What actually is an AI agent - is it just a bot?
Let's define the thing before we judge it. "An AI agent is an autonomous software system that can set its own goals, make decisions, and take actions to complete multi-step tasks without requiring constant human oversight. Unlike traditional AI models that only answer questions, agents actively execute workflows, use digital tools, and learn from their environment." Thank you, Gemini - yes, the Claude specialist just quoted Google. One problem: it's wrong on the point that matters most. AI agents don't set their own goals - they pursue the goals you set, the way you've trained them to. And that misunderstanding is where most of the hype starts: rubbish direction in, rubbish work out.
Here's the way I'd put it to an MD. Traditional AI works like an external consultant: you go to it for advice, for research, to draft a report - you bring it work, it completes the work while you're using it, and you put in exactly what you get out. Agentic AI operates like a member of your team: you train it how the work is done, it then does the work, it comes to you when your input is needed, and you review what it produces.
That distinction changes what "useful" means. A consultant is useful when you ask good questions. A team member is useful when they're trained, managed and given real work to own. Most of the disappointment with agents comes from paying for the second and using it like the first.
So why do so many agent projects fail?
Because most organisations are running exactly that mismatch, and the data is blunt about it. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing unclear business value, inadequate risk controls - and a market so full of "agent washing" that of thousands of vendors claiming to sell agents, Gartner estimated only around 130 actually do. MIT's Project NANDA went further: 95% of enterprise generative AI pilots deliver no measurable return - not because the models are weak, but because the pilots never change how the work is actually done. The tools don't learn the business, don't fit the workflow, and quietly die in month three.
Read those numbers again and notice what they are not saying. They are not saying the technology doesn't work. They are saying the adoption doesn't work. I don't think there's a person running a business, a department or a corporate job who wouldn't benefit from a trained team member completing work for them. But like people management, not everyone gets the benefit from a new hire - it's wholly dependent on how they're trained, developed and managed.
What does an agent look like when it's working?
The honest version first, because we've watched the failure mode up close. Organisations hand out access, a few enthusiasts prompt well, everyone else gets inconsistent results - and for every hour AI gives a team back, an hour gets spent checking its work. The team ends up doing the job twice and calls the technology overhyped. They're half right: that version of it is.
Now the co-designed version. An operations team managing incoming orders: the agent connects to the source of incoming paperwork, extracts the data, cross-matches purchase order against invoice and packing list, and flags exceptions. It generates the commodity code - a human approves it. It pulls duty and VAT from the authoritative source, not from whatever the internet offers. Approved, it generates the invoice and carries the same method through until the order moves to stock. The people who used to do that admin now direct it - and the organisations we work with get 6 to 10 hours back per person per week doing exactly this, against a public benchmark of about 2.2 hours for generative AI generally. Same technology. The difference is adoption.
The design work is knowing where the human sits: where their input genuinely adds judgement, and where an approval gate must sit before anything leaves the building. Get those two decisions right and the agent compounds. Get them wrong and you stay on the AI hamster wheel.
Are agents overhyped - or are we the variable?
Here's the line I keep coming back to: an agent's work ethic is non-negotiable. It shows up the same way every time - we all see how it works. It won't disengage, resist a new process, or quietly do the job its own way. So if your agents aren't delivering, there's only one place left to look - how they're being trained, developed and managed. The question "are agents overhyped?" mostly reveals that the main blocker to returns from AI isn't the technology. It's how equipped the organisation is to run it.
That should land as good news. It means the gap between the 95% and the 5% isn't budget or luck - it's behaviour, and behaviour can be trained. Your people don't get replaced by agents; they get promoted to directing them. That's the empowerment story underneath all the hype and all the backlash, and it's the one the data supports.
Frequently asked questions
Are AI agents safe to use at work?
Safe is a design outcome, not a property of the tool. Governed agents run with approval gates at every point where something leaves the business, connect only to the systems they need, and always have a named human accountable - under UK law, responsibility for AI output sits with a person or company, never the tool.
What's the difference between an AI agent and a chatbot?
A chatbot answers you; an agent works for you. One is a conversation, the other is a colleague completing multi-step tasks - reading the paperwork, updating the system, flagging the exception - and coming back when it needs your judgement.
How many agents will a team actually run?
Direction of travel: several per person. The skill that matters for 2026 isn't prompting - it's managing and auditing a small portfolio of agents the way a team lead manages people: clear briefs, checkpoints, and review.
Want to know what an agent should actually be doing in your business? Type out the workflow that eats your week - you'll get it back mapped, free. Or if you'd rather talk it through, book a discovery call.