So-called agents are meant to be the next big thing in artificial intelligence this year. If the breathless headlines about them are to be believed, you might think your job is on the line. It's not (for now at least) - but some back-office occupations probably are.
Last week, OpenAI announced its first agent-like tool called Operator, which carries out online tasks like navigating to a website and clicking on buttons, according to the company. But the idea is not unique. Alphabet Inc.'s Google, Anthropic and Salesforce Inc. have all launched platforms for agents, or AI systems that can act autonomously. Imagine, for instance, a customer service bot that doesn't just generate information, but can also book an appointment or lodge a complaint.
Mark Zuckerberg has said they'll replace mid-level software engineers this year, and earlier this month, Axios reported that a tech firm was preparing to release software that could autonomously handle complex tasks at a "PhD level." The hype around AI agents and their capabilities has reached fever pitch.
Yet while Silicon Valley dreams big, a small startup in London called REKKI offers a more realistic preview of how agents could transform business. It sells access to a bot dubbed Claire, which processes orders for restaurant suppliers. Using large-language models from vendors such as OpenAI and Anthropic, the software can convert midnight voicemails from chefs into standardized data for a food seller's enterprise-resource planning (ERP) systems.
The result is that one wine supplier cancelled plans to hire six seasonal workers during the 2024 Christmas rush, according to Orat Benyamini, REKKI's co-founder and chief operations officer, while other companies have cut their order processing teams from 15 people to just three or four. "We're seeing more and more suppliers being able to transition headcount, or not hiring," she says. Naming the agents makes it easier for suppliers to conceptualize their return on investment. "You can compare it to the cost of a salary of a person," Benyamini adds.
The startup has second agent called Aileen, designed to support the restaurant tendering process, scanning a spreadsheet of, say, 800 necessary ingredients and matching that data to a catalogue. And a forthcoming third bot will help suppliers boost their sales, scanning the menus of prospective customers and advising on what ingredients could be upsold to the restaurant.
Businesses don't have to wait for PhD-level agents, or even the next big breakthrough from OpenAI. "It's now about conventional software engineering, evaluation and iterating," says Raza Habib, the chief executive officer of San Francisco-based AI startup Humanloop.
It's also safer to keep tasks narrow, when large language models have been known to make mistakes. Combine their access to customer data with the ability to take action and you get more risk, and not just for making errors.
One of the earliest uses of AI agents has been to send personally-crafted sales emails. Downtown San Francisco is currently filled with billboards urging businesses to hire an "AI sales representative." But they've had mixed results, according to Habib: "I get so many of these (sales) messages now that I'm not sure if they're human-sent or not."
The reality of where AI agents are headed is that they won't be blank-slate geniuses that can handle lots of tasks. Instead, they'll be highly-specialized at repetitive chores, particularly where the risk of a mistake won't kill the business.
"A lot being sold about AI and 'agentic' in particular is a fantasy," says Matt Calkins, the CEO of enterprise software firm Appian Corporation, a company that has spent years automating processes for major banks. "AI shouldn't be making big decisions. It should be doing discrete tasks for which it is suited, as part of a team pointed toward a greater goal."
Sometimes the most powerful technology isn't the most ostentatious, but the kind that quietly gets the job done. ("Secret agents," anyone?) Companies that wait for super-intelligent AI tools promised by OpenAI might miss a more immediate opportunity focused on more focused automation. You don't need PhD agents to process orders and handle routine documents - but for now, handing such assignments over to software is more likely to help the bottom line.
(This story has not been edited by NDTV staff and is auto-generated from a syndicated feed.)