In 2025, skepticism among business leaders about AI's longevity has significantly reduced compared to 2024, when many viewed it as another crypto-like wave.

The CEO of Rackspace, Gajen, dedicates time from 4 a.m. to 6 a.m. daily to stay current on AI developments.

Building AI products is fundamentally different from building traditional software due to the non-deterministic nature of both user inputs via natural language and the probabilistic outputs from LLMs.

In some companies, subject matter experts are reluctant to collaborate on AI projects because they fear their jobs are being replaced.

The cost of implementing AI products will become 'ridiculously cheap' in the next few years, making design, judgment, and taste more valuable skills than pure execution.

A paper from researchers at UC Berkeley and Databricks, including Matei Zaharia and Ion Stoica, found that approximately 75% of surveyed enterprises cited reliability as their biggest problem with AI.

Major model updates, such as the deprecation of one API version for another, can break existing AI systems and require a full recalibration due to the new model's different properties.

Security issues like prompt injection will become a huge problem once AI systems go mainstream, but currently most teams are too focused on building to worry about it.

Developing agentic AI systems requires managing a trade-off between granting the system more agency to make decisions and relinquishing human control.

The claim of 'one-click agents' that work out-of-the-box for enterprise use cases is pure marketing and should be met with skepticism.

The term 'evals' in the AI industry is suffering from 'semantic diffusion', where it is used to describe many different activities, leading to a loss of its original, specific meaning.

Approximately 80% of the work for AI engineers and product managers involves understanding customer workflows, behavior, and data, rather than building advanced models.

An Air Canada chatbot hallucinated a refund policy that was not part of the company's official playbook, and the company was legally required to honor it.

Multi-modal AI experiences will be a major trend in 2026, moving beyond text-only LLMs to incorporate richer, more human-like conversational signals.

The AI product development lifecycle is very different from traditional software, which has broken established handoffs and collaboration models between PMs, engineers, and data scientists.

Replacing a critical enterprise workflow with an AI system typically takes four to six months of work, even for companies with optimal data and infrastructure layers.

The key to success in the current AI market is not being the first to launch an agent, but building the right data and feedback flywheels to enable continuous improvement.