Large Language Models (LLMs) and Car as a Service
By Simon Wilson, Chief Scientific Officer, GTS
FIFTH IN A FIVE PART SERIES
Last November, IDC published the results of a global survey into how businesses are starting to use AI. The report followed the explosion of interest in OpenAI’s ChatGPT chatbot, and the subsequent launch of similar products from Microsoft and Google. It found that for every $1 companies invest in AI, they are realising an average of $3.5 in return; 5% of organisations worldwide are seeing an average of $8 in return.
Because car subscription and rental programs generate so much data (vehicle and customer), they are perfectly positioned to take advantage of a new wave of data-driven technology that runs in the cloud. AI and machine learning should be part of their IT environment, ideally pooling data from various parts of the business to help make more informed strategic decisions.
Large Language Models (LLMs) are an example of an AI tool. Like all AI tools, they take data and use it to complete a learning task, and employ that learning for some purpose. In the case of LLMs, the data size is enormous (being a large part of what’s written down on the internet), and their purpose is to answer text queries in natural language. This is achieved by detecting complex patterns in the training data that covers a vast array of documents and countless subjects, using a form of neural network called a transformer. But what exactly can such models do for you?
The key observation is that the LLM then provides a way to interact with customers and other AI tools through natural language. This observation is behind their use in car sharing. Let’s look at a few examples.
Customer Relations:
LLMs have seen rapid adoption as a tool to draft text for publicity and press releases. They can also be used to improve the performance of chat-bots for customer queries, by improving the fluidity of the response and widening the subjects that the bot can handle before reverting to a human. When it is partnered with voice recognition, this can be an audio interaction as well as text-based. Convenience is a key selling point of mobility subscription and rental services using apps and digital tools to make getting the right car at the right time as easy as possible. Anything that increases convenience for the customer without hurting the operators bottom line is a big win and AI based customer support can deliver here.
Data Reporting and Querying Assistant:
The LLM can also act as an intermediary between the car subscription system manager and the AI tools used to monitor performance. While standard business information tools will create dashboards of simple summaries of system performance, any new query needs the help of the data scientist to consult literature, interrogate the system data and code up the analysis. We have discussed many of these in the previous blogs: statistical process control for identifying anomalous behaviour, fraud detection, etc. With an LLM, in principle one can ask it a text query e.g. ‘Are any parts of our operating area returning income per car that is well above normal this week?’. The LLM ‘understands’ this query, can identify relevant literature on the question, and propose a tool to answer the question with the system’s data, which it can pass on to the data scientist. It can then report any findings back in natural language. It allows for much greater access to data querying for the system owners. Such an intermediary can also improve the customer experience, for example by providing bespoke advice on route directions based on what the system knows about your preferences and past behaviour.
Generating Code:
Much has been made of the LLMs ability to draft code for specific tasks. This is more relevant to the likes of GTS who are providing software services for the mobility industry. Our view is that these abilities do not yet have the track record to be used in production software that needs to be completely reliable.
Human assistants not replacements:
I emphasise that LLMs do not yet have enough capability or track record to replace the human for critical tasks. Nevertheless, their use as a human assistant can enhance several different capabilities across a car share operation.
I finish by noting that more broadly there are ethical responsibilities around AI that we all need to be aware of, for example with legal issues to be resolved around its use of copyrighted material. All of this should be taken onboard, but it shouldn’t stop you from experimenting with the tools available.