Microsoft Copilot’s ‘Trash Bin’ Moment: Linas Beliūnas on Enterprise AI’s Data Problem

In a recent LinkedIn post, Linas Beliūnas offers a sharp critique of Microsoft Copilot, likening it, with a touch of hyperbole, to “the most expensive trash bin upgrade in the history of tech.” Beliūnas frames the widely anticipated AI tool not as the promised “AI coworker” or “productivity multiplier,” but rather as an “eager intern” that often proves intrusive and unreliable. He highlights common user frustrations, suggesting that the tool’s shortcomings stem from a fundamental misunderstanding of how AI interacts with complex enterprise systems.

The Gap Between AI Promise and Enterprise Reality

Beliūnas argues that the core issue lies in attempting to “duct-tape a large language model onto decades of messy SharePoint folders, half-broken Excel logic, email threads from 2012, [and] PowerPoints nobody owns.” He contends that this approach fails to deliver on the promise of “enterprise intelligence” because it overlooks the critical differences between probabilistic AI models and deterministic enterprise software.

“LLMs are probabilistic. Enterprise software is supposed to be deterministic. One guesses while the other files payroll 🤷‍♂️”

This fundamental disconnect, according to Beliūnas, leads to user experiences where Copilot feels “intrusive instead of invisible,” prompting jokes among finance and tech professionals about disabling it, a sentiment he notes was previously reserved for the much-maligned Clippy.

Beyond the ‘Failure’: A Transition Point for Enterprise AI

Despite his critical assessment of Copilot’s current state, Linas Beliūnas emphasizes that this is not the end of the AI journey for enterprises. Instead, he views it as a crucial “transition point.” He points to ongoing investments by companies like Salesforce and IBM in what he terms “knowledge pipelines, data plumbing, and context engineering.” This shift signifies a move away from a passive reliance on AI to “figure things out” towards a more proactive approach:

“Less ‘AI will figure it out’ → more ‘let’s structure reality so AI actually can.’”

This focus on foundational data infrastructure and workflow structuring is, in Beliūnas’s view, the key to unlocking the true potential of enterprise AI. He asserts that the critical lesson emerging is that AI does not replace existing systems but rather “exposes how broken they are.” The struggle of tools like Copilot, he explains, is not due to the inadequacy of AI models themselves, but because organizations have attempted to bypass the essential, albeit difficult, work of preparing their data, refining workflows, establishing ownership, aligning incentives, and solidifying their architecture.

The Future: Invisible and Embedded AI

Looking ahead, Beliūnas predicts that the next generation of enterprise AI will not be perceived as a distinct feature but will become an integral, almost imperceptible, part of the work environment. He envisions AI that is “invisible, reliable, and embedded where the work actually happens.” This future state suggests a seamless integration where AI’s capabilities are leveraged without the friction and intrusiveness experienced with current iterations.

In conclusion, Linas Beliūnas’s analysis suggests that while initial enterprise AI deployments like Microsoft Copilot may fall short of expectations, they serve a vital purpose in highlighting the underlying challenges of data and systems integration. The real progress, he argues, lies in the industry’s renewed focus on building robust data foundations and intelligent workflows, paving the way for truly transformative, yet invisible, AI integration in the future.