What is AI good for? Three broad trends are trying to answer that question.
First, there’s the discussion about replacing humans entirely. We have companies developing an AI that aims to replace roles such as a sales development representative (SDR) and venture capitalists creating funds that will only invest in companies replacing human workforces.
Second, there’s the discussion around artificial general intelligence (AGI) – AI that matches or surpasses human-level thinking. This is the topic that the likes of OpenAI are wrestling with, and there is a lot of debate over when we will see true AGI. Will it be in five years, fifty, somewhere in between?
The final thread is agentic AI, where the AI recognises words as clues to bring in other software and execute a task with minimal intervention. I tell the agent to book me a flight, and it does it.
Gartner named agentic AI one of its top strategic technology trends for 2025 at the end of last year. More recently, it predicted that agentic AI will autonomously resolve 80% of common customer service issues without human intervention, leading to a 30% reduction in operational costs, by 2029.
From humans helping AI to AI helping humans
It’s not hard to see the huge potential in AI agents. The capabilities are improving all the time; however good agents are today, they are going to be ten times better in a year.
The current focus of the conversation around agentic AI is on how humans can teach AI to become better at tasks. What if to get the most out of agents, we need to think about the complete opposite: how AI can tell humans what to do, ultimately helping them to become better and more capable at specific tasks?
This is already happening: Google Maps is powered by AI and helps humans navigate more effectively. It identifies the best routes and, while on the journey, provides suggestions and alternatives in real time if delays occur.
We have other coding tools that, given an objective, will write the exact code you need to actually build out your idea. I have recently started using a tool that will take a written description of a user workflow and build out an entire user interface. There used to be an old adage with startups that ideas are a dime a dozen – that execution is everything. Now, that adage may be getting turned on its head: executing is being commoditised while ideas are everything!
This is the real power of AI – to deeply understand your goals and guide you not only on exactly how you can achieve them but also on all the course corrections you need to make along the way as your environment and the tools at your disposal change. .
Combatting inherent model instability with what’s possible now
When we consider the debate around whether AI can replace humans, or when we’ll get to AGI, this idea that we should be using agents to improve our capabilities is a far better vision for the future of AI. Why? Because it harnesses what is possible now and also removes the inherent fear people have of losing their jobs and expertise to AI.
It’s important to remember that we’re at a point where everything we know about AI’s capabilities and limitations are changing on almost a weekly basis. We’ve got the foundational models, and we’ve got application developers building on those models at the same time. It’s like building a car, while the chemical nature of steel is changing every week. For most people, the application is where they interact with AI, so application developers optimise toward the user experience. But the underlying models are optimising differently; they’re trying to develop better reasoning and, through that, better generalized capabilities. While they test these new capabilities on benchmarks, the reality is not even the foundation model builders know exactly how the new versions of the model will perform in any given use case. This creates an inherent instability for application builders and their products, not to mention their economics.
This in turn, impacts trust in AI generally. We don’t trust what we can’t control, and we feel like we can’t control unstable AI applications. They can’t handle complex tasks without huge amounts of input, which can feel like it is defeating the point.
We can overcome this by flipping the perspective to this idea of not handing tasks off to AI but using its capabilities as a source of knowledge to make us more effective and using our capabilities as humans to fill in for any AI deficiencies.
Could AI launch a restaurant?
For example, say I want to open a restaurant. To do that I need to know a huge amount: the role of location, design, what happens during fit-out, licensing laws, finding suppliers, negotiating costs, hiring staff, marketing the business, launching it, managing reservations, getting the pricing right. Even the most highly experienced restaurateur will only be an expert in some of these elements, drawing on consultants and other specialists for support. That’s additional costs, plus there’s the challenge of everyone having the time to commit to the project.
Or I can use agents. Not one agent right now, because opening a restaurant is too complex, but I can use an agent to tell me what I would need to open a restaurant, before deploying other agents on certain aspects of that process; completing licensing forms for my review, identifying the best suppliers for the restaurant’s cuisine, providing guidance on costs to inform any negotiations.
I will still need humans for much of the work, but with agents, I’ll be able to work faster, find solutions to problems, and overcome other obstacles much more effectively. Suppliers going bust or increasing their prices that hurt my margin? An agent can find a new one. Licensing laws changing? An agent can keep me updated and review our existing agreements to ensure that we remain compliant.
Instead of everything depending on, and being held up by, my team’s capabilities and capacity, agents can gather the information we need to move forward quicker.
When we’ve launched, I can use other agents to handle certain aspects of customer service, such as managing bookings, noting the preferences of regulars, and monitoring for potential no-shows.
This is all work people can do, but to do it well would require a huge investment and a lot of employees; with agents, we would still employ people, but they’d be equipped to work much more effectively.
Maintaining the current rate of AI improvement
What’s key here is that there isn’t one solution, but very specific agents controlled and deployed by humans to give them superworker capabilities to drive real impact fast and at a fraction of the cost of what they would be able to do otherwise.
Will we get to a point where AI can do genuinely complex tasks that do away with the need for any human input? Maybe, but like with all past major technological waves, it’s far more likely that humans will instead upgrade their ambition using the new technology to do things we cannot even imagine today.
