It’s no exaggeration to say that AI tools like ChatGPT have had a huge impact on the world. However, they’re yet to change the office materially.

Yes, Microsoft Office now comes with AI-enabled Copilot, and Adobe Photoshop has AI-powered tools, but it’s easy to forget that after the launch of OpenAI’s phenomenally successful chatbot, big corporations banned it. JPMorgan, Apple, Amazon and many other firms have prevented staff from accessing it due to fears of accuracy and data leaks. AI agents are part of the plan to change that.

The hope and dream of AI agents is that they will go far beyond the chatbot approach – and, for the tech platforms, maybe open the door to deeper and more lucrative deals with other firms. Sick of emailing colleagues to get them to update a spreadsheet? Tired of filing your expenses? One day, an AI agent will do it for you, so the tech companies say, and realize huge productivity gains across corporations. Instead of having a back-and-forth with a bot about writing some code or improving an email, AI agents will write the code itself, deploy it to a testing site and give you the link for you to check and tweak it, as seen in one popular demo on YouTube. Or it might improve that email, let you see the draft, and collect and summarize the responses for you.

OpenAI, Anthropic, Microsoft, Google, Salesforce and other AI software companies are already selling or developing some version of this “agentic” approach to AI. But will this new generation of AI tools be able to handle complex tasks, and when will they be ready for prime time? And what might an office full of workers using AI agents look like in the future?

The end of the flunkies?

Dan Davies, a former regulatory economist at the Bank of England and the author of The Unaccountability Machine: Why Big Systems Make Terrible Decisions, explains how AI agents might fit into the workplace and the consequences for two different types of office workers.

In investment banking, senior executives often travel around with “three or four flunkies,” he says, who gather information and summarize things for their boss. “The boss only ever gets pure processed juice from people who are good enough to be on the top staff.” If AI agents become reliable enough, then they could potentially take over the flunkies’ tasks, such as digging through a set of accounts or reading through a meeting transcript to find relevant facts.

That would have profound implications for the bankers – and the flunkies.

“Senior bankers are always thinking about their next step up, so they might be quite happy with the idea of having an AI agent to do that ‘grunt work,’” he says. “But ambitious graduates wanting to get into the general staff as their path to the top job might be worried. Because those general staff jobs might not exist.” Fortunately for the flunkies, the technology is nowhere near reliable enough to take over these kinds of tasks – yet.

“AI agents hold immense promise, but I’m not really buying this ‘agent that takes over your life’ thing. I am still doing my own expenses,” says Albert Phelps, one of the world’s first full-time AI prompt engineers. He left Accenture in that role recently to co-found Tomoro.ai, a company which helps large companies embed AI tools into their workflows.

Phelps has previously tried to build fully automated AI agent workflows where three or four different AI characters are doing different parts of a task. “You let them all speak to each other and then you give them instructions,” he says. This was helpful for testing, but none were ready for deployment into live environments for his clients, who work in the insurance, biotech and gaming industries. “We’re waiting for the next step in the AI models to really make this valuable and reliable,” he says.

The rise of AI agents

The first reliable AI agent, if it can be created, is likely to come from one of the big labs, says Phelps. He’s looking to Anthropic’s “computer use” tool, Google’s Agentspace and OpenAI’s upcoming agent tool, reportedly codenamed “Operator.” As for startups, he’s interested in vertically focused ones like 11x, which is trying to build AI sales agents, Harvey AI, which is aimed at professional service firms, and the personal study tool Glean. He’s also interested in “horizontal” startups like Distil.ai, which is trying to do the “boring bits” of enterprise like writing Structured Query Language (SQL) for databases, or Project H, which is focused on quality assurance.

“I’m less enthused about the OpenAI-wrapper companies,” he says, referring to the often smaller, seed-funded companies from, for example, the startup accelerator Y Combinator, which commonly focus on general-purpose AI agents. “The virtual employee hype is pushed by people who haven’t spent much time interacting with big enterprises,” he says. “I don’t think that’s where the value is.”

The value, says Phelps, will come when firms create the data that will help future AI agents understand their tasks. A common mistake he’s encountered among businesses trying to integrate AI into their workflows is that they assume the models know things about their business. “You have to give as much context as you can in order to get better performance,” says Phelps. Davies, the author and economist, agrees with Phelps. “The people who will have the competitive advantage with AI agents are the ones who have better information to train up the AI,” he says.

As for when he thinks AI agents will be ready for prime time, Davies says you should ask yourself, “If you’re an iPhone user and Samsung and Android came out with an AI agent that actually worked, would people switch? I think the answer to that is maybe.”