Mirza Beširović was a translator first, which turns out to be the entire job

Every AI product team is currently discovering, at great expense, that the hard part was never the model. Mirza Beširović had a head start on that realization, because they did not come up through engineering or an MBA pipeline. They came up as a translator and interpreter, a linguist by training, in a job whose entire purpose is standing between two parties who cannot understand each other and making the meaning survive the trip.

Which, if you think about it for four seconds, is the job description for an AI product leader. A model can do a set of things. A person needs a different, adjacent set of things. Almost all the value lives in the gap between those two, and closing that gap is translation work, not machine learning work.

Pilot to a hundred million

The proof is the scoreboard at Zendesk, where they run product for AI Agents. It went from pilot to over $100 million in annual recurring revenue with the customer base quadrupling, and it now automates enormous volumes of customer conversation for Netflix, Levi's, Vimeo, and Reddit. That is not a demo that impressed a conference. That is a business, in production, where being wrong is expensive and extremely visible.

Getting there means solving the parts nobody posts about: what an agent should refuse to do, how it hands off to a human without the customer having to repeat themselves, what happens on the failure path, and how you convince an enterprise buyer to trust any of it. Reliability and handoff are product design problems dressed as technical ones, which is exactly the sort of thing a linguist notices first.

Why the pilots die

The talk of theirs worth your time is titled Why AI Pilots Die: You Can't Buy Your Way Out of an Org Problem, and it is the most useful contrarian take in enterprise AI right now. The industry's default explanation for a failed pilot is that the technology was not ready. Mirza's argument is that the technology was usually fine and the organization was not: unclear ownership, no one accountable for the outcome, teams that were never restructured to absorb what the thing actually changes. You cannot procure your way out of that, however large the contract.

It is the same insight as the translator's, scaled up. The bottleneck is never the machine, it is the humans and machines failing to understand each other, and someone has to do the work of making the meaning survive the trip. Their site says it in one line: building AI products and the orgs that ship them. The second half is the part everyone skips.