Julian Pintat, a freelance English-to-German translator has watched his 15-year career gradually unravel. Specializing in high-stakes fields like medical technology and pharmaceutics, his expertise has been repriced as an AI cleanup service. On a recent job, translating an operating manual for an oil rig, AI mistranslated “scale”—a mineral buildup—as both a musical scale and a device for measuring weight. Fixing such basic flaws, which now constitutes 95% of his work, often takes longer than translating from scratch, he says—a frustrating reality that has halved his income and put life plans including marriage and starting a family on indefinite hold. With Google Translate and later DeepL having burst onto the scene years before ChatGPT—professional translators have been feeling the effects of artificial intelligence longer than most. “I’m the canary in the coal mine,” Pintat says.
[time-brightcove not-tgx=”true”]
AI is changing the face of work, and Pintat is one of many looking at a very different future. While there has been much conversation about the technology replacing white-collar workers—with some CEOs like Ford’s Jim Farley and Amazon’s Andy Jassy predicting many corporate jobs will be wiped out by AI in the coming years—the first wave of AI adoption is already recasting workers in new roles and altering the contours of their jobs.
For some companies AI is enhancing efficiency. It has allowed London-based law firm A&O Shearman to effectively multiply its workforce, letting it take on projects it would have turned down, says partner and global head of the firm’s AI group, David Wakeling. To help a major U.S. bank comply with European law, the firm built a tool that scanned 20 years of license agreements and identified which needed amending. “Two years ago, we would have had 20 lawyers in a room, maybe some parallegals,” Wakeling says. But the tool whittled 2,400 regulatory requirements down to 900, halving the project cost even when accounting for the time to build the tool, he says. Still, he tempers his optimism. He says a basic off-the-shelf AI assistant probably won’t add much value, noting that real results often require customized or specialized solutions. “It takes a lot of elbow grease,” Wakeling says.
Meanwhile, AIG CEO Peter Zaffino told TIME in June that the insurance company is using AI to do underwriting work faster. It’s training a system to become a “junior underwriter” that can do the bulk of the underwriting, allowing the “more experienced practitioners” to do the rest. “Part of the cultural change is upskilling, retraining positions in a new world that enables them to be more productive than we were in the past,” he said.
An MIT report published in August concluded that 95% of AI pilots are failing to provide a return on investment. Even AI coding assistants, which have been held up as AI’s winning use case, have been called into doubt by a small preliminary July study published by Berkeley-based AI research group METR. The sample of 16 experienced developers were 19% slower when using AI despite estimating it made them 20% faster on average. Fueled by excitement—and perhaps fear of missing out—businesses are racing to use AI, even if in suboptimal ways. That creates a gap between what the market thinks AI can do and actual performance, squeezing both businesses and—in the case of translators—workers caught between the promise of superhuman efficiency and the reality of often flawed machine output.
AI’s rapid advance on white-collar tasks could soon erase that gap. Performance on tests devised by seasoned professionals across banking, law, and consulting has nearly doubled in little more than a year, according to a preliminary study published in late September by data firm, Mercor. The paper came on the heels of another report, authored by OpenAI, which sought to measure AI’s ability to do real-world tasks by comparing machine to human performance in blind tests. It found that the best models compare favorably to human authored work nearly half the time. Though, both reports note that such tests measure performance on well-scoped tasks—a quality lacking in the often messy real world. That means, for now, AI models might make a poor substitute for human workers, and implementation remains key.
Read more: AI Is Learning to Do the Jobs of Doctors, Lawyers, and Consultants
“Generative AI does a really fantastic demo,” says Kaitlin Elliott, head of Morgan Stanley’s Firmwide Generative AI Solutions, but making it useful is harder than it looks. The bank has built its own meeting transcription and summarization tool which Elliott says saves hours of grunt work. It’s also created an AI-driven search tool that makes it easier for staff to surface information. “In the early days, we thought that we could just give it all of our knowledge, and it would be able to give very accurate responses,” she says, but in practice, it took well structured data, and careful testing.
The challenge of implementing AI is creating demand for new kinds of expertise. “There’s still a need for technical skills,” Elliott says, who adds that while AI tools have automated work typically given to junior staff, younger generations are now being counted on to bring AI skills into organizations. “They’re all adopters of AI. They know how to effectively use it,” she says. That demand is also creating opportunities for companies selling that know-how as a service, like Scale AI, best known for its data labelling business but which has expanded into helping enterprises, including Morgan Stanley, with AI applications. When adopting AI, it’s crucial to work backwards from problems, says Felix Su, director of engineering for Scale’s enterprise AI arm. Applying AI for its own sake can backfire. Su gives the example of one of Scale’s clients, which had built four chatbots for slightly different tasks, forcing staff to constantly copy and paste between them. Su adds that identifying solutions means sometimes applying generative AI, but often it means using traditional machine learning or software engineering.
Improvements are coming to translation. DeepL offers features like creating a custom glossary and this year introduced a tool that asks follow up questions to clear-up ambiguities, which could help in niche domains. Translators that have successfully leveraged AI are working faster says DeepL CEO Jaroslaw Kutylowski, though the company only offers numbers comparing its tool to AI rivals, not unaided human professionals. Improved performance could allow translators to offset the lower cost-per-word by translating higher volumes of text, though Kutylowski notes that AI tools are allowing some businesses to bring translation in-house rather than outsourcing to professionals. “I think we’re just kind of upleveling here on this civilization ladder,” he says, acknowledging it will bring changes to how people do their jobs. “That is a change that we will have to go through,” he adds.
Elsewhere, companies are aiming to lower the technical bar for businesses to adopt AI by creating ready-made enterprise solutions. Most virtual meeting products now have an AI summarizer built in, for example. A&O Shearman arms its lawyers with AI legal tool Harvey, which Wakeling says is useful for general questions, though the firm now uses its own AI tool, ContractMatrix, for especially niche queries. And a new breed of enterprise-focused tools pull from internal documents, Slack messages, and emails to answer queries.
Canadian AI company Cohere, released North, its own such tool last month. The company’s co-founder, Nick Frosst, says he no longer sweats last minute meetings, because he uses the tool to prepare a brief on individuals based on their entire history with the company in seconds. North now tackles 90% of its general support tickets, though human operators are still in the loop. It isn’t just being used internally, with RBC, Canada’s largest bank, adopting the platform. (Salesforce, where TIME co-chair and owner Marc Benioff is CEO, is an investor in Cohere.)
While the bosses of OpenAI, Anthropic, and Google Deepmind all believe so-called artificial intelligence or AGI—a system that can automate most human work—could be just a few years away, Frosst’s outlook is comparatively conservative. He doesn’t believe we’ll reach AGI using anything resembling current technology. Still, he says even without AGI, the impact on labor will be disruptive, comparing it to the industrial revolution. “When there were massive transitions in the labor market, a lot of what was solved was at the government level, the union level,” he says. “This is a problem beyond any individual, and we need to address it as a collective.”