October 18, 2024
SINGAPORE – Artificial Intelligence (AI) has gone up another notch with the development of what are called autonomous AI agents.
These virtual assistants can in effect think for themselves and work on their own, putting them well ahead of regular AI agents such as chatbots, which often need a person to tell them what to do at every step.
An interaction on the communication tool Slack featuring an imaginary cake shop illustrates their potential.
The owner of the shop told a Salesforce digital assistant about a significant challenge: Customers often request personalised birthday cakes, which puts a huge strain on staff trying to keep up with demand.
The digital assistant responded by asking about customer behaviours, asking if they also enquired about delivery status, what an acceptable waiting time would be, and if a human should take over such requests when orders exceed 15 cakes.
The aim of the exercise is to help design an autonomous AI agent that will eventually take over the handling of customer requests for cake personalisations without troubling bakery staff, who can get on with making the actual cakes.
Many agents along these lines made their debut in various software companies in September and are in trial phases, with broader availability expected in early 2025.
These new AI agents can make decisions and take action based on what they know and what they are designed to achieve.
They essentially break down big projects into smaller tasks and tackle them one by one, getting better as they learn from what works and what doesn’t.
They would also coordinate with human staff or specialised agents from inventory, marketing and sales departments to address more complex customer inquiries.
What autonomous AI agents do that current gen AI doesn’t
The initial excitement surrounding generative AI has waned since the launch of ChatGPT in November 2022 on realisation of its limitations.
Automation platform Workato noted in September, when introducing its new sales and IT automation agents, that existing AI solutions have not met the needs of businesses.
“They can’t collaborate with other AI agents to execute processes end to end, can’t integrate with all enterprise applications and data sources, and provide a fragmented user experience across various lines of business,” it said.
Mr Wong Wai Meng, who chairs the joint Smart Technologies Action Committee (Stac) under the Singapore Business Federation and SGTech, is optimistic about autonomous AI agents.
“You can think of them as a highly efficient workers that operate at incredible speed with limited guidance,” he said.
“While Co-pilot is great for tasks like summarising and drafting e-mails, agentic AI offers far more to businesses by enhancing efficiency and productivity.”
Three significant developments for businesses
Mr Duncan Kenwright, managing director of global solutions at telecommunications firm Verizon, said it is no surprise that companies are experimenting with agents on sales and customer experience, given their potential to deliver customisation to big numbers of customers.
“The main advantages of autonomous agents are their ability to offer highly curated products or services to individual customers based on millions of data points instead of making recommendations on broad demographics,” he added.
“This approach allows companies, regardless of size, to operate at scale while treating each customer as a unique segment.”
Software-as-a-Service providers are also offering AI agents as a way to make the technology cheaper and more accessible.
Mr Gavin Barfield, chief technology officer for Salesforce Asean, said companies that have ploughed resources into building their own AI are struggling to keep up.
”What we’re trying to talk to customers about is, don’t DIY (do it yourself) your AI. Many of the large companies in Singapore have really invested in building these massive sort of teams of data scientists,” he added.
“That will continue, but you’ve got to build a lot of stuff in order to start delivering from AI. That might not be the best use of your resources. Here, we provide that out-of-the-box ready.”
The third significant development is the promise of a new payment model.
Customer relationship software firm Zendesk announced in August that it will charge customers only for issues resolved autonomously by its AI agents.
Salesforce, which is still finalising its payment terms, has signalled a similar outcome-based payment model.
Whether a pay-per-use or subscription model works better would depend on the business involved, said Stac’s Mr Wong.
He said: “This approach may encourage businesses to experiment with low usage initially, which keeps costs down. However, this model may not be suitable for all companies as utility-based pricing can complicate forecasting, and many still prefer a more predictable cost structure.”
Adoption will be cautious
A recent National Prompt Competition organised by the Association of Small and Medium Enterprises (Asme) and Straits Interactive attracted more than 100 SMEs, indicating strong interest in AI, said ASME president Ang Yuit.
However, unlike the initial rush to adopt generative AI tools such as Microsoft Co-pilot, companies are going to wait and see before signing up for autonomous agents, he said.
Mr Ang said: “There are some stuff that AI does very well, and is almost magical, but yet, there are other use cases where the AI can be just plain stupid. So, SME owners who have tried it before also know to take these announcements with a pinch of salt.”
Mr Anurag Rana, senior technology analyst at Bloomberg Intelligence, sees potential use cases in customer service, human resources and supply-chain management.
The biggest challenge is ensuring the system doesn’t make mistakes or “hallucinate”, he said.
While traditional chatbots pass unresolved questions to humans, generative AI autonomous agents might try to solve problems even when the solution isn’t clear. This can lead to costly errors and damage to a company’s reputation.
“We foresee at least one to two years of experimentation before this technology is deployed on a full scale,” Mr Rana said.
Mr Scott Bickley, research practice lead at Info-Tech Research Group, is likewise cautious about enterprise software firm SAP’s agents launched under its AI named Joule.
He noted in CIO magazine: “SAP customers must look under the hood here: What data structure is required? What level of complexity can this engine handle without error? How standardised does your process need to be for this to work?”
What next?
Costs, accuracy, copyright, explainability, cyber security and inaccuracies caused by hallucinations are most commonly cited as barriers to gen AI adoption. These problems do not go away with AI agents.
“While the technology is capable of operating autonomously, ethical considerations remain a key challenge,” said Mr Wong, adding that chief technology officers (CTOs) would need to set up safeguards as reliance on data increases.
Technology consultant Bensen Koh urged CTOs to address weaknesses such as poor data quality and partnerships.
He believes that a viable working model will emerge over time, as such AI solutions are essential for realising smart cities, autonomous vehicles and advanced digital twins.
Mr Shan Moorthy, CTO for Asia-Pacific at human resources and finance platform provider Workday, said the firm’s survey in 2023 showed 51 per cent of Singaporean businesses investing in AI and machine learning, higher than the global average of 34 per cent.
“But a lot of people are in that trough of dissolution because they bought into the hype of gen AI but didn’t see the value realisation out of it,” he added.
“I think the agentic experience will change that.”