AI Agents in 2025: The Quiet Revolution Behind the Loud Promises

2025 was supposed to be the year when generative AI grew up. 

Back in January, OpenAI cofounder Greg Brockman first heralded “the year of AI Agents”. Ever since then, consulting firms have been falling over themselves to demonstrate how AI can act as a “digital worker”, working tirelessly alongside humans to turbo-boost productivity. 

As the year wraps up, it’s time to take stock. Have agents proved their worth?

‍

What are agents?

The promise of agents has been a bit hard to grasp. Software is something we all understand. We click around, we set up a profile, we plug some data, and out pops a nice dashboard. What about an “agent” though? What exactly does it do? How do I interact with it?

Well, in theory, you should be able to instruct an agent in the same way you instruct an intern - it could be via Teams/Slack, email, even phone call. Everything short of tapping it on the shoulder and saying “would you mind just helping me with…”

The magic is typically created by connecting two things. First, you need an AI language model (an LLM, like ChatGPT) so that your plain English instructions can be understood. This should then have access to “Tools”, so that the instructions can be carried out. A tool could be any modern web SaaS, like Excel, Gmail or access to internal info / data. Like an intern, the Agent can then access tools if it needs to, in order to carry out its task. You could instruct it to “Put all my emails into a spreadsheet”, “Set up a meeting with Julie” or “Summarise my customer feedback from last week”.

Yes, these use cases are generic and not exactly transformational. Where things begin to look interesting is looking at business-specific use cases. Where are the gaps in the workflow? What is causing high-value team members to spend hours of their time on a menial admin task? Perhaps it’s a contract addendum that needs to be filled out every week or so with project-specific data, or a reporting pack that needs to be pulled together.

‍

Have agents lived up to the hype?

We've entered a world in which the mere mention of AI can split the room. The naysayers claim that humans will always be needed, and trusting an AI with your core business processes is a recipe for disaster. AI “hallucinates”, meaning the unquestioning faith we put into the accuracy of our software is not applicable in this new world. 

The bulls can be equally doom-mongering, with AI always one step away from replacing everyone’s job, excepting pro sports stars and few other lucky exceptions.

Conflicting reports amplify the noise. In August, an MIT report became infamous by claiming that 95% of AI pilots by corporates had been unsuccessful. The bulls ignored it, claiming that the pilots had been poorly implemented, and that the study’s data was insufficient anyway. 

On the other hand, both Walmart and Citigroup have doubled down on their early agent pilots, having seen significant ROI.

Beneath the surface froth though, a few key trends have emerged.

1. Most of the corporate gains are in the realm of “Shadow AI”. Essentially, staff using ChatGPT to help them with their jobs, often circumventing the outdated tools that their company gives them in favour of the latest AI models that they’re used to using at home.
2. “Moonshot” AI transformation is incredibly difficult. Most businesses simply aren’t ready - processes aren’t codified, data is dispersed and of varying quality. Even VC darlings Klarna had to roll back on a plan to lay off staff in favour of AI customer support agents.
3. Tightly defined use cases with real, measurable goals are succeeding. The quiet wins are boring - automating compliance reports, reconciling invoices, drafting customer support emails - but they’re the ones that are showing ROI.

‍

How should businesses approach this?

1. Clarify your AI strategy with your team. With some high profile mass layoffs hitting the headlines, some people will inevitably be concerned for their jobs. Now, more than ever, leaders need to communicate effectively with their teams to clarify their intentions. Generally, it is unlikely that entire roles will disappear due to technology as it exists today. Measured steps to embrace AI gives teams the time to adjust in stages. A forced one-shot transformation in several years’ time may not.
2. Upskill your team. Depending on company size, it’s unlikely that the CEO will know all the pain points in all the processes their company runs. Businesses should treat AI transition like a high priority internal project, and ensure that all relevant team members are properly equipped to find opportunities.
3. Focus on quick, cheap wins. There will inevitably be opportunities to speed up your processes and make your team’s lives easier - gaps in the system where an LLM could indeed help. Where is time spent on report writing? Document generation? What might the marketing team do with 10 highly capable interns? Where is Excel a bottleneck? Where are internal sign-offs holding things up? Where are we still copy-pasting data? Who needs better information at crucial moments? If in doubt as to whether AI can help, ChatGPT is your friend.
4. Recognise what AI is and isn’t. The “digital worker” framing has a lot to answer for here. AI is a tool to be understood and used. It doesn’t think like a colleague, doesn’t learn like a colleague and it has no initiative like a colleague. Like any tool though, if you use it in the right way, it can improve productivity.
5. Plan the transition when ready. When the easy wins have been captured and the potential has been understood, it’s time to ask larger questions: “What does an AI-first company in our industry look like? If we were starting from scratch, what systems would we put in place? How would the data flow?” Getting ready for an AI transition is no small task, and beyond the scope of this article. Suffice it to say though, most businesses will need help.

‍

The Final Verdict

Despite all the noise, there will be ROI to capture for almost every SME. The winners will be those that take the time to understand the tools, and begin with small, deliberate steps to automate and augment what their teams already do. It might seem that the world is moving fast right now, but those that thrive will treat this as a long game: experimenting, documenting what works, and upskilling their people along the way.

The coming years will sort the talkers from the doers - and the doers will be the ones who quietly figure out how to make AI serve them, one practical use case at a time.

‍