Most software waits to be used. You open it, you click, you get a result, you close it. An agent is different in a way that takes a while to sink in. You give it a goal and some access, and it goes and does the work, checking in when it needs you. That sounds like a small change. In practice it feels a lot more like adding a member of staff than installing a tool.

That is why a growing number of teams have stopped talking about buying AI and started talking about hiring it. The language is only half a joke. The way you bring an agent into a company is starting to look like onboarding.

From tools to teammates

The scale of this shift is easier to see in forecasts than in any single product. Gartner expects that by 2028 a third of enterprise software will include agentic AI, up from less than one percent in 2024, and that around 15 percent of day to day work decisions will be made autonomously by these systems, up from none at the start of 2024. Whatever the exact figures turn out to be, the direction is clear. Software is moving from something you operate to something that acts.

Once a system can take actions on its own, the questions you ask about it change. You stop asking whether it has the right features and start asking what it is allowed to do, who reviews its work, and what happens when it gets something wrong. Those are management questions, not procurement questions.

Where agents start

The first roles tend to be the same ones a growing company hires for early. Customer support is the obvious one, because the volume is high and many questions repeat. Klarna's assistant handled the equivalent of 700 agents worth of conversations within a month of launch, which shows how quickly an agent can absorb a function that used to need a large team.

Operations is another early home. Agents are good at the connective work that holds a business together, things like moving data between systems, chasing missing information, drafting routine documents, and flagging anything that looks off. Analysis is a third. An agent that can pull numbers, summarise them, and answer follow up questions gives a small team the kind of support that used to require a dedicated analyst.

Managing a workforce you cannot see

The mistake is to treat an agent like a finished product and walk away. The teams that get good results treat it like a capable new hire who is fast but green. That means giving it a clear scope, setting limits on what it can touch, and reviewing its output closely at first before loosening the reins as trust builds.

It helps to write down what the agent is responsible for, just as you would for a person. What does a good result look like. What should it never do without asking. Who owns the outcome when it acts. A junior employee with no brief and full access to your systems would be a risk. An agent is no different.

Onboarding an agent

Bringing an agent into a team works best when it looks like a real onboarding rather than a switch being flipped. Start narrow. Give the agent one clear job, a defined slice of access, and a human who owns its output. Let it run alongside the existing process for a while so you can compare results, the same way you would let a new hire shadow someone before handing them the keys.

From there the path is familiar. You correct the early mistakes, tighten the instructions, and widen the scope as the work proves reliable. The companies that skip this step, dropping an agent straight into production with broad permissions and no review, are the ones that end up in the cancellation statistics. The technology is rarely the reason these projects fail. The lack of a sensible onboarding usually is.

The honest limits

It would be dishonest to pretend this always works. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, usually because the cost outran the value or the controls were never there. Klarna itself is a cautionary tale as much as a success one. After leaning hard into automation, the company said in 2025 that it had gone too far in places and began reopening some human support roles, noting that quality human help is worth paying for at the premium end.

The lesson is not that the AI workforce is hype. It is that an agent is a hire, and hires can be wrong for a role, badly briefed, or pushed into work they are not ready for. The companies doing this well are not the ones automating everything. They are the ones who think carefully about which jobs to hand over, bring the agent on with the same care they would give a person, and keep a human accountable for the result.