From Copilot to Autopilot

There's a lot of noise around AI agents, and much of it treats "using AI" as a milestone a company either reaches or doesn't. That isn't how adoption works. It runs along a curve, from software that helps a person complete the work to software that performs the work itself. Almost every company is already somewhere on that curve, whether it has named the progression or not.

The automotive industry mapped a similar problem years ago. The SAE levels of driving automation range from systems that assist the driver to systems that operate without continuous human control, and the useful insight is that "self-driving" isn't one capability. It's a progression defined by how responsibility is divided between the human and the system. Researchers studying AI agents have reached much the same conclusion. One framework for the levels of agentic autonomy describes increasing independence by the role the human keeps, moving from operator toward observer.

For a business, that progression is a useful operating lens, and at Muan Group we think about it as the Muan Autonomy Curve.

The Muan Autonomy Curve

Assisted. AI drafts, suggests, summarizes, and analyzes, but the person still performs the work and makes the decision. This is the copilot stage: the system might write the ad, summarize the call, or prepare the report, but the human stays responsible for execution. Productivity rises while the labor and the judgment stay human.

Supervised. AI proposes or prepares an action, and a person approves it before execution. The system can draft the customer response, recommend the next step, or tee up an offer, but nothing meaningful happens until someone confirms it. The human is still in the loop, inside the operation.

Delegated. AI performs the task end to end and involves a person only when something falls outside expected conditions. Here the human moves to oversight of the loop: no longer reviewing every step, but watching performance and handling the exceptions the system flags. This is where the role shifts from working inside the process to supervising it.

Autonomous. AI operates toward an objective within predefined boundaries. The human sets the goal, the economics, the permissions, and the guardrails, and the system decides how to execute inside them, so the person reviews outcomes rather than individual actions. This is the autopilot.

The distinction matters because most conversations about AI adoption fold all four stages into one word. A company using generative AI to draft emails and a company using an agent to run thousands of customer interactions might both say they "use AI," but operationally they're doing very different things.

The real gap is pilot to production

The direction of enterprise technology is clear. Gartner expects up to 40% of enterprise applications to include task-specific AI agents by the end of 2026, compared with less than 5% in 2025. In its best-case scenario, agentic AI could reach around 30% of enterprise application software revenue by 2035.

But adoption and production are different things. McKinsey finds nearly nine in ten organizations use AI in at least one business function, yet only about a quarter are scaling an agentic system anywhere in the enterprise. Demonstrating an agent in a controlled setting is one thing; running it continuously across real customers, real money, and real processes is another.

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, on rising costs, unclear business value, and inadequate risk controls. It has also warned about widespread "agent washing," where ordinary automation or a chatbot is repositioned as an autonomous agent without meaningfully changing what the system can do. That's the real divide in enterprise AI: not proving an agent can work once, but building one that performs reliably, economically, and safely thousands of times. The distance between a stage-two demo and a stage-four operating system is where much of the value, and much of the execution risk, sits.

The objective is appropriate autonomy

Reaching stage four is not the goal for every process. AI maturity isn't measured by how many operations a company can make fully autonomous, but by whether it can match the right level of autonomy to each one.

Some workflows suit high autonomy. A system that schedules appointments, qualifies leads against defined criteria, routes information, or follows up on predictable outcomes can often operate within clear boundaries. Others involve decisions where human judgment remains essential, such as a financial exception, a legal decision, a complex dispute, or a high-impact personnel decision. Those processes may belong at the supervised or delegated stage indefinitely. The goal isn't to remove the human from every process. It's to place human judgment where it creates the most value.

The goal is not maximum autonomy. It is appropriate autonomy.

— Muan Group

A highly autonomous system with no escalation rules isn't mature, it's just hard to control. The strongest systems know both how to act and when to stop.

How we apply the curve

For Muan Group, this curve is more than a view of where enterprise software is heading. It is an operating framework for how we build and evaluate the portfolio.

Global Connect runs toward the delegated and autonomous end, one outcome at a time. MIA can place the call, qualify the lead, book the appointment, manage follow-up, and re-queue no-shows. That workflow can move further along the curve precisely because the objective is bounded. The system knows what outcome it's after, what qualifies a customer, when to book, and when to escalate to a person. The human doesn't need to sit in every successful interaction, only the ones that need judgment.

ASIO moves non-technical users further along the curve by turning complex campaign execution into software-managed operations. The operator defines the objective and the system runs the work underneath, so instead of coordinating every step by hand, the user focuses on the outcome and the constraints around it.

And we climb outcome by outcome, not all at once. Autonomy is earned process by process: identify an operation, define the objective, set the acceptable boundaries, measure the result, and decide where a human still belongs. Then the system can move further along the curve. That approach builds something more valuable than an impressive demo. It builds infrastructure that can actually operate.

What comes next

The next phase of enterprise AI won't be defined by how many companies use it, since that threshold is already crossed. The more important question is how responsibility gets divided between people and software across each operation. Some processes still need a person at every step. Others can move to approval-based supervision, run independently until an exception appears, or operate toward a defined objective with humans focused on outcomes.

Those answers will increasingly shape how organizations design their teams, their software, and their operating models. The advantage will not come from having more AI. It will come from knowing where autonomy creates leverage, where human judgment creates value, and how to design the handoff between the two on purpose.