AI agents doing real work: what is realistic in 2026, and where it pays to wait

If you got three emails this week offering an "AI agent" that will run your company by itself, you are not imagining things. The word wore out. Gartner gave the practice a name: "agent washing", which is taking a chatbot, an old automation, or an assistant and relabeling it as an agent. Of the thousands of vendors claiming to sell agents, Gartner estimates around 130 offer something genuinely agentic. The rest is paint.

This matters because you are the one paying the bill. A real agent decides and executes several steps toward a goal: it reads a case, queries systems, makes a decision, acts, and accounts for what it did. A reskinned chatbot only answers questions. Paying the price of the first for the second is the most expensive mistake of the year, and it is easy to make when every piece of sales material uses the same three words.

It helps to read two figures side by side. McKinsey, in its November 2025 State of AI report, found that 62% of organizations are at least experimenting with AI agents. It sounds like the train already left. But the same report adds that in no business function have more than 10% of companies managed to scale those agents, and that only 39% attribute any earnings impact to AI. Plenty of trials, few that pencil out.

The most honest projection came from Gartner itself: it estimates that more than 40% of agentic AI projects will be canceled before the end of 2027, due to runaway costs, unclear business value, or inadequate risk controls. This is not a columnist being gloomy. It is what happens when a technology gets bought on the promise instead of on a measurable problem.

For an SME, the takeaway is not "this does not work". It is: the average fails, and the average fails by choosing the wrong place to start. The difference between companies that get value and those that cancel is not the model they use. It is having picked a concrete, repetitive, measurable process before signing anything.

There is ground where agents already do real work, and it is almost always the same kind of ground: high-volume tasks, clear rules, available data, and a tolerable cost of error. Picture a distributor that receives two hundred purchase orders by email a day, each in a different format. Read the email, pull the data, load it into the ERP, and flag the ones that do not match: a well-scoped agent pays off from the first week, because the work is repetitive and a human reviews the exceptions.

The same applies to first-line support, to reconciling payments against invoices, to drafting quotes from a catalog, or to classifying and routing incoming tickets. These are tasks where the agent speeds up a person rather than replacing them. The Klarna case made it plain: its assistant came to handle two-thirds of support conversations, the work equivalent of 700 people. And even so, in 2025 the company rehired humans for complex and sensitive cases, because quality dropped there. The lesson is not that AI failed. It is that it worked exactly where it was supposed to, and no further.

A useful rule for deciding: if you can write the procedure on one page and a new hire would understand it, an agent can probably assist with it today. If the procedure lives in your operations manager's head and changes depending on the client, not yet.

Waiting is not falling behind. It is choosing not to pay to be the guinea pig. There are three zones where, in 2026, getting ahead costs more than it returns for an SME.

The first: high-impact autonomous decisions without supervision. An agent that moves money, grants credit, fires suppliers, or replies to a regulator with no human signing off. Today's models do not hold complex goals or nuanced instructions over time, and the cost of an error there is not an annoyed customer, it is a fine or a lawsuit. The second: processes that sit on top of messy data. If your information lives split across the ERP, Excel sheets, WhatsApp, and the memory of three people, no agent will tidy that chaos for you. You put the house in order first, then you automate.

The third: buying the full "agent platform" before you have a single process solved. It is tempting to sign the big suite that promises to cover everything. But through the end of 2026, most of the real value sits in a handful of narrow tasks, not in a layer that orchestrates your whole company. Starting small is not timidity; it is how you avoid ending up in that 40% that gets canceled.

Before evaluating tools, it is worth lining up three questions. Which concrete task costs you hours and repeats every week? Is the data for that task somewhere a system can reach, or does it live in heads and loose files? What happens if the machine is wrong one time in twenty: does someone catch it in time, or does it go straight to the customer?

If all three answers are good, you have a real candidate to start with, and it is worth treating it as a trial with a measurable goal: how many hours it saves, how many errors it prevents, how quickly it pays for itself. McKinsey was clear on one point that often gets lost: the biggest earnings impact comes not from the model but from redesigning the workflow around it. Buying the tool and pasting it on top of the old process is the most common recipe for seeing no return.

If the answers are not good yet, the best AI decision of 2026 might be to fix the data first, or to buy nothing this quarter. That is also strategy.

The difference between companies that win with agents and those that cancel them is rarely budget. It is judgment: knowing which process to pick, what to leave for later, and what not to buy yet. That gets decided by looking closely at your operation, not by reading a vendor's brochure.

If you are looking at AI agents for your SME and you are not sure where to start, let's talk it through with no commitment. A short diagnostic is usually enough to tell apart the two or three tasks where getting ahead truly pays from the ones better left to mature. Sometimes the best recommendation is "not yet", and we would rather tell you that before you sign anything.