Blog article
See all stories »

How to prevent AI induced hallucinations

In the financial services sector the transformative power of artificial intelligence (AI) promises to, and in many cases already is, driving real-time business success. From enhancing data-driven insight and productivity, to improving fraud detection and customer experience.

However, in the rush to adopt AI many haven’t taken into consideration how having poor quality data on customers could lead to gibberish, or worse, biased and inaccurate outcomes. What is commonly called AI induced ‘hallucinations’, which leads to poor results. 

To provide an example, if someone living in a prosperous part of a city applies for a loan, and the postcode the bank has on their system for them is in an area of the city that is less wealthy, this could impact on the value of the loan offered and the rate of interest determined by AI. This may lead to the prospect sourcing a quote elsewhere. 

Data decays rapidly  

Data decay is a significant factor impacting on the effective implementation of AI. Data decays fast with customer contact data lacking regular intervention degrading at 25 per cent a year as people move home, die and get divorced. Additionally, 20 per cent of addresses entered online have errors; these include spelling mistakes, wrong house numbers, and incorrect postcodes.

To avoid the scourge of inaccurate contact data it’s important to have verification processes in place at the point of data capture, and when cleaning held data in batch. This typically involves simple and cost-effective changes to the data quality process.

Address autocomplete / lookup for correct data in real-time

A valuable piece of technology to use at the customer onboarding stage is an address autocomplete or lookup service. These deliver accurate address data in real-time when onboarding new customers by providing a properly formatted, correct address when they commence inputting theirs. It also reduces the number of keystrokes required, by up to 81 per cent, when typing an address. This results in the entire onboarding process being speeded up, reducing the probability of the user not completing an application to access a service, for example. This approach to first point of contact verification can be extended to email and phone, so that these valuable contact data channels can also be verified in real-time.

Deduplicate data with an advanced fuzzy matching tool

Many customer databases have duplicate rates of 10 to 30 per cent. It’s a significant issue and frequently occurs when two departments merge their data and errors in contact data collection take place at different touchpoints. Not only does duplication have the potential to confuse an AI application, but it adds cost in terms of time and money, particularly with printed communications, and it negatively impacts on the sender’s reputation.

Using an advanced fuzzy matching tool to deduplicate data is the answer. It’s only by using such a service that it’s possible to merge and purge the most challenging records and create a ‘single user record’ which delivers an optimum single customer view (SCV) that AI can make learnings from. Additionally, organising contact data in this way will maximise efficiency and reduce costs, because multiple outreach efforts will not be made to the same person. An added benefit is that the potential for fraud is decreased because a unified record will be established for each customer.

Data suppression supports AI tools

Data suppression, or cleansing, using the appropriate technology that highlights people who have moved or are no longer at the address on file, is a vital element of the data cleaning process, and consequently in supporting efforts with AI. Along with removing incorrect addresses, these services can include deceased flagging to stop the distribution of mail and other communications to those who have passed away, which can cause distress to their friends and relatives. By employing suppression strategies financial institutions can save money, protect their reputations, avoid fraud and aid their AI efforts.

Use a data cleaning SaaS platform

Today, it’s never been easier or more cost-effective to deliver data quality in real-time to support AI and wider business efficiencies. It’s possible to source a scalable data cleaning software-as-a-service (SaaS) platform that entails no coding, integration, or training. This technology cleanses and corrects names, addresses, email addresses, and telephone numbers worldwide. It matches records in real-time, ensuring no duplication, and offers data profiling to help source issues for further action. A single, intuitive interface provides the opportunity for data standardisation, validation, and enrichment, resulting in high-quality contact information across multiple databases. This can take place with held data in batch and as new data is being collected, and can also be accessed via cloud API or on-premise, if required.

AI has the ability to give your financial institution a competitive edge, but this is dependent on the quality of data fed into the AI models. Inaccurate data leads to AI ‘hallucinations’ with unreliable predictions and hence bad outcomes. To maximise the success of your AI efforts apply best practice data quality procedures. 



Comments: (2)

A Finextra member
A Finextra member 10 July, 2024, 18:21Be the first to give this comment the thumbs up 0 likes

data quality has got nothing to do with 'hallucinations'. AI hallucinates perfectly well on it's own with data that's impeccable.

A Finextra member
A Finextra member 10 July, 2024, 18:22Be the first to give this comment the thumbs up 0 likes

maybe the blog was written by AI! lol.

Barley Laing

Barley Laing

UK Managing Director


Member since

28 Jan 2019



Blog posts


This post is from a series of posts in the group:

Artificial Intelligence

After the successful launch of the Chat GPT 4.0 chatbot by OpenAI at the beginning of 2023, many businesses started testing the tools provided by artificial intelligence and the areas of their application.

See all

Now hiring