(Artificial Intelligencer is published every Wednesday. Think your friend or colleague should know about us? Forward this newsletter to them. They can also subscribe here.)
As we head into the holiday break, this will be the last Artificial Intelligencer of the year. We’ll be back in the first week of January, and until then, I hope you get some well-earned rest, good food, and a little distance from model releases and AI trades. Thank you for reading along since our launch this summer.
The Reuters tech team gathered this week for our year-end brainstorm — a chance to step back and look at what surprised us most in 2025 and what we’ll watch closely in 2026. This year was dominated by incremental model improvements , a continued investment frenzy in data centers, and the breakout winners with AI-generated revenue. Looking ahead, you can expect us to stay focused on the intersection of Wall Street bets and research labs looking for new directions, as well as the real-world impact of these new AI products.
One of the themes we continue to press on is the AI bubble and enterprise adoption. Everyone loves talking about pilots, but far fewer companies are willing to talk about what didn’t work. We’ve been asking exactly that. While it’s hard to get executives on the record about AI failures, we managed to surface a few telling anecdotes. Scroll on.
As a reader of our newsletter, you’re only seeing part of the story. Reuters subscribers unlock unlimited access to global coverage, including our biggest scoops and deep-dive reports. Become a subscriber to get the complete picture. Email me to share any thoughts.
Our latest reporting in tech & AI:
Why AI didn’t work at work (yet)
AI companies ultimately need other companies to pay for using their technology. Sam Altman has declared enterprise a huge focus for OpenAI and a $100 billion opportunity, and Anthropic’s $180 billion valuation is built almost entirely on business customers. Public software companies, from Salesforce to ServiceNow, are racing to reinvent themselves as “AI agent” platforms. Walk around San Francisco and you’re likely to see more than one billboard promising that AI will transform how work gets done.
Using AI at work is the key to unlocking productivity from this once-in-a-generation technological shift. But beyond general tools like ChatGPT Enterprise or Microsoft Copilot, many companies are still trying to figure out how this actually helps their top or bottom lines.
Over the past few months, my colleagues Deepa, Supantha and I spoke with companies ranging from Verizon, which employs nearly 100,000 people, to a niche wine-recommendation app, to understand where AI has fallen short. What we’ve learned is that it takes more than selling access to a state-of-the-art model and hoping it works like magic.
At CellarTracker, the wine-collection platform, the team built an AI sommelier designed to give honest recommendations. The problem? The chatbot was too polite. “It kept saying nice things instead of telling users they probably wouldn’t like a wine,” CEO Eric Levine told us . It took six weeks of tuning prompts and guardrails before the model learned it was allowed to say no.
That tendency — known to AI researchers as sycophancy — shows up everywhere. Models are optimized to be agreeable, which keeps users engaged but often produces worse advice. In safety-critical settings, the consequences are bigger. At Canadian rail operator Cando, an internal chatbot struggled to summarize a 100-page safety rulebook. Sometimes it forgot rules. Other times, it invented them. After spending about $300,000, the company paused the project. “We thought it would be the easy button,” one executive told us. “It wasn’t.”
Customer service was supposed to be AI’s fastest win, but even here companies are walking back their boldest claims. Fintech app Klarna once famously said its chatbot could replace hundreds of agents; a year later, the company acknowledged many customers still want humans for complex issues. Verizon reached a similar conclusion. AI now handles triage and routine tasks, but empathy — and trust — remain hard to automate.
Researchers describe this mismatch as the “jagged frontier” of AI: models can ace math Olympiad and still fail at mundane tasks like understanding dates, geography, or context buried in long documents.
That’s why enterprise AI is increasingly about hand-holding. We're seeing a surging demand for specialized engineering teams that work directly with customers to adapt the technology for their needs. Anthropic hires applied AI specialists who embed with companies. Startups like Writer put engineers on calls with customers to rebuild workflows together.
AI works when treated less like magic and more like infrastructure. It needs domain expertise, specific datasets and humans in the loop. The technology is advancing fast — but turning it into profit remains a very human problem.
Chart of the week:
Usage of OpenAI’s AI video generator and social app Sora surged quickly after launch, then cooled just as fast. According to Similarweb estimates, Sora’s daily active users across iOS and Android peaked at over 1 million in early November, before settling at roughly 750,000 DAUs in recent weeks.
The chart shows a familiar pattern for breakout AI consumer products: a sharp spike driven by novelty and curiosity, followed by a normalization phase as users figure out whether the tool fits into their daily habits. Turning AI video into something people open every day requires more than impressive demos; it needs social loops, clear use cases, and creative identity. The early data suggests Sora is finding a smaller but steady core audience, rather than sustaining mass-market momentum out of the gate.
