Phrase, AWS, and Uber on The Future of AI Agents & Multilingual Content

At SlatorCon Silicon Valley 2025, Phrase convened a panel on one of the most pivotal questions in enterprise AI today: what AI agents mean for the way companies operate and for how global content is created, adapted, and delivered.

AI agents go beyond simple task execution or automation. Instead, an AI agent pursues a defined goal, decides which tools and steps to use, reflects on outcomes, and adapts its strategy until it succeeds.

The panel brought together three complementary perspectives on how the agentic future of enterprise AI is unfolding. Semih Altinay, VP AI Solutions at Phrase, articulated a vision for how agentic workflows can link global content teams and make localization a core enabler of global enterprise performance. 

Juan Marcano, Software Engineer at Uber, offered the SlatorCon delegates a rare inside look at an in-house AI agent that autonomously explores the Uber app and finds bugs.

Govind Varadan, Manager of Solutions Architecture at AWS, provided a panoramic view of enterprise adoption, drawing on decades of experience in digital and cloud transformation to show how organizations can now take the next step into AI.

The Concierge

The panelists started by establishing a shared understanding of what AI agents are. “An AI agent doesn’t just accomplish a task,” Altinay told the SlatorCon audience, “Rather, it needs to figure out what tools to use to accomplish that goal.”

On top of these two characteristics of autonomy and goal orientation, Marcano added a third: adaptability. “An agent thinks on its feet and works around whatever comes up,” he said.

The panelists cautioned that the term AI agent is used loosely across industries, spanning everything from narrow task-based assistants and collaborative copilots to fully agentic, concierge-like applications.

Surprise Handlers

An example of a genuinely agentic workflow was shared by Marcano. The Uber software engineer told the audience how he developed an internal AI agent for testing and bug finding in the Uber app.

“We taught the agents to look at the apps and then execute actions to reward themselves or punish themselves if they were getting close to the end goal,” — Juan Marcano, Software Engineer, Uber

But what can AI agents do that automation cannot? As Marcano explained, traditional automation is effective for routine testing but struggles with surprise.

“Imagine you’re testing the action ‘request a trip,’” Marcano posited. “The system logs in and presses the right button, but then unexpectedly it [gets] a pop-up. Traditional automation breaks at that point. It sees something unexpected and just, boom, bye.”

This kind of variability is inherent to any global app that spans multiple markets, languages, devices, and sees constant feature updates. It’s exactly the kind of complexity that AI agents are well-suited to handle.

“The agents we’ve fine-tuned see those things and respond: they can cancel or sign up [when a promotional pop-up appears] and then keep going until they accomplish the goal,” Marcano explained.

An LLM interface was added to make the AI agent more accessible for internal testers, which “opened the floodgates” for easy onboarding and a surge in bug reports, so the team added further agents to identify and prioritise real bugs for fixing.

The results were dramatic. “Whenever we wanted to release an app globally — a new product, a new feature — it used to take about a quarter,” Marcano said. “Now we can release features weekly, going from roughly twelve weeks to just one.”

Testing & Trust 

For most enterprises, the shift toward agentic workflows is a work in progress. As AWS’s Varadan explained, many are currently in the “copilot” phase.

Varadan recommends starting with something simple: “I tell customers that the first thing that you can do to adopt AI is to see if you can put a natural language interface in front of your application or system,” he said.

In this way, enterprises get the benefit of AI-assisted interactions in everyday language on top of existing business logic and legacy systems.

“A lot of customers also have agents that work in the background,” Varadan added.

The biggest obstacle to enterprises moving into fully agentic workflows, Varadan said, is trust.

“Until agentic AI came about, people were only using AI for querying. Now it’s about changing the data in the system. With this, trust is obviously a big thing,” — Govind Varadan, Manager of Solutions Architecture, AWS

Varadan expects progression, however. “It is all about time and experimenting, seeing how the responses are and getting feedback from their customers, and then determining if they want to go to the next phase or not,” he said.

Governance is key, Varadan explained. This involves setting clear limits on what an agent can do and controlling its access to systems.

AI as Throughline

AI agents have powerful transformative potential for localization and global content creation, according to Phrase’s Altinay. 

According to the VP of AI Solutions, the next move beyond AI-enabled orchestration will be agentive workflows that produce multilingual, multimodal content aligned with business goals.

“Can I understand the intent? Can I understand the audience? Can I understand where this content is going to live? And then have the AI agent deliver the right output from that understanding? I want the agent to handle all of that seamlessly.” — Semih Altinay, VP AI Solutions, Phrase

Given a goal like “launch a new feature in Brazil next week,” an AI agent could plan actions, route and adapt content, align tone with the audience, track progress, measure impact, and self-correct as new data emerges.

A key implication for localization teams is that multilingual content creation must be clearly tied to a company’s top-level business objectives.

“It’s not just translating content for X purpose but helping the entire company achieve its goal at the executive level, whether that’s increasing product adoption, improving customer service, or something else.” — Semih Altinay, VP AI Solutions, Phrase

“This is one way for localization departments to elevate themselves,” Altinay went on, “by getting into those conversations.”

Organizational structures will transform, too. Business goals cut across departments, so an agent’s logic will naturally span silos and coordinate actions across traditional boundaries.

“Agents connect the silos and data that weren’t connected before,” Altinay explained. “We can link the marketing, product, and localization functions so we truly understand what we’re working with: what type of content, which language pairs, where it’s going to live. Then we let AI figure out exactly how to achieve that goal.”

AI agents thus become a connecting layer between functions — linking tools, data, and people — to produce outcome-aligned multilingual content. 

Altinay identified this as Phrase’s current direction: to be a language technology platform (LTP) that can serve as the AI backbone of the global content ecosystem. 

The human layer is not redundant, however. Experts will still create and adapt content and make qualitative judgments, while AI scales productivity. 

“So we’re now talking not just about hyper-automation, but hyper-personalization, reaching and resonating with audiences much more effectively,” Altinay concluded.