Topline · 27 Aug 2026 · From the week of 24 August
SPOTLIGHT: AI Isn't Just Faster Translation, It's a $40B Tug-of-War for Global Attention. | Bryan Murphy, CEO @ Smartling
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Bryan Murphy, CEO of Smartling, an AI translation company whose customers include Apple, IBM and Disney, discusses how the company moved from translation that was faster and cheaper to translation that is also better. The episode covers the size of the translation industry, the quality gap between machine and human translation, how Smartling split its offering into human-in-the-loop and fully automated AI translation, and the organizational changes it made after GPT became usable. The central argument is that a founder-style reset, structured R&D tied to customer outcomes, and proof-of-concept selling were needed to bring AI into a traditional business without sacrificing quality or brand voice.
For founders
- Bryan said that when OpenAI made GPT usable, he reset Smartling's operating plan that same week to hire AI talent and reorganize the tech team.
- He ran R&D on time-boxed bets with a stated confidence level, and dropped ideas that did not reach that level within the set weeks.
- He set up R&D as a separate production line alongside the engineering team so that the team's release process was not disrupted.
- He said top-down change did not work at Smartling, so he backed off, listened to the team, and adjusted while holding firm on the vision.
- He won skeptical customers by running proofs of concept at Smartling's own cost, including a $100 bet that the company could deliver what it promised.
For revenue leaders
- Bryan said Smartling's pillars for AI work are quality, speed, cost and ease of translation, and R&D proposals had to show a material impact on one of them to proceed.
- He said companies ration translation to control cost, which he argued hurts conversion, since 87% of people globally will not buy from a website that is not in their language.
- He said machine translation scored about 86 on the MQM quality scale versus about 98 for human translation, so content like medicine, landing pages and marketing emails needed human work.
- He said he expects a combination of fully automated and human-in-the-loop translation, depending on the type of content and the level of quality required.
- He said the fully automated AI translation part of Smartling's business is now growing over 50%.
What was said 19, most useful first
AI-assisted human translators went from about 2,000 words a day to 8,000 to 10,000 words a day. Listen
He described this as AI with a human in the loop, which makes Smartling's translators vastly more productive. He compared it to developers using AI to write code, where the output rarely goes straight to production. He said the productivity figure is still climbing.
“It's like 8 ,000, 10 ,000 words a day and climbing”
Fully automated AI translation scores around 96 or 97 and matches brand style guides at a fraction of the cost. Listen
He described this as purely automated AI translation that is on brand, matching terminology, tone of voice and localization. He placed it a click below human translation's 98 on the same scale. He said it delivers at a fraction of the cost and turnaround time.
“that is matching style guides, so terminology, tone of voice, localized, all these things at a fraction of the cost and turnaround time.”
Smartling organizes its AI work around four customer-outcome pillars: quality, speed, cost and ease of translation. Listen
He said these pillars are the North Star for everything, including the product roadmap, and that every AI initiative has to link to them. He said anything that cannot show a material impact on one of them does not get off the ground in the BRD process, a discipline he learned from a mentor at eBay who asked what the change would do for the customer and whether it was material.
“So in our particular case, it's quality, speed. cost and ease of translation.”
Smartling time-boxes R&D bets and requires each to reach a stated confidence level within a set number of weeks, or it is dropped. Listen
Bryan said each idea is framed around a customer problem, given a set number of weeks to prove out, and measured against a confidence level. He said they give a little latitude but then move on to the next idea on the list, so they keep moving forward. He said that once an idea reaches a high confidence level, giving 80% as an example, it moves into the product and engineering roadmap.
“and you're going to be able to measure it with a confidence level”
He wins skeptical customers with proofs of concept that Smartling pays for itself. Listen
He said customers did not believe the claims and thought they sounded too good to be true, so Smartling put its money where its mouth was. He described betting one customer $100 that Smartling could deliver what it said, running a proof of concept on its own dime, and delivering, and said that customer is now a multimillion-dollar account.
“I bet the customer $100 that we could deliver what we said we could deliver.”
Bryan put the translation industry at about $40 billion and said code generation, projected to reach $27 billion by 2027, is about half its size. Listen
He said he had no idea it was that large when he started looking at the space four years ago. He compared it to code generation, which AI companies are fighting over, and said the most competitive market in tech is about half the size of translation.
“I had no idea that it was a $40 billion industry”
Bryan cited data that 87% of people globally will not buy from a website that is not in their language. Listen
He used a German shoe site as an example, saying he would probably not buy because he does not read German. He said Smartling's B2C and B2B customers translate to drive conversion rate and GMV.
“87 % of people globally will not buy from a website that's not in their language.”
The traditional translation rack rate is about 20 cents a word, and the workflow is still largely manual. Listen
He described the usual process as extracting content into a spreadsheet, emailing it to a translator, and revising it back and forth. He said this has always been expensive, slow and resource-heavy, and that he has seen large enterprises still work this way.
“rack rate is like 20 cents a word, and it's fairly slow.”
He rationed translation at past companies to control cost, and that this hurt conversion in foreign markets. Listen
He said most companies ration what they translate because of cost and turnaround time, and that he did this himself throughout his career. He said this is to the detriment of the company, since many prospects will not buy in a language they cannot read.
“do it as infrequently as possible which is really to the detriment”
Listen to the episode Pipeline & demand generation Link to this
Large retail customers with millions of SKUs and constantly changing product content need fast, automated, high-quality translation. Listen
He said Smartling's big retail customers have millions of SKUs with tens of millions of product descriptions that change all the time, which he described as a product information management problem. He said fast fashion customers in particular need automation and very rapid turnaround at high quality.
“they have millions of SKUs with tens of millions of product descriptions”
Machine translation scored about 86 on the MQM quality scale, while human translation scored around 98. Listen
He described MQM as a 0 to 100 score combining qualitative and quantitative measures, and he acknowledged that the score should not be read as percent accuracy. He said at 86 machine translation worked for lots of content but not for medicine, a landing page or a marketing email. He added that human translation is never perfect because translators have different preferences.
“we could do machine translation that would give you a score of like 86”
The fully automated AI translation business is now growing over 50%. Listen
He called this the big unlock for Smartling, saying it is the part of the business that is now growing over 50%.
“that business now is growing over 50%”
Small foreign-language websites lose search ranking, and that translation can grow a company's digital footprint in new markets. Listen
He said a US company's sites in markets like Germany, France and Japan are tiny compared to its home site, and in his view Google reads that as weak engagement and ranks it down. He said larger localized sites with better engagement and conversion help a company compete in those countries.
“Doesn't have great engagement. So I'm going to knock it down in terms of SEO, right?”
Bryan expects the future of Smartling's business to be a combination of fully automated and human-in-the-loop translation. Listen
He said customers need high volume, which automation handles, and very high quality for certain types of assets. He said the change will be less about moving everything to AI and more about doing more with the right type of translation.
“I think it's going to be a combination”
Bryan reset Smartling's operating plan the week OpenAI made GPT usable, hiring AI talent and reorganizing the tech team. Listen
He said the leadership team had the same realization at once and was messaging each other about it. He said that week he pulled the plan and started a project to hire the right talent and reorganize the tech team and its process for rapid innovation tied to customer outcomes. He noted Smartling already had over a decade of neural machine translation, which is a form of AI.
“So that week, I pulled the LTN and immediately reset the operating plan”
Bryan set up R&D as a separate production line within the tech organization, rather than disrupting the engineering team. Listen
He said the engineering team is one of the best he has worked with, with 3,300 production releases a year, 99.99% uptime and 98% on-time delivery, so he did not want to disrupt it. He hired AI talent and structured R&D as a process that feeds into the product and engineering machine.
“create almost like a production line within our tech organization that started with R &D”
Going top-down too hard created friction at Smartling, so he backed off to listen to the team. Listen
He said the founder in him went top-down too hard at first in an organization that had invested heavily in doing things the old way. He said top-down change does not work and creates a lot of friction, so he adjusted by listening and communicating while holding firm on the vision, which he called a tough balance.
“And you can't go tops down. It doesn't really work. It creates a lot of agita.”
He expects AI to dwarf, or at least match, the commercialization of the internet as a technology shift. Listen
He described three big moments in his career and said the next one, AI, is going to dwarf the commercialization of the internet or at least be on par with it. He said this is why he went into founder mode when GPT became usable.
“the next I think is going to dwarf both of those or at least be on par with them”
Founder-led organizations can act faster and with more conviction than non-founder-led ones when a platform shift arrives. Listen
He said he went into founder mode, calling timeout and breaking glass to throw out the plan, which he said would seem crazy in a big company. He said his leadership team shared the conviction, and that the speed and conviction with which founder-led organizations act is something he has noticed is very different from other organizations.
“It's that's the one thing that I've noticed is very different in founder led organizations versus not”