OpenAI for Wealth Managers – An Experiment

We’ve put AI to the test, building an app with Microsoft Azure

The use of OpenAI’s GPT-3 in the financial sector is becoming more common, but what other applications could there be for wealth and fintech firms? Technology expert Ross Brown from Sionic’s award-winning specialist team decided to explore OpenAI API, build a customised AI model, and use it to build a proof-of-concept AI chat web application using Microsoft Azure. Here’s what he found.

A fast-moving target 

The main problem with talking about OpenAI is the speed of change. The last few weeks alone have seen OpenAI’s GPT-3 based model integrated into Microsoft’s Bing Search engine, Google making a late response with Bard/LaMDA, and a host of other big players and AI start-ups following closely behind with their own Large Language Models (LLM) implementations. And if that wasn’t enough, OpenAI’s GPT-4 LLM is imminent, while Microsoft are preparing to offer a ‘no code’ option for companies to create customised, branded versions of ChatGPT.

Nevertheless, consensus seems to be emerging that the AI juggernaut involving the ‘big tech’ firms will continue making this technology accessible and pre-integrated into more and more areas of our everyday application ecosystems. Individuals across all organisations who understand both the power and limitations of this technology, and can harness it appropriately in an ‘assisted AI’ mode, will become significantly more efficient than those who try to dismiss this as just another technology fad.

Testing alternatives

Perhaps the more interesting question is whether there might be other, more specific and practical applications for wealth and fintech firms? To find out, I undertook a ‘deep dive’ into the OpenAI API, with the aim of understanding how to train and make use of specialised models using the base GPT-3 building blocks. I wanted to understand whether this approach revealed valid use cases that would fall within the reach of regular wealth and fintech firms.

To achieve this, I started work on a proof-of-concept to build a simple AI Chat web application hosted in Microsoft Azure linked to my personal OpenAI free tier account. Note that since I started, Microsoft have provided an integrated OpenAI service with a variety of pre-trained models directly into Azure.

I then set about frantically copy/pasting code from the example documentation and tutorials to see how it worked. I very quickly managed to write code to train one of the base models (text-Da-Vinci-003) and read through the sample code to understand how best to create custom, specialised models over and above the baseline.

What quickly became clear is that two concepts in particular are crucial when it comes to training your own models within OpenAI –

  • Fine tuning – pre-training the model with (preferably a lot of) specific prompt/completion data. This is effectively a templated upload in JSONL format that radically scales up the concept of ‘few shot’ prompt engineering that many will be familiar with from experimentation with the ChatGPT front end. This is then used to pre-train a new model with specialised understanding and context. Key takeaway is while this can be used to overlay a degree of domain specific knowledge into an existing model, it is not applicable for bulk teaching, importing large scale new knowledge, or processing large text documents.
  • Embedding – augmenting your model with larger blocks of text such as websites, articles, internal knowledge base data, corporate documentation. This does require an understanding of the basic machine learning concepts, as it involves creating a local database of the embedding vector structures generated by the API from your data. These are then analysing to add unique context to your subsequent OpenAI queries. If you are not sure how it works, don’t worry: there are plenty of online tutorials and examples to help get over the initial steep learning curve.

Much of this is actually achievable within the OpenAI playground environment. But beyond that, I wanted to show that it was possible to quickly train and create custom AI models using OpenAI as a base, and then make the specialised models available from within an application development environment not dissimilar to many wealth and platform technology firms.

Back to the beginning

Time to return to my original challenge: discovering whether there specific and practical applications of being able to do this for wealth and fintech firms?  Several ideas were suggested:

  • Train the model using a large history of templated adviser letters, embed this information, then use the AI to generate out ‘stylistically consistent’ content for review – essentially a drafting service for adviser review.
  • Create an adviser support team knowledge base integrated Chat App by embedding internal knowledge and training articles, and then fine-tuning it, using a long history of support queries.
  • Generate email responses to standard queries – effectively an extension of the previous point but also improving more general individual productivity.
  • Classify and categorise client sentiment including emails, secure messaging, and potentially transcribed video calls.

As an aside, areas where this approach is currently weak includes anything that involves calculations. In fact this is a general weak point of LLMs (large language models), or any direct customer-facing interaction on regulated activities. This is not surprising, given the nature of LLMs means that they will not be ‘right’ 100% of the time, and do need oversight and supervision.

Key questions

I’d recommend anyone interested in these challenges to undertake your own experimental learning first hand.  As a starting point, I’d suggest asking:

  1. Do you have existing reservoirs of valuable, but siloed corporate ‘training data’, directly relevant/applicable to current, high volume, time-heavy manual processes?
  2. Do you have the technical skills to transform that data into embedded and fine-tuning datasets and templates to create your own specialised models? Are you confident in your capability to iteratively train and fine tune these specialised models to a meaningfully high target accuracy level?
  3. Have you carried out analysis of the most cost-effective combination around base model selection, predicted token utilisation, fine tuning and embedding?
  4. Do you have the technical coding skills to integrate the models into existing processes and communications channels to expose the resulting ‘specialised’ AI models to the correct audience?
  5. Do you have appropriate moderation/supervision built into your new processes?
  6. If yours is a regulated firm, have you checked the OpenAI licensing and hosting agreement for compliance with your data and information security policies, particularly if client data being used?

With the increasing number of start-ups and the focus of the major technology players you might also want to consider whether your use-case is really unique. If not, and you are therefore unlikely to get any ‘first mover advantage’, then you may waste time building something which is already outdated before it is released. For example, a meeting minute tool is not unique, but a tool identifying missing disclaimers in pension transfer letters is more focussed and unlikely to be as rapidly commoditised by the likes of Microsoft.

In conclusion

While OpenAI’s technology ecosystem is incredibly simple and enjoyable to use, it’s the ability to recognise and exploit ‘locked up’ corporate data and use it to train genuinely useful AI models, and then embed these appropriately into organisational processes, that will be the key driver of benefit realisation.

And make no mistake, those benefits are real, achievable, and almost certainly already being worked on by your more innovative competitors.

We help wealth management and private banking clients tackle their toughest technology challenges. To discuss how we can help with your specific requirements, please contact us.

Meet the experts

Ross Brown

Principal consultant

I’ve enjoyed a number of years delivering change, covering regulatory change, product launches, platform upgrades and a full platform launch.