Starring AI agents: In and as employees

On a new employee’s first morning, someone hands them a badge, explains who they report to, tells them what they can and can’t do, and quietly starts watching how good they are. None of that is optional. It is how every company on earth turns a stranger into a productive colleague.

Now ask: when your company turned on its first AI agent, did anyone do any of that?

For most organisations, the honest answer is no. The agent was deployed. It was not onboarded. BCG’s own research on AI transformation found that just 10% of success comes from the algorithm and 20% from the technology—the remaining 70% comes down to people and process. And that gap—not the underlying technology—is where the value is being lost.

This is not a doomsday story about agents replacing the humans who do the work. It is a mental model for leaders trying to organise a company that can thrive in an AI-first world, and the argument behind it is simple: organisations that treat AI agents as employees, rather than software, will extract the most value from this transformational resource.

There are three reasons why.

First, agents act. Every generation of enterprise technology before this one—the ERP system, the CRM, the dashboard tools—has done one thing: hand a human the facts and wait. An AI agent can look at the same facts, decide what to do, and go do it, across multiple systems, without a person approving every step. That is a different category of resource, and it deserves a different category of management. You don’t instal judgment. You supervise it.

Second, agents cost money every time they work, not once, like software. Traditional software has a steep setup cost, then a marginal cost near zero: build it once, and running it a million more times barely moves the bill. An agent is the opposite. Every time you call on it—even for a task it has “already learnt”—it burns tokens, and the meter runs

again. That is closer to a salary than a software licence, which means agents deserve the ROI scrutiny of a hire, not a system upgrade. BCG’s work with retail banks found that moving from simple AI copilots to agents running full, end-to-end workflows could lift profitability by up to 30% and cut costs by 30-40% by 2030—but only for the organisations willing to treat the shift as a change in how work is organised, not a software rollout.

Third, agents learn, and they make work visible that used to be invisible. Every judgment call and manual workaround that once lived quietly inside one employee’s head now runs through a system that can be watched, measured, and improved, because an agent leaves a trail every time it acts. In a BCG and MIT Sloan Management Review survey of more than 2,100 executives worldwide, 76% said they now think of agentic AI less as a tool and more as a co-worker—one expected to learn and improve over time. 66% of organisations already using agentic AI expect it to fundamentally change their operating models, roles, and career paths within three years.

But there’s a trap here worth flagging first. “Treat agents as employees” should never mean giving them a name, a personality, and most importantly a pass. It means giving them what a real employee gets: a defined role, a manager, and a clear boundary around what they can do without asking first. Singapore’s technology regulator made almost exactly this case when it published the world’s first governance framework for agentic AI at Davos this January: every agent, it argued, needs a traceable identity and a human accountable for what it does.

So what does fully integrating an agent into your organisation look like? It comes down to four organisational interventions:

1. Establish guardrails—for the enterprise and for every agent within it: exactly which data and systems it can touch, and which it categorically cannot.

2. Define clear goals—a specific outcome/s each agent owns, and a clear line between which decisions are its and which stay with a human.

3. Set the operating model—how the agent works with the team: does it act only with sign-off, act and then report back, or run independently with spot checks? Any of the three can be right. None of them should be accidental.

4. Build a monitoring system for drift—a way to catch an agent quietly drifting from what it was built to do, with one named, accountable human on the hook.

Most companies spend enormous energy onboarding an analyst and almost none integrating the agent now doing adjacent work. That asymmetry is backwards, and no slide full of bullet points will fix it. Only structure will.

The real pivot here is not from thinking about AI as a technology to thinking about it as magic. It is from thinking about AI as a technology to thinking about it as a colleague you are responsible for managing well. Get it right, and agents become some of your most productive investments. Get it wrong, and you end up spending a fortune on an employee no one bothered to train.

(Chitkara is India Leader, People & Organization Practice at BCG; Priyadarshi is Partner and Associate Director – People Strategy at BCG. Views are personal.)