A look back at the predictions for 2025 and eight new ones for 2026. Let's see what came true last year and what lies ahead.
How did the predictions for 2025 turn out?
Several things were confirmed in 2025. Reasoning models and compound artificial intelligence (AI) systems truly dominated. For example, OpenAI showed with its o1 and o3 models that giving a model more time to "think" during testing unlocks capabilities that larger base models cannot achieve. The focus shifted from standalone models to systems in which multiple models, tools, and verification steps work together simultaneously.
AI also changed the economics of software. The "services as software" thesis—where AI delivers outcomes instead of selling licenses—became the standard for B2B investment. Startups are focusing on the services market, not just software.
OpenAI's dominance weakened. Models such as Gemini, Claude, and open models captured a share of the market. Enterprise customers are opting for a multi-model strategy. OpenAI remains the leader, but it is no longer the automatic choice.
NVIDIA faced challenges. Cerebras gained traction, AMD delivered at scale, and Anthropic bet on TPUs. The chip battle expanded to multiple fronts.
Robotaxis gained public trust. Waymo's white Jaguars are everywhere in San Francisco, and the company plans to quadruple its service area in 2026.
Some predictions will come true later. AI startups have not yet toppled the major players, but the pressure is growing—Google has returned in full force. AI interfaces are moving beyond chat windows, but more slowly than expected. Traditional search with 10 links is weakening, AI Overviews have reached 2 billion monthly users, and clicks have declined. The real shift will come with agentic commerce, where people are not involved at all.
One prediction failed: Llama did not become the "Linux of AI." Open models are thriving, but the ecosystem fragmented instead of unifying around Meta.
Eight predictions for 2026
Enterprise AI will finally begin to work at full scale. For two years, it seemed that enterprise AI was just around the corner. The problem is not the models, but the fact that AI cannot see how work gets done within companies—scattered across tools, protected by permissions, full of exceptions and human judgment that is not documented anywhere. In 2025, the focus was on agents that actually perform work from start to finish. Reaching 80% is easy for pilot agents, but 99% requires enormous effort. In 2026, startups will catch up with their ambitions, and companies will move from testing to real-world deployment. Major players such as Salesforce, ServiceNow, and Microsoft are supporting this with their offerings. Startups will work with companies on environments where AI observes and improves processes. Small, customized models running on internal infrastructure will outperform large ones because they are faster, cheaper, and keep data inside the organization.
Decision traces will become the new data advantage. When an agent carries out a process, it gathers context, applies rules, and resolves conflicts. Most systems discard this, but if you store it—inputs, policies, exceptions—you gain a structured history of how context turns into action. This is a context graph: a record of decision traces connected across entities and time. The more processes you handle, the better you become at automation. Startups have an advantage because they see the full context during decision-making, unlike legacy systems.
AI security will take center stage. Agents hold sensitive information about how companies operate. Threats span models, agents, and systems. Standardization such as MCP (Model Context Protocol) creates risks if verification fails. In 2026, AI security will become a key metric for executives. The principles of least privilege, controls, and monitoring will slow some deployments but strengthen the industry. Another major incident is expected, similar to the Salesforce flaw that allowed attackers to extract data, or the Mixpanel breach linked to OpenAI.
Major SaaS players will fight back. Salesforce is rebranding around Agentforce, and ServiceNow is making a similar push. They see customer churn and want to control the data. In 2026, they will restrict API access, add rules, and integrate AI into their platforms. For startups, this creates dependency risk.
Agents will take over e-commerce. Consumers are getting used to AI, and commerce will feel the impact first. AI learns preferences directly from descriptions. Visa says 2025 is the last year people will shop on their own. Mastercard, PayPal, and Google have protocols for agents. In 2026, this will take off with small purchases. Brands will optimize for agents, and aggregators such as Expedia will feel the pressure. Amazon and Walmart will remain strong thanks to logistics.
Gemini will overtake ChatGPT. ChatGPT grew by only 6% to 810 million monthly users, while Gemini grew by 30%, and the Gemini 3 model forced OpenAI to respond. In 2026, Gemini and Grok will gain market share thanks to integration into Search, Chrome, Workspace, Android, and X.
AI labs will go public. Anthropic is planning an initial public offering (IPO), with revenue growing from $1 billion (approximately CZK 22.5 billion) to $7 billion (approximately CZK 157.5 billion) in 2025, and is targeting $26 billion (approximately CZK 585 billion) in 2026. OpenAI is heading toward $20 billion (approximately CZK 450 billion) in revenue, but losses will reach $115 billion (approximately CZK 2.587 trillion) by 2029. The IPOs will be enormous but turbulent.
Interfaces like Cursor will become standard. Chat windows are on their way out. Cursor sees code directly and edits it. This year, the approach will spread to law, finance, and marketing. AI will propose solutions directly within the work environment.
Scaling laws are multiplying: pre-training, optimization, and test-time compute are being combined. Verification is key—AI excels where results can be checked. Pricing is shifting toward outcomes rather than usage.
