example “Bereavement”, “Income shock” or “Severe or long-term illness”.
Olive has learnt from this training material to analyse customer interactions and detect language cues that may suggest a customer is suffering from vulnerability, or possibly multiple vulnerabilities simultaneously.
These are then flagged to customer service agents in real-time, with customers displaying possible vulnerabilities prioritised for human support. Agents can then supplement these regulatory categories with additional information and descriptive notes.
These insights are also available to clients, enabling them to see in granular detail the issues that are affecting their customer base.
Applying empathy and automation with Large Language Models
Once a vulnerable customer has been identified, what then? How do you ensure they are supported with personalised solutions and support?
Enter, Large Language Models (LLMs).
A form of generative AI, these models utilise deep learning, drawing from vast datasets to interpret and generate new content in the output of text, audio, images or code.
Whilst foundation models like ChatGPT and DALL-E may instantly spring to mind, proprietary fine-tuned LLMs are revolutionising the way customer service teams are handling customer support.
Fine-tuned LLMs take foundation models like ChatGPT, which have already been trained on huge amounts of data and are adept at a wide range of text generation jobs — and further refine them with domain-specific data.
This is exactly what we did with Olive 2.0 — upgrading it from a supervised machine-learning model to a fine-tuned LLM.
This included training Olive 2.0 on a wide range of paired texts including sample customer interactions and analyses covering different types of vulnerabilities, whether they are long or short-term, their impacts, nuances, possible causes and how this affects how we communicate with them.
To ensure the model was fit for purpose, we cross-referenced its performance on over a thousand examples that had been labelled by our experienced customer services team — reducing the incidence of vulnerabilities flying under the radar.
The outcome? Olive 2.0 can now not only detect intersectional vulnerabilities within every customer interaction but also generate summaries of customers’ nuanced circumstances whilst providing recommended next steps to customer service agents in real-time.
Engaging vulnerable customers with full spectrum support
The potential of using a powerful, fine-tuned LLM like Olive is enormous. Detecting which customers are vulnerable with a high degree of accuracy means we can automate responses to interactions that aren’t flagged as vulnerable.
This frees up time for human agents, who can spend more time with complex cases or customers in need of more support.
Using an LLM like Olive also allows for a far more nuanced insight into a customer’s immediate situation, taking into account how quickly vulnerabilities can change and offering recommended next steps for our customer support team to action in real-time.
For example, Olive may pick up and flag that a customer has recently lost their job. Whereas previous model architectures would stop here, simply categorising the overarching vulnerability, e.g. ‘Income Shock’, Olive will take into account entire interaction histories to build a more robust picture.
For example, the new flag might read, ‘Customer is homeless, unemployed, and receiving only £350 in benefits per month to support herself and her son.’
Olive may suggest following up via email for a customer who suffers from anxiety and has implied they prefer not to speak on the phone, or suggest industry-specific resources to send over that might be particularly relevant to that customer's case.
Once the model has flagged possible vulnerabilities, our customer services team can tailor their support to a customer’s needs, like offering breathing space, providing alternative communication methods or signposting them to relevant resources.
Going further than just customer care, Olive 2.0 can be further instructed to provide recommended next actions for specific clients when required. This ensures businesses can signpost to chosen resources or solutions — overall minimising friction for their customers across a range of difficult situations.
Why supporting vulnerable customers is so important
It’s likely that most organisations have far more vulnerable customers using their services than they are accounting for.
Ensuring vulnerable customers are identified and subsequently supported, with solutions adapted to their needs, is a business critical, as well as ethical consideration.
Supporting vulnerable customers is fundamental to building an exemplary customer experience and glowing brand reputation that extends far beyond regulation and industry compliance.
Fine-tuned LLMs that adapt and build on the already powerful capabilities of foundation LLMs are rapidly transforming how many organisations can approach customer care.
Rather than viewing AI within this space as a point of contention, we should view it as a tool that can enable organisations to stringently uphold regulatory requirements whilst going above and beyond their duty of care for each and every customer.
Intrigued to see how Olive 2.0 could revolutionise your vulnerable customer detection and care? Get in touch today.
