What is agentic AI?

Agentic AI refers to systems that don't just answer questions — they take actions to complete a goal. Instead of waiting for each instruction, an AI agent can plan a sequence of steps, use tools and software, and adapt as conditions change, all in pursuit of an outcome you defined.

The difference from a standard assistant is simple: an assistant responds, an agent acts. Ask an assistant to draft an email and it writes one. Give an agent the goal of resolving a customer request and it can check the order, apply a policy, update the system, and send the reply — end to end.

That capability is what makes the market interesting. It's also what makes the autonomy question unavoidable.

How big is the agentic AI market?

The numbers point in one direction: up, and steeply.

According to MarketsandMarkets (2025), the agentic AI market is projected to grow from USD 7.06 billion in 2025 to USD 93.20 billion by 2032, a compound annual growth rate of 44.6%. For context, the same market was valued at USD 4.81 billion in 2024.

Adoption inside companies is moving just as fast. Gartner (2025) projects that by 2028, 33% of enterprise software applications will include agentic AI — up from less than 1% in 2024 — and that at least 15% of day-to-day work decisions will be made autonomously by agents by the same year.

The trajectory is clear: what's less clear, and more important, is how much of that growth will be fully autonomous, and how much will keep a person in the loop.

Fully autonomous agents vs. human-in-the-loop: what's the difference?

Neither model is "better", they answer different questions. The mistake is treating autonomy as a finish line instead of a setting you tune per process.

Why the market is bending toward human oversight (for now)

The hype says fully autonomous, but data says something more measured.

Gartner (2025) also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, and inadequate risk controls. The same analysts warn about "agent washing" — vendors relabeling older tools as agents — estimating that only around 130 of thousands of vendors offer genuinely agentic products.

Meanwhile, the companies actually getting value are the ones keeping humans involved. Several investigations found that high-performing organizations are more likely than their peers to have defined processes for when a model's output needs human validation. Oversight isn't the thing slowing them down. It's part of why they're succeeding.

This matches what we see in practice: the failures rarely come from agents being too weak. They come from agents being trusted too soon — with no one accountable for the outcome and no clear point where a person steps in.

So how should you decide how much autonomy to give?

Autonomy isn't a single switch, it's a decision you make process by process. A few questions worth asking before you hand over control:

1.    What's the cost of a wrong action? If an error is cheap to reverse, more autonomy is reasonable. If it's expensive or irreversible, keep a person in the loop.

2.    How well-defined is the task? Clear rules and clean data support autonomy. Ambiguity and judgment call for oversight.

3.    Who is accountable for the result? Autonomy doesn't remove responsibility — it relocates it. Name the owner before you deploy.

4.    Can you see what the agent did? If you can't audit the decision, you can't trust it at scale. Visibility comes before autonomy, not after.

5.    Does the process touch a regulated or high-trust moment? Money, health, legal, and customer relationships usually warrant a human checkpoint.

The pattern across strong deployments is the same: start with oversight, earn autonomy. Give the agent room where it's proven itself, and keep a person where the stakes are real.

The real projection isn't autonomous vs. human

The market forecast is easy to summarize: agentic AI is growing fast and isn't slowing down. The harder truth is that the winning model won't be fully autonomous or fully supervised. It will be a spectrum — set deliberately, where each process gets exactly as much autonomy as it has earned.

The companies that get this right won't be the ones that hand over the most control. They'll be the ones that decide, with clear judgment, where control belongs — and keep the human where it matters most.

That's the part no forecast can automate for you.

Frequently asked questions

• Will AI agents be fully autonomous?

Some will, for high-volume, low-risk tasks. But most enterprise deployments are expected to keep a human in the loop for high-stakes or regulated decisions. Gartner (2025) projects around 15% of day-to-day work decisions will be made autonomously by 2028 — meaningful, but far from total autonomy.

• What is human-in-the-loop AI?

Human-in-the-loop AI is a model where an AI agent proposes or takes actions while a person reviews or approves the critical steps. It balances the speed of automation with human judgment and accountability.

• Why do agentic AI projects fail?

Most failures when implementing AI appear to respond to escalating costs, unclear business value, and weak risk controls — often made worse by deploying autonomy before the process, data, and oversight are ready.

Deciding this alone is the hard part

Choosing how much autonomy each process should have isn't a one-time decision — it's the real work of putting agents into production. It's what we do every day.

At Lumen, we design and implement AI agents with the right level of oversight for each process: autonomous where it's safe and proven, human-in-the-loop where the stakes are real. No agent-washing. No handing over control you can't audit. Just automation that gives your team back hours, speeds up decisions, and keeps a person accountable where it matters.

You don't need more tools. You need a clear map of where agents fit in your operation — and how much control to keep. Let's build that map together