AI agent autonomy advances, but enterprise control lags

By Tom Whitfield

62% of IT leaders reported building, deploying or developing AI agents in KPMG's July and August survey, up from 53% in the previous quarter. The article does not state the survey sample size, so the rise indicates respondent momentum, not a precise market adoption rate.

Enterprise interest has outpaced willingness to surrender control. Employers must decide which processes agents may run, which employee approvals remain mandatory, and how access, audit trails and accountability will work; these are governance choices, not proof that oversight can disappear.

What does agent autonomy change for employers?

Agentic AI promises systems that can maneuver through enterprise work, make decisions and take actions autonomously. Few IT leaders appear ready to give bots widespread control, with low-risk tier 1 IT support among the few plausible early exceptions.

ModelOp CTO Jim Olsen sees limited autonomous use for processing customer refunds or credits without human intervention, but says early lessons keep full-autonomy risks in view. The exposure depends on what the agent can reach.

Olsen said: “We saw the large adoption of Clawdbot and OpenClaw by individuals where they would allow it full access to their email and gave it full autonomy.” He added that success for some users still left them with “a large risk exposure.”

Olsen recommends putting autonomous bots in a strict sandbox governed by well-defined policies. He does not expect unrestricted autonomy in enterprises in the near future, because restricted autonomy is already in use.

He expects non-mission-critical work without access to sensitive information to receive full autonomy first. As trust improves, he expects the range of use cases to widen.

Takeaway: Employers can grant autonomy now, but only inside boundaries suited to low-risk work.

What controls make autonomy safer?

IT leaders are concerned about “autonomy creep,” in which agents quietly gain more authority than intended. Human-in-the-loop checks offer limited protection when employees lack the authority or time to correct errors, according to AI and security experts.

Available controls include API gateways, operating-system enforcement and credential management to keep behavior within acceptable goals. Olsen says leaders need clear policies tied to each agent and use case; available tools still need a cohesive strategy to produce enough trust for minimal autonomy.

The source supplies no independent error rates or evidence that these controls work together. That omission matters because technical restrictions cannot, by themselves, establish who is accountable for an agent's decision.

Takeaway: Technical controls can limit behavior, but leaders still need a coherent policy for each agent and task.

How much authority should agents receive?

With about 2,000 agents deployed within her enterprise, Barracuda Networks CIO Siroui Mushegian still has employees checking bot actions. She describes many agents as close to autonomous and places ticketing automations in a “trust-but-verify” category, with employees checking the agents that oversee them.

Successful results have increased her comfort, but not her willingness to eliminate accountability. She remains particularly cautious about auditable processes and says she does not want zero human accountability.

Mushegian expects autonomy to expand when agents demonstrate reliable performance. In her view, the likely consequences of a mistake should determine how much evidence leaders need before granting more authority.

Her boundaries favor narrow, correctable work and keep high-impact changes under review:

• Routine internal support requests can run autonomously when the answer is well documented, the action is limited and errors are straightforward to correct.
• Agents that change access permissions or production systems are less likely candidates for autonomy.
• Mixed workflows can combine action and approval. Mushegian would give an agent only the authority needed for the task, with human approval where mistakes could have serious consequences.

Takeaway: Wider authority should follow reliable performance, while approval remains necessary for decisions with serious or hard-to-audit consequences.

Which actions should still require approval?

Taupia co-founder and CTO Joshua Rowe would draw the line at the action rather than the agent. The same agent can act alone on one step and require sign-off on the next.

He sees opportunities for near-autonomous work now, including data gathering, triage, drafting and comparing options. Taupia is testing an agent that compares energy deals and helps customers switch providers without employee approval at each step; customers must still review and confirm the change before it takes effect.

Rowe says autonomous authority depends on making each action reversible or bounded, not on whether AI models become smarter. His controls include:

• spending limits;
• cooling-off periods;
• rollback;
• an audit trail sufficient to undo a mistake.

Once those protections exist, he says the action can proceed autonomously. He would retain human approval for the foreseeable future where an action has no undo path, such as signing a contract or moving money from an account.

Takeaway: Reversibility can justify autonomy, while irreversible actions should keep a person accountable for approval.

What does meaningful human oversight require?

Text CTO Maciej Walaszczyk identifies customer service as an area where IT leaders have begun to cede control to agents. He says agents can answer customer inquiries without constant supervision as they learn from knowledge provided by humans. Permissions, consequences and tailored response limits determine the scope of autonomy.

Walaszczyk does not advocate handing an entire department to an agent. Leaders should instead delegate defined processes with clear goals, boundaries and verifiable outcomes; an agent acting for someone cannot exceed that person's authority.

He frames this as deliberate design of the environment, permissions, instructions and oversight. Organizations should build trust through real work and verified results, with people checking outcomes in many cases.

He said: “Oversight must be real. It takes time and attention, and it means much more than clicking approve.”

Saying an AI made the decision does not remove the organization's responsibility. Irreversible actions or decisions with major consequences still require human review.

Takeaway: The likely next step is not hands-off departments, but narrower autonomy that expands as permissions, reversibility and meaningful oversight improve.