Build vs. buy in AI Translation: Choosing the path that drives ROI

Insights from Laszlo Varga, Head of Research at Nimdzi Insights and Bryan Murphy, CEO at Smartling

As AI reshapes translation, localization leaders face a pivotal question: is it better to build your own AI translation stack, or buy a purpose-built platform?

In Session 4 of Smartling’s AI Translation 101 series, Smartling CEO Bryan Murphy sat down with Nimdzi’s Laszlo Varga to unpack one of the industry’s most impactful decisions. Together, they shared data, hard-won lessons, and a realistic path for teams under executive pressure to use AI—without risking quality, brand, or budget.

📺 Redo att titta? Streama hela sessionen on demand.

🎧 Föredrar du att lyssna? Lyssna på podden här.

📘 Go deeper with our ebook: Navigating the shift: Why, when, and how to adopt AI translation

Why ChatGPT alone isn’t enough for enterprise translation

It’s tempting to ask: “Why can’t we just use ChatGPT to translate?”

Laszlo explained why that approach rarely works at scale:

“Even the most advanced large language models still produce twice the errors that machine translation does,” he noted, citing Smartling benchmarking.

ChatGPT also isn’t trained on your brand voice, style guide, or terminology and carries a high risk of hallucination, especially across multiple prompts or sessions. Most importantly, some of the most important—and most complex—parts  of translation lie in the workflows: integrating content systems, preserving HTML tags, handling context, and reusing linguistic assets. ChatGPT on its own offers none of that.

Key takeaway: ChatGPT is powerful for consumers, but without workflow integration and brand safeguards, it’s not fit for enterprise-scale translation.

How top performers actually use AI translation

Bryan described how leading companies are structuring AI translation today:

• Automating workflows
• Creating custom-trained engines that use translation memory, glossary, and style guide
• Adding quality estimation, post-editing, and hallucination detection
• Routing content to a human when needed

When structured properly, these approaches deliver dramatic outcomes:

Bryan shared that Smartling customers have nearly tripled the content they’ve translated, while reducing cost per word by about 60% and reducing turnaround time by over 50%—making Smartling 4x faster than the average LSP.

Key takeaway: Purpose-built AI translation platforms combine automation with content governance, unlocking scale, cost savings, and quality gains.

The hidden costs of building your own solution

Building an AI translation solution internally can sound appealing. But as Bryan cautioned:

“You’re going to need full-stack engineering. You’re going to need integration engineers… data scientists… a computational linguist… and QA resources to validate and train what you just did.”

It’s relatively easy to prototype a translation demo. But the hard part (and the expensive part) is molding it into a scalable, compliant, and brand-safe enterprise solution.

Bryan added the opportunity-cost angle:

“Isn’t there something better that [your internal teams] could be doing that would actually drive revenue for your company?”

Key takeaway: DIY translation stacks often underestimate the true cost and resource burden. The combination of scaling, updating your solution to keep up with technological developments, and continued maintenance is often as expensive as the initial build.

The risks of skipping human-in-the-loop

Both speakers stressed that AI without review introduces serious risks:

• Brand damage from inconsistent or even offensive translations
• Legal and compliance exposure: GDPR, regulated content, and sensitive phrases in certain markets
• Operational failures: No audit trail, version control, or SLA management

As Laszlo warned:

“What are some of the risks of using, say, raw LLM output without a platform or post-editing? …It’s hallucinations. It’s bias. It’s mistranslations or omitted translations, or very simply put, no translations. Sometimes the large language model will not return anything—one in a thousand times perhaps. But if you’re doing it at scale, that’s a large enough number… And that’s the operational risk you’re running.”

And Bryan summed up the brand risk succinctly:

“No great brand was ever built on ‘good enough.’ That’s the whole point of this.”

Key takeaway: Purpose-built tools, quality evaluation, and post-editing remain essential for high-risk, high-visibility content.

Saying “yes” to the C-suite—the smart way

Both agreed that the right answer to executive pressure isn’t “no,” it’s “yes, and here’s how.”

Bryan advised framing AI translation the same way companies think about AI for software engineering: it accelerates output, but nothing ships to production without review. Back up your recommendations with metrics executives care about:

• Content volume: 2-8x more translated, depending on your workflows
• Cost savings: 60% lower cost per word
• Speed: 4x faster turnaround times

Key takeaway: Position AI as a controlled accelerator, not a shortcut.

Summan av kardemumman

Session 4 of AI Translation 101 made it clear: the build vs. buy decision is really about risk, resources, and results.

• ChatGPT alone isn’t ready for enterprise translation use.
• Working with a purpose-built AI platform delivers tangible gains and measurable ROI.
• Building your own is costly, complex, and rarely sustainable.
• Quality measurement and post-editing tools are non-negotiable for brand and compliance.

As Laszlo reminded the audience, “Clients don’t want to buy translation. Clients want to buy customers. That’s the bottom line.”

With the right approach, localization teams can say “yes” to AI while protecting brand trust, accelerating output, and proving ROI.

If you missed part of the conversation, replay the full webinar anytime, or catch the podcast on the go.

Want a practical roadmap for adopting AI translation? Download our ebook, Navigating the shift: Why, when, and how to adopt AI translation.