If 2024 was the year of AI experimentation, 2025 is shaping up to be the year when organizations are expected to make it real. But for many enterprise leaders, the question remains: how do you actually move from experimentation to something scalable, measurable, and safe?
That was the focus of the recent Session 4 of the Elevate Innovate series, hosted by Dave Ruane and co-piloted by Olga Blasco of Lion People Global on LinkedIn. Joining them were our own Georg Ell, CEO, and our Chief Product Officer Simone Bohnenberger-Rich, PhD.
Together, the group tackled one of the most urgent (and often misunderstood) challenges in enterprise AI: what it really takes to move from pilot projects to production-ready systems.
Over the course of 45 minutes, the conversation moved from failure rates to governance frameworks, innovation culture to agentic AI. And rather than offering sweeping predictions, the panel kept things refreshingly grounded, sharing lessons learned, missteps encountered, and the everyday realities of operationalizing AI at scale.
Watch the full session now:
Here are the key takeaways from the session.
1. Why most AI projects still don’t deliver
Rather than jumping to celebrate the potential of AI to transform the language industry, Host Dave Ruane began the session with a question that offered something of a reality check for those of us working with AI for enterprise: Why is it still so hard to make AI work at scale?
Simone Bohnenberger-Rich responded honestly to this, with data to help frame the reality of current AI development and deployment.
“If you consult Gartner or other surveys, 90% of AI projects don’t make it from a POC into production,” she said. “That’s a pretty sobering number.”
It’s not for lack of enthusiasm. As Simone explained, the C-suite has made its position clear: AI is a strategic priority. But in many organizations, that mandate has translated into a wave of disconnected pilots, what she described as “cities of robots.” Experiments are everywhere, but successful deployment is rare.
So what’s going wrong?
One of the biggest misconceptions, Simone argued, is the belief that LLMs are plug-and-play. “They look deceptively simple,” she said.
“They work in the lab. They feel intuitive. But they’re not solutions. They’re more like sous-chefs in your kitchen, telling you what you could cook with what you have in the fridge.”
That’s not enough when consistency, accuracy, and accountability are required at scale.
The deeper issue, she said, is data. Without a clean, structured, well-maintained internal data layer, even the most sophisticated AI won’t produce reliable results.
“Your proprietary data is your competitive edge,” she explained. “But it’s often messy, scattered, or simply not ready for production.”
And then there’s the question of risk. As Simone reminded the audience, LLMs are probabilistic by design. They might get something right one day—and completely wrong the next.
“You’re dealing with systems that aren’t deterministic…and that unpredictability can have serious consequences.”
- Simone Bohnenberger-Rich, PhD., Chief Product Officer, Phrase
She referenced a recent development from Lloyd’s of London: a new insurance policy created specifically to protect businesses from AI errors. That, she said, should be a wake-up call.
Building a reliable AI solution isn’t about tinkering with a few prompts or launching flashy demos, but about how you accomplish deeper integration that’s grounded in data governance, risk tolerance, and realistic expectations.
As Simone put it: “The challenge isn’t experimenting with AI. It’s engineering it into something that can stand up in the real world.”
2. From pilots to production: What most teams miss
Later in the conversation, Dave steered the panel with a key question: Once you’ve launched a pilot, how do you move to something that’s truly operational?
Again, Simone offered a word of caution. Too many teams, she said, assume the model will manage itself.
She explained that automation alone isn’t enough. Real-world deployment requires safeguards, especially when reputational risk is on the line.
As an example, Simone talked about how you might prevent an AI from accidentally naming a competitor in branded content. “That’s a mistake that you could easily catch with a simple rule.”
These are the kinds of implementation details that get overlooked when AI is treated like a lab experiment, rather than a business-critical tool.
Simone’s core argument was simply that AI isn’t magic. It’s software. And like any software, it needs testing, oversight, fallback systems, and governance. That means combining human-in-the-loop practices with layered safeguards, and designing end-to-end workflows where AI adds value, but doesn’t run unchecked.
3. Making innovation real: Culture, cadence, and the role of leadership
Shifting the conversation from technical AI challenges to something more foundational, our host asked the panel: how do you actually build a culture where innovation doesn’t just happen once, but becomes sustainable over time?
With both a CEO and a CPO in the hot seat, it was an ideal moment to explore not just strategy, but behavior. Dave framed it simply: “Let’s pivot into the how. How do you drive and build a culture within an organization?”
Phrase CEO Georg Ell responded with characteristic candor. “We’ve certainly been on a journey,” he said, reflecting on how Phrase evolved from selling individual tools to operating as a full platform. But he was quick to point out that the shift wasn’t just about technology, it was cultural.
“It starts at the top. I join every new hire orientation. We talk about experimentation, about making mistakes. That tone is set from day one.”
- Georg Ell, CEO, Phrase
But it was co-host Olga Blasco who pressed the conversation further. She acknowledged the buzz around innovation in tech, but challenged the panel to go deeper.
That led to a revealing exchange about Phrase’s own structured approach to launches, something Georg described as a “cadence process” inspired by rocket science. With set “T-minus” milestones and fixed release dates, the entire company moves in sync every quarter.
“People assume too much structure kills creativity, but we’ve actually found the opposite. The structure eliminates friction. It gives people the clarity to experiment.”
Yet the hosts kept the discussion grounded. Dave pointed out that plenty of companies try to systematize innovation… and fail. So what makes it stick?
Georg’s answer brought it back to ownership and iteration.
“We haven’t missed a launch in 11 quarters,” he said, “but we’ve learned from every one. Each cycle, we get better. And everyone’s involved—engineering, marketing, sales, customer success. It only works if it works for everyone.”
- Georg Ell, CEO, Phrase
This portion of the session didn’t deliver a silver bullet, but it did offer something useful: a glimpse into what happens when innovation is treated as a discipline, not a buzzword. Through the lens of both leadership and critique, the panel made it clear: culture doesn’t come from slogans. It comes from what people are empowered to do, week in, week out.
4. Productivity, not headcount: What AI really enables
During the conversation, host Dave Ruane brought up a familiar tension: is AI just a vehicle for efficiency, or can it become something more transformative? For Georg Ell, the answer was clear.
It’s a mindset that runs counter to the short-term thinking seen in some corners of the industry, where AI is often used to justify job cuts. But at Phrase, Georg emphasized, the goal isn’t leaner teams. It’s more capable ones.
AI, in his view, shouldn’t be reserved for technical roles or power users. It should be in the hands of everyone, from junior hires to senior leaders. That’s why Phrase equipped its entire workforce with access to tools like ChatGPT nearly two years ago, along with training and support.
“This is an innovation arms race. The companies that embrace AI at every level—from junior to senior—will win.”
Olga Blasco supported the idea, adding that real change happens when productivity becomes collective. Not just faster output, but smarter decisions, better orchestration, and more room to experiment without fear of failure.
5. Why it’s time to rethink the ‘TMS’
At one point, Dave posed what’s become a loaded question in the language tech industry: “What exactly is a TMS today?”—referring to the once-standard label for Translation Management Systems.
Georg didn’t mince words.
“TMS is a tiny subset of what enterprises actually use today. We don’t even sell a TMS anymore. We haven’t for years. What we offer is a platform.”
- Georg Ell, CEO, Phrase
That distinction matters. The term TMS, as Georg explained, is often too narrow to capture the full complexity of today’s enterprise localization workflows, especially when AI, APIs, orchestration layers, and real-time automation are involved.
Referencing Slator’s 2025 language industry market report, which formally repositioned vendors (including Phrase) under a new term: Language Technology Platforms (LTPs).
“It was a seminal moment,” Georg said. “It was the industry finally catching up to what’s actually being built.”
Olga echoed that the rebrand reflected the shift toward broader ecosystems that can plug directly into enterprise content strategies, from CMS integrations to AI-driven review pipelines.
She noted that future platforms will need to be subscription-based, flexible, and predictive, as well as fundamentally aligned with how modern content gets created, distributed, and localized.
6. The hype (and reality) of agentic AI
“Agentic AI” rapidly gaining buzz across tech circles. Dave invited Simone to unpack where the real value (and the real risk) lies.
She began with the vision: intelligent agents that can localize content, automate workflows, and make context-aware decisions in milliseconds.
“There’s a lot of potential in agentic AI. It can condense multiple workflow steps into one. It’s fast, personalized, and incredibly appealing.”
But she also cautioned against blind enthusiasm. Each action an agent takes adds risk. If one part of a multi-step process is only 80% accurate, chaining three together drops total accuracy to around 50%.
“That’s a coin toss,” she said bluntly. “It’s worthless.”
Simone explained that agentic systems don’t just amplify the capabilities of LLMs—they also amplify their unpredictability. If the agent pulls from an unvetted data source, or invokes a risky tool, things can go sideways quickly. She even referenced the potential for security vulnerabilities if a bad actor injects a rogue tool into the system.
Her advice was pragmatic:
Agentic AI clearly holds massive promise, but only for teams who are disciplined enough to constrain it.
7. Ecosystems over egos: A new approach to partnership
As the conversation shifted toward partnerships, Georg made it clear: Phrase doesn’t see success in terms of how many leads a partner sends their way.
That humility is strategic. It’s why Phrase has opted to remain a pure software company, even as others in the space drift toward offering managed services.
“I’ll resign as CEO before Phrase becomes a services company,” Georg said with emphasis. “We don’t want to compete with our partners. We want to enable them.”
Olga underscored how meaningful that clarity is in a market full of blurred lines. Customers don’t want every vendor to become an all-in-one provider. They want cohesive ecosystems that work seamlessly, and know their lane.
The conversation also touched on product. Phrase’s API-first approach, Georg noted, is about innovation as much as extensibility. Partners and customers are encouraged to build on the platform, adapt it to their needs, and co-create value.
“We want to be the Microsoft of localization, where for every dollar we earn, our ecosystem earns 17.”
That, he said, is how you build a network that scales and lasts.
8. Advice to the C-suite: leading through uncertainty
As the session drew to a close, Dave asked the panel to distill everything down to one thing: What should senior leaders do now to guide their organizations through this moment of transformation?
Georg didn’t overthink it:
“Get started. You’ll fail. That’s fine. There’s gold in them thar hills.”
The message was clear: don’t wait for perfect clarity or the perfect roadmap. Action beats analysis.
Olga brought a voice of experience, especially for those navigating change at scale:
“Test and learn. Fail fast. Improve. Repeat. And get comfortable with the pivots.”
Simone offered a strategic lens, cautioning leaders not to fall in love with the tech:
“Start with the business case, not the tool. If there’s no clear success metric, it’s a moonshot. And moonshots rarely pay off.”
Together, their advice formed a practical playbook: ground your AI strategy in reality, give your teams the tools and the freedom to explore, and focus on what matters most: value, not novelty.
Final Thoughts: Turning AI ambition into enterprise reality
What stood out in this session was honesty about AI’s potential, and limits.
Rather than dwelling on abstract visions of what AI might become, the panel focused on what it actually takes to make it work today. From Simone’s breakdown of why projects stall, to Georg’s insights into cadence, partnerships, and productivity, the message was consistent: success is about designing systems, teams, and cultures that can adapt.
Whether you’re part of a global enterprise trying to scale AI across content workflows, or a leader figuring out how to steer your organization through a shifting tech landscape, the real lesson is simple: start small, build thoughtfully, and don’t wait for perfect conditions.
The future of AI in localization—and enterprise more broadly—won’t be driven by hype. It’ll be shaped by the organizations that learn how to operationalize it, piece by piece.
