Welcome to The Median, DataCamp’s newsletter for August 22, 2025.
In this edition: OpenAI rolls out a $4.60 ChatGPT Go plan in India, DeepSeek launches V3.1 with hybrid inference, Anthropic partners with the U.S. government on nuclear AI safeguards, Google expands AI Mode globally with new agentic features, and an MIT report finds AI pilots stalling as tech stocks slide on bubble fears.
Important note: Our newsletter will be taking a short two-week break. We’ll be back in your inbox on Friday, September 12.
This Week in 60 Seconds
MIT Report Shows Most AI Pilots Fail as Tech Stocks Sink
A new MIT State of AI in Business 2025 report finds that while nearly 90% of employees use AI tools at work, only about 5% of enterprise pilots make it into production, with most companies seeing no measurable impact due to weak integration. Early ROI is showing up in back-office automation and agency spend cuts, while agentic systems are framed as key to bridging the “learning gap.” The report’s findings rippled into markets this week: Nvidia fell 3.5%, Palantir nearly 10%, and the Nasdaq dropped 1.2%. We’ll take a closer look at the MIT study in the Deeper Look section.
OpenAI Rolls Out $4.60 ChatGPT Go Plan in India
OpenAI has introduced ChatGPT Go, its most affordable tier yet, priced at 399 Indian rupees per month. The plan offers ten times more messages, image generations, and file uploads than the free version, plus double the memory span for better context retention. It also supports local UPI payments, making subscriptions easier for Indian users. Initially launched in India, OpenAI may expand the plan to other regions based on user feedback.
DeepSeek Launches V3.1 With Hybrid Inference
DeepSeek has released V3.1, introducing a dual “Think/Non-Think” mode that lets users toggle between fast outputs and deeper reasoning. The update delivers significant gains in efficiency, stronger multi-step agent capabilities, and a 128K context window across both modes. Benchmarks show improvements over earlier versions, especially on SWE-bench Verified (66.0 vs. 44.6), a benchmark consisting of real-world software engineering issues from GitHub. Open-source weights are available on Hugging Face, while new API pricing—$0.56 per million tokens for input and $1.68 for output—will take effect on September 5.
Anthropic Partners With U.S. Government on Nuclear AI Safeguards
Anthropic has announced a first-of-its-kind public-private partnership with the U.S. Department of Energy’s National Nuclear Security Administration to mitigate nuclear proliferation risks in AI. Together, they developed a classifier that distinguishes between harmful and benign nuclear-related conversations with 96% accuracy in preliminary testing, already deployed on Claude traffic. Early results suggest it can catch concerning content while allowing legitimate discussions in education, medicine, and policy.
Google Expands AI Mode Globally With Agentic Features
Google has rolled out AI Mode in Search to more than 180 new countries and territories, building on earlier availability in the U.S., U.K., and India. The update introduces new agentic capabilities, letting users book restaurant reservations, with plans to add local service appointments and event tickets. AI Mode can now personalize results based on user preferences and past interactions, such as suggesting restaurants that match dietary or seating preferences. A new sharing feature also enables collaborative use cases, like planning trips or events directly inside Search.
New Course: Building AI Agents with Haystack
A Deeper Look at This Week’s News
The Five Myths of GenAI in the Enterprise
MIT’s new State of AI in Business 2025 report opens with a sobering finding: while almost every company is experimenting with AI, 95% say they’re not actually making or saving money from it.
Pilots are everywhere, but business transformation is rare, and only a tiny share of projects scale beyond the demo stage. To unpack why adoption is so high but results so thin, we’ll explore the study through five persistent myths about generative AI in the enterprise.
Myth 1: AI will replace most jobs in the next few years
The data doesn't support the idea that AI is about to wipe out large swaths of the workforce. The report shows that GenAI’s impact is selective, not sweeping.
Layoffs are appearing mainly in functions historically treated as non-core—customer support, administrative processing, and standardized development tasks—where companies reported 5–20% reductions.
In most industries, however, hiring patterns remain stable. Executives in healthcare, energy, and advanced industries reported no expectation of reducing hiring in the next five years, with some noting they couldn’t even predict when AI might alter their workforce needs.
By contrast, more than 80% of leaders in technology and media—sectors already disrupted by GenAI—expect to reduce hiring volumes within two years.
Myth 2: Generative AI is transforming business
Adoption is undeniably high—nearly 90% of employees use AI tools at work—but transformation is rare. Only about 5% of enterprises have integrated GenAI into workflows at scale.
Source: State of AI in Business 2025 Report
Furthermore, most industries show no significant structural change. The report shows that clear disruption is confined to Tech and Media, while sectors like healthcare, finance, and retail remain largely untouched.
The maximum score is 5. Source: State of AI in Business 2025 Report
Myth 3: Enterprises are slow to adopt new tech
The report argues that the stereotype of sluggish enterprises doesn’t hold up, pointing out that 90% of large firms have seriously explored purchasing AI solutions and often dedicate more staff and budget than smaller firms.
However, the same study shows that mid-market companies take an average of 90 days to scale a pilot, while enterprises stretch to nine months or longer. In other words, enterprises may want AI badly, but their size and complexity still make them slower at turning pilots into production. This myth may not be entirely a myth after all.
Myth 4: Model quality is the main bottleneck
Executives sometimes blame poor model quality when pilots fail, but the report makes clear this is not the primary issue. Failures stem mainly from poor workflow integration, limited customization, and brittle user experiences.
Barriers to core workflow integration. Source: State of AI in Business 2025 Report
Users happily draft emails with ChatGPT but hesitate to trust enterprise AI for mission-critical processes when the systems can’t remember context or adapt to specific needs.
Myth 5: The best enterprises are building their own tools
It’s tempting to think the most advanced enterprises succeed with GenAI because they’re building everything in-house. The report shows the opposite: organizations that try to build their own tools often stall, while those that buy or partner are about twice as effective at getting measurable results.
The playbook looks less like classic SaaS adoption—plug-and-play software with minimal setup—and more like BPO or consulting engagements, where tools are customized, integrated, and co-evolved with vendors.
Early ROI comes not from broad layoffs but from reducing external costs, like trimming agency spend or eliminating outsourcing contracts, and this happens faster when companies work with specialized providers rather than reinventing the wheel internally.
Takeaways for companies
The report makes clear that the gap between experimentation and real impact comes down to how enterprises structure adoption. The companies that make it across the divide:
Start narrow and specific: Choosing well-defined use cases (document automation, voice AI for calls, repetitive code tasks) rather than trying to “AI-enable everything.”
Think like BPO, not SaaS: Treating vendors less like off-the-shelf software providers and more like long-term partners who co-develop workflow solutions.
Empower frontline managers: The best pilots are driven by people closest to the work, not top-down mandates from a central innovation lab.
Demand learning capacity: Tools that can remember, adapt, and persist with context are far more likely to scale than static chatbots or one-off pilots.
In short, the winners are those who stop chasing demos and start embedding AI into the messy but critical workflows where money is actually spent and saved.
Takeaways for individuals
The clear takeaway for individuals is that AI literacy is no longer optional. Executives consistently emphasized that proficiency with AI tools is a competitive advantage, with some even saying recent graduates surpass seasoned professionals because they’re more fluent.
Practical steps for individuals:
Develop AI fluency: Learn to use GenAI in your day-to-day tasks, whether for drafting, analysis, or workflow automation.
Stay close to workflows: The most resilient employees are those who understand the context of how work gets done and can pair that with AI tools.
Build adaptability: Companies are experimenting with many tools; being the person who can pick up new systems quickly is itself a valuable skill.
The jobs least at risk are those where workers combine domain expertise with AI proficiency. In other words, the safest move for individuals is to become the colleague who not only knows the business but also knows how to make AI work for it.
Industry Use Cases
DINOv3 Helps WRI Monitor Forest Restoration
The World Resources Institute (WRI) is using Meta’s DINOv3 vision model to improve how it tracks global reforestation and agroforestry projects. Traditional satellite datasets could spot large trees being cut down but struggled to monitor new saplings. By training DINOv3 on satellite and drone imagery, WRI can now detect a growing tree as early as eight months after planting and verify project progress across 27 African countries. The model cuts canopy height measurement error from 4.1 to 1.2 meters and enables a single workflow across multiple satellites, reducing costs and speeding up analysis.
Travel Advisors Use AI to Boost Bookings Without Losing the Human Touch
Some travel agents use AI to plan trips more efficiently. Athena Livadas, who runs a luxury travel agency, reports that her business grew about 40% after adopting AI for route ideas, emails, and client outreach. Digital agency Fora has rolled out AI-powered tools for drafting proposals, building itineraries, and even a chatbot called Sidekick, which 25% of its advisors now use monthly. One advisor said AI cut her trip planning time in half and tripled her booking capacity. Still, agents stress that AI can’t replace their relationship-driven edge—like securing upgrades or personal touches at hotels—which remains central to their value.
Startup Builds AI Tool to Fight Health Insurance Denials
Counterforce Health (a startup) has developed an AI assistant to help patients and clinics appeal denied insurance claims. The tool scans policies, medical journals, and review commission data to draft customized appeal letters, cutting the hours patients typically spend navigating complex denials. Founder Neal Shah, who previously launched CareYaya, describes it as “AI versus AI,” countering the algorithms insurers already use to deny claims. Clinics like Wilmington Health are testing the service, and pharmaceutical companies have begun reaching out. For now, the platform is free for patients, with potential future monetization coming from charging clinics and building a database of insurance claim outcomes.
Tokens of Wisdom
The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted. We're processing some contracts faster, but that's all that has changed.
—Mid-market manufacturing COO (anonymous), interview for the State of AI in Business report