A couple of weeks ago, Manus, the Meta-owned agentic AI platform, launched a desktop version, an AI agent that runs locally on your machine. You give it access to a folder, and ask it to organize, rename, summarize, etc., and it does that, running locally. Just a few days later, Anthropic released their version, called CoWork, an AI agent that you can deploy, just install, actually, on your PC. It can code, as well as perform similar tasks.
Agentic AI is now accessible to businesses and individuals alike. Use it to code, prepare reports, generate highly personalized textual content, find patterns, or even send emails and calendar invites. AI agents are not just chat interfaces running locally. They can operate across tools, execute multi-step tasks, and act on instructions with minimal structure. Just natural language will do. These are assistants, in the more literal sense of the word, and can do what you allow them to do. That’s the catch.
For the first time, we have a situation where it’s not the ability of technology that’s the constraint, but our acceptance and willingness to harness it. Data governance is the single biggest reason that keeps enterprises from doing pilots with AI chat or agents, let alone full-fledged deployment.
Data Governance: Rational vs Paranoid (Expanded)
As individuals explore these tools with curiosity, enterprises remain cautious. Their caution is backed by good reason. Anthropic’s positioning of CoWork is telling. It is not being framed as a system for decision-making or autonomous execution in critical workflows.
IDC reinforces this caution more bluntly. In an analysis of agentic AI in the enterprise published in December 2025, it highlights that early deployments without governance can lead to serious consequences, including compliance failures and operational risk. Some organizations, it predicts, will face regulatory and leadership fallout by the end of the decade due to poorly controlled AI autonomy.
As valid and timely as the caution is, it is often taken to mean “Treat all data as sensitive,” when what it is actually saying is, “Be careful with sensitive data.”
Better safe than sorry, yes. But in this case, fortune favours the discerning.
If a business treats all data as equally sensitive, it has effectively opted out of this technology. Oftentimes, the line between caution and paralysis is thin and dynamic. It needs active calibration. The ability to manage this line is what defines a calibrated start and unlocks efficiencies from the word go.
IDC frames this as a dual reality: risk and reward coexist, and the differentiator is governance maturity.
Consider this: In most organizations, a large portion of internal documents are low-sensitivity, and significant work can be done with anonymized or masked inputs. Many workflows involve public or already abstracted data.
This leads to a more pragmatic framing: The question is not whether to use agentic AI. It is where you can safely start.
Organizations that adopt a rational classification approach, instead of blanket restrictions, unlock immediate gains:
- Faster turnaround on internal reporting
- Reduced manual effort in documentation and synthesis
- Improved utilization of skilled employees
Those that don’t will wait for perfect policies, perfect tooling, perfect certainty. We don’t know when and if such a state will come.
What are AI agents being used for?
A recent review of Manus by MIT Technology Review noted that while the system is not yet reliable enough for high-stakes autonomy, it performs well on structured, bounded tasks with clear instructions and human oversight. In other words, exactly the kind of work that consumes a disproportionate share of knowledge worker time today. And that’s the uncomfortable truth enterprises tend to ignore: 20–40% of knowledge work is operational glue work, necessary, repetitive, and low-risk.
Early usage patterns and independent reviews, including MIT Technology Review’s analysis of Manus point to a consistent set of use cases: organizing files, drafting reports, and synthesizing information. Summarizing a few files works well in a chat interface. But when the task scales, multiple folders, layered data, cross-references, agents become significantly more effective.
This reduces not just effort, but context switching, one of the biggest hidden drains on productivity.
Competitive Edge Through Early, Controlled Adoption
IDC makes a critical point: “The next era belongs to organizations that govern autonomy, not constrain it.”
That distinction matters. Businesses where governance is pragmatic and evolves through adoption will have an edge over those that see set-in-stone governance as a prerequisite. Not all deployments need to be “AI transformation” stories. They can be incremental across the business, with advantage coming from compounding small efficiencies.
Perhaps the right way to approach agentic AI in the enterprise is by asking the right question. Instead of “Is the technology ready?”, we should ask, “What low-risk work are we still doing manually that we don’t need to?”
