Agents aren’t ready yet. Not overhyped forever, not a con: just not ready yet, for one specific and fixable reason. This isn’t a sceptic’s piece, it’s a timing argument. There is a single condition that, once it’s met, flips the whole thing, and I’ll be the first to say so when it does. Until then, you are not as far behind as the headlines are working to make you feel.
When I ask people for their best autonomous-agent use case, I get process management. Plumbing. Genuinely useful, occasionally, but a long way from the revolution the headlines promise. And when I watch frontier models work, I watch them make errors and, more tellingly, fail to finish what they start. For most businesses today, it is still cheaper to hire a human to do the work an AI agent will begin but not complete. There’s a technical reason for this, and it’s why I’m not just being a sceptic: until we have a better model for memory and context management, an agent can’t hold enough of the world reliably enough to be trusted to finish. That’s a falsifiable claim. When the models can remember and manage context properly, I’ll change my mind. They can’t yet.
This matters because of the fear it should dissolve. The mid-market IT leaders I talk to have CEOs reading agentic-AI headlines and asking why they aren’t turning the agents loose like everyone else. The honest answer is that the safe, governed, human-in-the-loop approach isn’t costing them a revolution, because the revolution isn’t here. The FOMO is partly manufactured. You are not missing as much as you fear, and you’re protected from a downside that is entirely real.
You should never trust an AI to work fully autonomously, because responsibility and accountability can never be delegated. A human in the loop is always necessary.
Because here is what I believe most deeply, and it’s why my company is called what it’s called: you should never trust an AI to work fully autonomously, because responsibility and accountability can never be delegated. A human in the loop is always necessary. The autonomous crowd looks faster today only because the bill hasn’t come due. The first major legal case built on a decision no human can account for, the first serious data leak, the first significant fine for an unethical automated decision: when those land, the whole market will reprice risk overnight, exactly as it did before every governance framework we now take for granted. Throughput without accountability is a short position, and it will blow up. The businesses left standing will be the ones that redesigned their processes to get the throughput with the human still accountable.
None of this means I’m a pessimist about AI. The opposite. AI is the future, a new one we get to build, the way we built the internet. I was there for that one, and I remember the excitement. But the internet didn’t become the internet through better HTML and CSS. It became the internet through Wi-Fi, the iPhone, 3G, cable modems: everything built around the core technology that made it usable and present in ordinary life. AI is at exactly that stage now. The value won’t come from the LLM. It’ll come from everything we build around it: the platform, the governance, the interface, the architecture that lets a real organisation actually use the thing.
I learned this the hard way. I once watched a very smart head of operations look at a model picker and have no idea what to choose or why it mattered. The tooling was asking her to understand AI before she could use it, and dressing that demand up as empowerment. “Do what you want” is not help. A blank chatbot in front of someone who just wants to know what they can expense for a hotel is not a strategy. The expertise has to be baked in upstream (the model choice, the prompting, the guardrails decided by the people who built it) so the user never has to carry it. The software should do the work, so the business gets value fast, without asking everyone to first become a prompt engineer.
That’s the whole bet. Not a better model: better everything around the model. Discipline built into the build loop instead of bolted on after. And a human, always, in the loop. The companies that get this right won’t be the ones who moved fastest in 2025. They’ll be the ones still standing, and still trusted, when the reckoning comes.