BREAKINGVIEWS-AI agents have clear mission, hazy business model

The author is a Reuters Breakingviews columnist. The opinions expressed are her own.

By Karen Kwok

LONDON, Feb 20 (Reuters Breakingviews) - In late 1999 the online fashion startup Boo.com launched its website, which featured a virtual shopping assistant. The avatar, known as Miss Boo, was supposed to offer styling tips to customers and allow them to view jeans and shirts on three-dimensional models before buying. The concept may sound familiar to anyone who has recently listened to technology companies extoll the potential of agents using artificial intelligence. Boo.com collapsed a few months after it launched, in part because the website proved slow and unreliable. The technology industry’s hopes are riding on a bet that AI-powered agents will be more successful this time.

“Agentic AI” is the latest buzzword on the lips of executives at large tech firms and their enterprise clients. Since OpenAI launched its ChatGPT chatbot in late 2022, companies like Microsoft MSFT.O, Alphabet GOOGL.O and Amazon AMZN.O have poured tens of billions of dollars into startups building large-language models, which use vast amounts of text to learn and then generate human-like language. However, investments in companies like OpenAI and Anthropic face uncertain returns, especially after Chinese rival DeepSeek released a model which performs as well as its Western rivals at a fraction of the cost. That’s one of the reasons Big Tech firms are now eager to talk about AI agents, which can in theory replace tasks historically done by human.

Like many technology buzzwords, the term lacks a standard definition. Indeed, many software systems already act like agents: retail giant Walmart WMT.N has long used machine learning to predict inventory and demand for its supermarkets based on certain rules. The 2025 version of Agentic AI differs in three ways, however. First, agents can in theory act independently and proactively. Second, instead of following explicit human instructions, they focus on achieving goals and broad directions, such as, say, maximising sales. Third, they can perform multiple tasks, from searching databases to making decisions to approving a transaction, all without human intervention.

The hope is that the latest generative artificial intelligence models developed by OpenAI and Anthropic will power AI agents capable of imitating human logic and reasoning in nuanced scenarios. These agents could in theory plan their actions, complete tasks, communicate with people - or with other AI agents - and learn to improve their performance. So instead of asking Google for a list of nearby restaurants, a user could instruct an AI agent to locate the best option, analyse the best available time, and book a table – all within a single platform.

This level of autonomy will take a while to materialise, though. Most of the existing agents are fancy customer service tools which perform specific tasks. For example, when a flight is cancelled, Lufthansa's LHAG.DE online chatbot - built by startup Cognigy - can rebook passengers onto other flights instantly, helping the German carrier's customer service department manage sudden spikes in traffic. A typical airline spends between 5% and 10% of its revenue on customer services. If the AI agent can help keep costs flat while revenue rises, the airline’s operating profit margin will improve.

Many companies are already using technology to replace repetitive manual tasks, thereby saving time and lowering costs. Amazon boasts it uses autonomous ​​agents to upgrade 30,000 production applications, saving the e-commerce giant $260 million annually. Productivity gains also ease the pressure to hire more staff. Salesforce CRM.N CEO Marc Benioff said in a recent podcast that, due to a 30% productivity boost from AI, the enterprise software group would not add any new engineers in 2025.

Better customer service should also translate into extra revenue. If an airline provides a relatively painless cancellation experience, customers are more likely to choose it when booking future trips. Online chatbots like Pandora's "Gemma", which can spell out answers such as the nature of lab-grown diamond, make it easier for the Danish jewellery retailer’s customers to buy presents. Investors see vast benefits: fund manager Ark Investments estimates AI agents will enable $9 trillion, or 25% of global e-commerce sales, by 2030. Some businesses may also be able to charge a subscription fee for their agents, or increase prices for bundled services. Microsoft is testing its users’ appetite by hiking subscription prices for its personal and family users of its Microsoft 365 suite of software which includes AI features.

Businesses which have long built their own software and collected vast amounts of data will find it easier to develop use cases and build their own agents. DoorDash DASH.O, for example, has used Amazon's technology to improve the U.S. food delivery firm's contact centres to answer questions from couriers. Pets At Home PETSP.L built an agent using Microsoft's Copilot technology to help the retailer detect fraudulent transactions.

Companies that do not share such advantages are instead more likely to buy ready-made AI agents from software groups like SAP SAPG.DE and Salesforce. The German group’s Agentic AI tools raise the rate of detecting flaws in products for manufacturers by up to 90%, according to data from JPMorgan analysts who surveyed SAP's clients. Startups are also fighting for a slice of the growing pie. Grocery retailer Instacart CART.O and ride-hailing firm Uber UBER.N are partnering with OpenAI's "Operator" agent. Lyft LYFT.O, Uber’s arch-rival, is working with Anthropic's version. Consultant BCG expects the market for AI agents to generate $52 billion of revenue by 2030, almost ten times the $5.7 billion they produced in 2024.

Companies hoping for instant productivity gains will be disappointed, though. They must prepare employees to use the tools in their daily workflows. It also takes time to train AI agents to be ready for enterprise use as these systems could easily go wrong. Financial firms run the risk that an agent might approve a high-risk loan, leading to financial loss, Mckinsey reckons. Air Canada AC.TO last year was forced to refund a customer who was misled by its chatbot into buying an expensive ticket.

Further out, Agentic AI could upend some established business models. Agents that, say, help speed up audit processes or draft legal documents will shrink the number of billable hours charged by accounting and legal firms, forcing those entities to dream up different ways of charging customers. Replacing tasks normally done by junior workers also means companies need to rethink their approach to hiring new and developing staff.

Microsoft Chief Executive Satya Nadella recently declared that building AI agents should be a simple as creating an Excel spreadsheet, freeing up time for humans to do other useful tasks. Whether Agentic AI opens their wallets with similar ease is less clear. The agents may have a clear mission, but the return on investment remains fuzzy.

Follow @karenkkwok on X