Roy Singh heads up Advanced Analytics globally at consulting powerhouse Bain & Company. He reveals 3 ‘Opportunity Buckets’ that your business can leverage, and 5 ‘Risk Buckets’ to avoid.
By all accounts, British-born Roy Singh was early to AI. Born in England, he started working in the field in the late 90s, after studying Mathematics at Oxford. In 2003, Singh joined Microsoft’s London office. He moved to the Bay area a decade later and signed on as a partner at Bain in 2017, leading AI at one of the world’s most prestigious consulting firms.
“I’ve been working in Natural Language Processing (NLP) which is the broader field in which live language models sit, since the mid-2000s,” Singh tells Forbes Australia in an exclusive interview at the Bain office in Sydney.
LLMS — as Large Language Models are known — act as the foundation of the technology that gives rise to ChatGPT, introduced to the world by OpenAI last year.
“Since the advent of deep learning in 2012, followed by innovations at Google in 2017, and then Open Al, we’ve dramatically improved the effectiveness in speech, images, and so on,” says Singh.
The Bain partner saw promise in the work OpenAI had been working on since 2015, and signed a deal to offer the services to Bain enterprise clients long before the technology became public in November 2022.
“We had our first conversations with OpenAI back in 2021 and partnered with them early in 2022,” says Singh.
Buckets AI can help with
That partnership, and those with other machine learning providers Microsoft and Amazon, helps Bain’s enterprise clients develop AI solutions in their own companies.
“Companies are using the base technology via a so-called API — application programming interface — to build new experiences.”
One of the areas Singh sees as having the most potential to change the way businesses operate and innovate is through leveraging unstructured data.
Text, images and other kinds of media fall into the unstructured data category. Historically, structured data, not unstructured data, has been used to develop AI systems.
“I think what’s been exciting over the last year is just been the dramatic progress that has really democratized the ability to leverage these unstructured datasets,” says Singh.
It has opened up opportunities to optimize productivity in a company, make customer service more efficient, and aid in product development and innovation.
Opportunity Bucket 1: Productivity
“There are tasks that have historically been done by human beings,” says Singh. Productivity can be optimized by taking some of those tasks off the plate of human workers, and having them done by analysing unstructured data, for example, text.
“One of our retail clients was able to shave off about 20% of the end-to-end procurement process by [using AI to] draft the RFP, look at the responses from vendors, and so on,” says Singh.
He sees opportunities to use AI to increase productivity across multiple industries.
“Contact centres, procurement, legal finance, HR. Research into documents, drafting the applications of relatively formulaic rules,” says Singh.
Opportunity Bucket 2: Customer service
“One way to think about it is to say that most businesses that have customers have not generally been able to afford to have a personalized concierge, for every consumer. What this technology allows you to do, is to create that.”
Singh advises that enterprises can utilise advanced technology like LLMs to tailor customer service to meet the individual needs of the consumer.
Opportunity Bucket 3: Product development and innovation
“The third bucket is clients that have some sort of product development process — for example, in software engineering, drug discovery, life sciences.”
Singh advises his enterprise clients that AI can be put to work to innovate toward solutions that may not have been apparent before this technology was available.
“They might be accelerating the existing innovation process, or getting products to market more rapidly. Or, coming up with a completely new business proposition to say, that with this technology, we can serve our customers in new ways.”
With great reward comes great risk
In addition to advising client companies on how they can use AI to drive growth, Bain is integrating OpenAI technology internally to maximize efficiency.
“We have 15,000 knowledge workers that do research, they synthesize documents, they do analysis, they draft you know, strategy documents and and carry out customer surveys, they develop software.” Those areas are ripe for innovation fueled by AI, freeing up employees to work on human-essential tasks.
Bain was founded in 1973 and now has a presence in 40 countries. The firm espouses a strong commitment to environmental, social, and ethical issues, and says its global supply chain is in the top 1% of sustainability ratings globally. Safety is always top of mind in the projects Bain undertakes with AI says Singh.
“Machine learning technology is inherently probabilistic – it typically does not follow a completely sort of set path. You need to evaluate, under different circumstances, what is this algorithm likely to produce,” Singh advises.
He sees the risks as falling into five categories:
Risk Bucket 1: Untruthful answers
“There are risks of inaccuracy on truthfulness, If someone is putting in a question on which the model has not been trained, you can potentially have so-called ‘hallucination’ where you get an untruthful answer,” says Singh.
Risk Bucket 2: Bias
“There are risks around bias. Inherently in some data on which the language models have been trained, and in images as well. Images of different genders, nationalities, ethnicities, that have certain associations. This is not a sort of new issue — for example, historically, as you search for a CEO on Google — the gender balance of the images that come back, is not 50/50.”
Singh advises these biases need to be taken into consideration when optimising machine learning models.
“In language models and other AI models, there can be unforeseen, and unanticipated bias that needs to be tested and controlled for in a disciplined way.”
Risk Bucket 3: Copyright Infringement
“There are also risks of inadvertently infringing copyright. Many research labs have made strong efforts to make sure that web scraping excludes copyrighted content to make sure that they are not at risk of violating anyone else’s IP. But that’s an imperfect science.
Risk Bucket 4: Impact on employees and consumers
“There are risks around applying this technology in a way that is not seen to be in employees’ and consumers’ interests. We’ve seen some of that in social media over the past few years, in terms of targeted advertising,” says Singh.
Risk Bucket 5: Systems stability and security
“There are risks in terms of systems stability, and security. These are very complex models. And as with any complex software systems, the more complex the system is, the more ways that it can go wrong,” says Singh.
“We’ve seen things like flash crashes in the financial markets in the US. Utility systems can fail because of complex software – the systems can create additional risks and instabilities as there are many different categories.”
Strategies to mitigate these risks is an area that Bain also advises its clients on.
“As well as the underlying modelling technology, you also have to build technology to test and verify the risk of the model. You have to have machine learning models that are simply there to assess the risk of other machine learning models that you’re deploying.”
How enterprises can manage the risks
“With any uncertain or risky endeavour, you need to have a strategy,” says Singh.
Fundamental to an AI risk strategy, is a robust testing process.
“As well as the underlying modelling technology, you also have to build technology to test and verify the risk of the model. You have to have machine learning models that are simply there to assess the risk of other machine learning models that you’re deploying.”
Singh says he has concerns about how AI can be used and misused. He advises clients that building AI models comes with great responsibility.
“It boils down to having a very solid testing process to identify, assess and minimise the categories of risk. There need to be roles and accountabilities, where people are responsible for testing before [the model] goes into production and is deployed,” advises Singh.
OpenAI, Microsoft, and now, Amazon Web Services
It is not just OpenAI that Bain partners with to help enterprises build AI solutions.
Bain announced a partnership to use Amazon Web Services last year and now offers its enterprise clients an AWS AI platform to build AI on, as well as Microsoft.
“What’s been exciting over the last year is the dramatic progress that has really democratised the ability to leverage these unstructured datasets.”
Roy Singh, Bain Partner and Advanced Analytics practice leader
“Over the last year since the launch of ChatGPT, OpenAI has continued to be the market leader,” says Singh. “We’ve seen a lot of a lot of demand for OpenAI and Microsoft. But we’re also increasingly seeing a diverse range of other cloud providers, Google, Amazon and other model vendors, too.”
Facilitating the development of AI has become a key part of Bain’s enterprise business.
“An increasingly large chunk of our services business – where we advise enterprise clients – is helping them get value out of data analytics, and machine learning AI,” says Singh.
Look back on the week that was with hand-picked articles from Australia and around the world. Sign up to the Forbes Australia newsletter here or become a member here.
