More autonomy was supposed to mean more value
For the past two years, a common belief in enterprise AI has been simple: the more an agent can do on its own, the better. Give it room to plan, decide and act across many steps, and the results will follow.
A recent VentureBeat article by Midhula Mariyam Jeevan argues that this assumption is now being tested in real production environments, and that in many deployments it is failing. The companies that benefit from agentic AI, the article says, will not necessarily be the ones that gave their agents the most freedom. They will be the ones that built agents with specific responsibilities and made sure those agents operate within clear rules.
This post walks through the evidence the article presents, the four patterns it sees in teams that are getting this right, and how to apply them to your own agents.
Capability is outrunning control
The article points to two data points. First, it cites Gartner's forecast that more than 40 percent of the agentic AI projects running today will not survive to 2028. The reasons given are escalating costs, unclear business value and inadequate risk controls, not weak models.
Second, it cites McKinsey's 2026 AI Trust Maturity Survey, which found that average responsible AI maturity sits at 2.3 out of 4, and that only about 30 percent of organizations have reached level three or higher in governance and agentic AI controls.
Put together, the article's conclusion is that capability is outrunning control. It frames the next phase as a trust race. The question is no longer who can build the most capable agent, but who can get an agent approved for production by risk, legal and compliance teams, and keep it approved once it is live.
Why full autonomy breaks down
According to the article, Gartner describes a repeatable failure pattern. Projects launch with broad, highly autonomous workflows, run into integration complexity within weeks, and then stall without a clear path to return on investment. The article also cites Gartner's count that only around 130 of the thousands of products sold under the agentic AI label have real autonomous capability behind them.
There is a deeper structural problem too. The article puts it plainly: autonomy and accountability move in opposite directions. An agent that can plan and execute a long chain of steps on its own is also an agent whose individual decisions are harder to trace afterward. When something breaks several steps into a chain, working out why it happened and who is responsible becomes a project in itself.
The article adds that nearly two thirds of businesses now say security and risk issues are their greatest challenge in scaling agentic AI, ahead of regulatory uncertainty and technical barriers.
Four patterns of governed orchestration
The teams that are leading, the article says, are not slowing down. They are changing how autonomy is distributed. It highlights four patterns.
1. Narrow agents over general ones
Break an end to end process into single responsibility agents with tightly bounded jobs. A smaller scope of work means a smaller scope of failure, which is far easier to audit.
2. Human checkpoints before the outcome
Place review at decision boundaries, before a sensitive action happens rather than after. That means a person checks before sensitive data moves, a transaction posts or an outside system is triggered.
3. Traceability by design
A full record of what each agent did, and why, should be available on demand. It should not need to be reconstructed under pressure during an audit. The article notes that the EU AI Act's human oversight requirements for high risk systems are still coming, with the compliance deadline moved to December 2027.
4. Data placement as containment
Where an agent's data lives, and who can reach it, decides how contained a failure can be. Tighter access limits how far a misbehaving agent can reach.
Calibrated control, not maximum control
The article is careful to note that the risk runs both ways. An agent that needs a person to approve every minor step has not removed any friction. In its words, that is automation wearing a manual process as a costume. The goal is calibrated control, concentrated where the cost of an error is actually high.
That is a useful way to think about any agent you build. Routine, reversible steps can run on their own. Steps that spend money, contact a customer, change an important record or move sensitive data deserve a checkpoint.
Four questions to ask about your agents
The article offers a simple review framework. You can use it on any agent, whether it is already live or still on the whiteboard.
- Can you reconstruct why an agent took a specific action months from now? If the honest answer involves digging through raw logs, traceability is an afterthought.
- Does every agent have one clearly bounded job? An agent told to "figure it out" across a broad task is where compounding errors start.
- Are human checkpoints placed before key decisions, or only after the agent has acted? A review after the fact catches consequences. A checkpoint before prevents them.
- If an agent malfunctioned right now, how much could it touch before anyone noticed? Access scope is your containment plan.
How to apply this in practice
You do not need an enterprise governance program to start. A few habits go a long way.
- Write each agent's job in one sentence. If you cannot, split it.
- Give each agent access only to the tools its job requires.
- List the actions that are hard to undo and put an approval step in front of each one.
- Decide where the record of each run lives and who reviews it.
- Start with routine, reversible work and widen the scope only as results hold up.
Where Innflow fits
Innflow is the all in one workspace for AI agents. It is built around the kind of bounded autonomy this article describes. You give each agent a focused task with clear responsibilities, connect it only to the tools it needs, and choose where approval is needed before work goes out. Workflows let you set conditions and assign the next step to a teammate or an agent, so predictable steps stay predictable. And you can follow every agent's progress from one workspace instead of piecing it together afterward.
The aim is not to slow agents down. It is to let them handle the routine volume while people keep the final say on the decisions that matter.
The real advantage
The article closes with a clear point. The winners will not be whoever deployed the most autonomous agents fastest, but whoever built agents trustworthy enough that risk and compliance teams stopped being the bottleneck. Scoped jobs, checkpointed decisions and a clear record are what make that possible. If you want one place to build agents that way, you can try Innflow at innflow.ai.
Frequently asked questions about AI agent deployments
Why do the best AI agent deployments limit autonomy?
Broad, highly autonomous projects tend to break down, while bounded agents with clear controls are easier to trust and scale.
Does limiting autonomy mean adding approval to everything?
No. The aim is calibrated control. Too many approvals slow the work, while too few raise risk.
How can I review my own agents?
Write each agent's job in one sentence, list the actions it may take alone, and decide which ones need a person.
Related reading
- Orchestration: The New CX Challenge for AI Agents
- 5 Steps to Give AI Agents Clear Permission Boundaries
- AI Control and Execution Planes: Make Permissions Traceable
Source: Midhula Mariyam Jeevan, VentureBeat, "Enterprises winning with AI agents are limiting how much the agents can do alone" (August 22, 2026). Read the original.












