Prosus Predicts: AI trends for 2026

Euro Beinat, Global Head of AI at Prosus - What we got right, what we missed, and what's coming next in 2026

As an AI-first company, we’ve embedded AI solutions throughout our global portfolio of lifestyle ecommerce companies, from food delivery and online classifieds to fintech and experiences. We’ve built a team of 750+ AI specialists, investing $100M+ per year in AI talent, models and infrastructure – giving us a unique vantage point on emerging AI technologies and the best performing models.

It’s against this backdrop that we make annual predictions about where AI is headed next. Last year, in November 2024, we presented our five predictions to 3,000+ AI experts at our AI Marketplace conference. Having recently wrapped another AI Marketplace in November 2025 (this time with 17,000+ registrations), we can reveal which came true – and which were pretty far off – and share our new predictions for 2026.

Twelve months ago, we made five bold calls. Let's see how they held up:

1. 50% of new online content will be AI-generated (from 14% in 2024).

Outcome: We were on the money! The prediction that 50% of online content would be AI-generated in 2025 was surpassed, with 52% of content now created by machines. This surge can be attributed to the widespread adoption of advanced AI tools that not only excel in mimicking human-like writing but have also expanded into generating hyper-realistic images and videos. As AI-generated content has become easier to produce and more widely accepted by audiences, machines have overtaken humans in content production.

1. AI agents will carry out upwards of 10% of the customer interactions in e-commerce (from 0%).

Outcome: It depends. Agents currently complete ~0% of end-to-end transactions, but 50% of consumers are now using AI shopping assistants to search, compare and decide.
AI agents are becoming integral in e-commerce, streamlining buyer-seller interactions. However, the technology isn’t yet ready for agents to complete the full end-to-end transaction, where a consumer will state what they want to buy and the agent will find, purchase and have the product delivered. There are even more fundamental hurdles to be solved, for instance how to solve agentic payment in a secure and trustworthy manner.

1. There will be no open-source large-language models in the top 10 of LLM leaderboards (from 3 in 2024).

Outcome: We missed the mark, with three open-source models in the ProLLM top 10 in 2025. In 2024, proprietary models from major Western players had grown increasingly powerful and looked set to remain dominant. However, in 2025, high performance open-source models launched by groups like DeepSeek and other Chinese labs entered leaderboards and reshaped these rankings. Since DeepSeek’s breakthrough, we’ve seen a wave of high-quality open-source models coming out of China – a stark contrast to the US where companies have scaled back on open-source models. The surge in token consumption due to the rise of sophisticated AI agents has meant that open-source models are now also a more cost-effective option than proprietary models. 

1. A GPT-4-level large language model will be running on mobile devices.

Outcome: We missed the mark. Smaller models are running on mobiles, but not of a GPT-4 class.

While the trend toward on-device AI is clear – users want models running locally for privacy and speed rather than sending data to external servers – we underestimated how difficult this would be to achieve. The main technique for fitting large models onto phones is "quantisation," which reduces the precision of calculations to save memory and processing power. While it does enable smaller models to run on devices, the tradeoff is significant: these models work, but they're nowhere near GPT-4's quality. Reaching that level will likely take another year or two and will require not just better software optimisation, but new mobile chips specifically designed for AI workloads. The trajectory was right, but our timeline was too optimistic.

1. A $10B company will be created, based on a biological foundational model such as AlphaFold.

Outcome: We missed the mark. There are multiple $1-3B companies, but no $10B yet.
Our prediction was rooted in the potential of AI as a tool for science – where AI doesn't just analyse data but creates hypotheses, then tests and develops them. Following breakthroughs like AlphaFold in late 2023 and early 2024, there was widespread expectation that AI-powered biotech and material science companies would emerge as the next wave of billion-dollar startups. Several factors contributed to this: the technical challenges proved harder than anticipated, and investor attention shifted toward agentic AI applications rather than foundational scientific research.

Here are the five trends we're betting on for 2026:

1. 20% of AI agents for knowledge workers will complete multi-step workflows (30+ minute tasks) without human intervention (up from <5% in 2025).

At Prosus, we’re building the world's largest agentic workforce, encouraging everyone to identify workflows that they can automate by building a specialised AI agent. With 30,000 already in use, we’ve created the equivalent of 1,000 full-time employees across the group, unlocking major organisational capacity. A small but growing number of agents (around 250) have been used more than twice per day for several months, showing sustained value. Consider a senior agent deployed at iFood in Brazil. Each Monday, this restaurant account executive agent automates a critical task for account managers by extracting restaurant performance data from a database, compiling it into actionable reports, and helping account managers prepare for meetings with our restaurant partners. What once required manual effort equivalent to 40 full-time employees (FTEs) is now handled seamlessly and is used daily by 200 of our associates.

The real impact comes when agents can handle complex, multi-step tasks that take them 30 minutes or more: iterating through different loops, making mistakes, correcting course, testing alternative patterns, and persisting until they find a solution. Currently, agents tend to collapse after a few minutes or around 600 iterations. So when an agent successfully completes a 30-minute task end-to-end, it’s a significant breakthrough. We're not consistently there yet, but we are confident that we’ll get there in 2026.

1. A major retailer will report that 10% of sales come from fully agentic checkout, where consumers authorise the agent to make purchases on their behalf using their saved payment methods.

The prediction envisions a scenario where a consumer can tell an AI agent they want a pair of Adidas sneakers in a specific size, and the agent handles everything, including the payment authorisation that they’ve pre-approved. While last year’s prediction that AI agents would handle 10% of ecommerce transactions didn’t come true, we see ecommerce as being on the cusp of a fundamental transformation. A shopping transaction requires multiple steps: product search, adding items to a cart, confirming delivery details, processing payment, and handling tracking and support. In 2025, AI agents could handle search reasonably well, but little else. Now, all the necessary building blocks are emerging but haven't been fully connected yet. Google recently announced a Universal Commerce Protocol to streamline agent-based shopping across multiple retailers, Stripe has announced payment protocols for AI agents, ChatGPT has integrated with Shopify to enable cart experiences, and various protocols like MCP (Model Context Protocol) and agent-to-agent (A2A) and agent-to-payment (A2P) frameworks are being developed. While no single protocol is sufficient yet, the path is visible and investment is flowing into the space.

1. More than 40% of models in the top 10 leaderboards will be Chinese.

DeepSeek's breakthrough moment in early 2025 demonstrated that Chinese labs could produce frontier-quality models at a fraction of the cost, challenging assumptions about the dominance of Western AI companies. This wasn't an isolated achievement—it signaled the maturation of China's AI ecosystem, which benefits from massive scale, significant investment, and a large pool of AI talent. As Chinese labs continue competing at the highest levels of model performance while innovating on efficiency and cost, we expect this momentum to accelerate, fundamentally reshaping which models dominate the global leaderboards. With 800+ models in production across the Prosus portfolio, it's important for us to look at where the best models are coming from in the world.

1. A major player will launch a new fully AI-powered consumer device that achieves mass adoption, fundamentally changing how we consume and interact with AI.

The way we interact with AI is evolving beyond screens and keyboards. While smartphones remain our primary AI interface, we're seeing early signals of a shift – from smart glasses with integrated AI assistants to wearable devices that offer hands-free interaction. Meta's Ray-Ban smart glasses and a rival pair from Samsung represent initial attempts to reimagine how we access AI throughout our day, and we expect these to become mainstream in 2026. As AI models become more capable of understanding context and anticipating needs, and as hardware catches up to deliver seamless experiences, the conditions are aligning for further breakthroughs.

1. At least three major AI systems will be delayed or unavailable in Europe, laying the blame on the AI Act, whether an entire model, a suite of agents, or a critical business feature that remains accessible elsewhere in the world.

The EU AI Act enters its most consequential implementation phase in August 2026, when requirements on “high-risk” AI systems take effect. The regulations will impact fintech, education, food delivery, and numerous other services, creating significant compliance challenges for AI companies. There's precedent for caution: both OpenAI and Anthropic have already delayed certain releases in Europe even in the context of earlier AI Act implementation phases, though they eventually launched. Proposed amendments to delay the implementation of the EU AI Act and simplify elements of it could potentially soften these requirements, but if the current framework remains unchanged, it's reasonable to expect some companies will choose to pause or withhold features for Europe rather than navigate the regulatory uncertainty.