I run a group of companies staffed mostly by agents. So this is not a forecast, it is a field report: what an AI agent really is, where it earns its keep, and where it quietly burns money.
An AI agent for business is software that takes a goal, works out the steps, uses tools and acts, instead of just replying to a question. That last part is the whole difference. A chatbot answers; an agent does. In our companies an agent might read an incoming enquiry, check a system, draft a response and log the result, with a person reviewing anything that carries risk. I want to be useful rather than breathless here, so most of what follows comes from running these things day to day, not from a deck.
The market noise is deafening and worth cutting through. The agentic AI market is roughly $9.9 billion in 2026 and growing more than 40% a year, and Gartner expects 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. But intent is running far ahead of production. Only 17% of organisations have deployed AI agents so far, while more than 60% expect to within two years, and only about 23% of organisations have scaled an agentic AI system in production. The gap between wanting agents and running them is the whole game.
Strip the hype and an agent is three things bolted together: a model that can reason, a set of tools it can use, and a loop that lets it act, check the result and try again. That is it. The journal piece on what an agentic workflow is, without the hype lays out the parts in plain terms. What makes agents useful in a business is not intelligence in the abstract, it is that they can be handed a whole small job rather than a single step.
After running these across several companies, the pattern is consistent. Agents pay off on work that is repeated, has clear inputs and outputs, and needs reading plus a decision but not a human relationship. First-line support and triage. Lead qualification and routing. Document handling and data entry. Scheduling and follow-ups. The WhatsApp deal manager we built for an estate agency, described in the custom CRM case study, is really a set of agents doing intake and chase work a person used to forget.
The honest test is simple. If a task can be described as "when X happens, read the details, decide between a few options, and do the next thing", an agent can probably carry it. If it needs genuine human trust, judgement under ambiguity, or accountability that cannot be delegated, keep a person there. Most jobs are a mix, which is why we map them into steps first, as described in breaking a job into steps a digital workforce can staff.
The failure rate is real and I have contributed to it. Over 40% of agentic AI projects will be cancelled by the end of 2027, per Gartner, from escalating costs, unclear value and weak governance. Not one of those causes is the model. The projects that die are the ones pointed at a vague ambition rather than a specific process, built without clean access to the data the agent needs, and shipped with nobody owning the output. Every agent we run has a defined job, a human checkpoint, and a number it is supposed to move. The ones I killed early were the ones I could not describe in a sentence.
Pick one painful, repeated process. Give one agent that one job. Keep a person reviewing its output while you learn what it gets wrong. Prove the number, then widen. This is the shadow-mode-then-cutover approach in the piece on human oversight that does not become a bottleneck, and it is the single biggest reason our agents run in production rather than sitting in a slide. Autonomy is earned one proven task at a time, and by 2028 Gartner expects at least 15% of day-to-day work decisions will be made autonomously by agentic AI, up from 0% in 2024. The businesses that get there will be the ones that started narrow.
The framing that trips people up is "replace the team". That is not how it plays out, and chasing it is how budgets vanish. Agents replace tasks, not people, and the value shows up as a small team doing the work of a much larger one, with the humans moved onto the parts that actually need them. That is the model our companies run on, and the plain daily version of it is in a normal Tuesday. Get the framing right and the technology mostly takes care of itself.
An AI agent is software that can take a goal, decide the steps, use tools and act, rather than just answer a question. For a business that means it can work a process end to end, like reading an email, checking a system and drafting a reply, with a person reviewing what matters.
On repeated, rules-plus-judgement work with clear inputs and outputs: intake and triage, first-line support, lead qualification, document handling, scheduling. Anywhere a task needs reading and a decision but not a human relationship works well.
Often yes, because a small team feels the cost of routine work most. The trick is to point one agent at one painful process, keep a human in the loop, and prove the number before doing more. Small and focused beats broad and vague every time.
Not because the models are weak, but because projects skip the boring parts: a specific use case, clean data access, governance and an owner. Gartner expects over two in five agentic AI projects to be cancelled by 2027, almost always for those reasons.
If you want to point one agent at one real process and see it work before committing to more, that is exactly how we start. For the wider category picture, including the market numbers and the agent-washing problem, see the complete guide to agentic AI. The agentic workflows guide covers the shape of the work, and AI workflow automation covers the ROI. To scope your first agent, come to contact.
Tell us one process that eats your team's time and we will tell you whether an agent can carry it, where a human still has to sit, and what proving it would take.
Book a 30-minute fit call