Lior Pozin had an epiphany about AI infrastructure in early 2025. As CEO of AutoDS, an AI-powered e-commerce automation platform, he had pushed his team to deploy AI features quickly, betting that speed would define success.
AutoDS was bootstrapped and eventually reached 1.8 million users, generated more than $1 billion in user revenue, and exited successfully to Fiverr. From its earliest days, the company was fast moving, the kind of place where speed was strategic and rapid implementation felt like the natural way to operate. But as Pozin’s team moved from pilots to production, they learned that speed alone was not enough. AI only delivers results when the right data foundations and ownership structures are in place.
“Without the right governance, data organization, and access, AI can’t scale,” Pozin tells Fast Company. “Once we built that foundation, everything changed. AI stopped being a feature and became part of how we operate.”
That experience was not unique to AutoDS. In 2025, across several industries, companies quickly realized that deploying AI at scale required confronting uncomfortable truths about their infrastructure, their assumptions about what AI could do, and their willingness to solve unglamorous problems before chasing transformative ones. While the year began with big promises, it turned out to be less about breakthroughs and more about a reckoning with reality.
The lessons that emerged reveal an industry growing up. Instead of building ever more powerful models or simply raising more capital, the industry is maturing by figuring out what actually works when the demos end and the real work begins.
INFRASTRUCTURE FIRST, OR NOTHING ELSE MATTERS
In early 2024, database company RavenDB explored building an AI assistant for its documentation in collaboration with Microsoft. The project ultimately fell apart. According to founder Oren Eini, the problem was not the AI model itself but everything surrounding it.
Data had to move through multiple systems before reaching the model, and updates required manual intervention. The entire setup depended on fragile connections that could break at any moment. For a database company, the irony was hard to miss.
The experience clarified something essential for the team: AI needed to be integrated far more deeply into the database itself to be reliable, predictable, and scalable.
For Eini, it wasn’t a setback so much as a signal that the surrounding architecture mattered as much as the model itself. That realization informed RavenDB’s more recent work on AI agents in and capabilities built directly into the database layer, where models operate closer to the data they rely on and can behave more predictably in production environments.
At AutoDS, that shift translated into a more deliberate approach. The team focused on building a shared data layer into its drop-shipping platform and clearer ownership around AI initiatives, which later enabled products like its AI-powered store builder to scale more reliably across the business.
The shift required patience. Pozin’s team stopped chasing what looked impressive and started tracking what mattered: time saved, accuracy improved, and decisions accelerated. “Success now means AI actually improves how we work, not just that we’re using it,” Pozin notes.
EFFICIENCY BEATS RAW POWER
While much of the AI industry chased larger models and more compute in 2025, Oculeus, a software-for-telecom company with deep experience in AI, spent the year prioritizing efficiency. The team focused on designing and refining systems that deliver reliable performance without excessive computational overhead. That focus is central to how Oculeus applies AI in telecommunications, where its systems are used to detect fraud patterns and anomalous behavior in real time.
In those environments, Arnd Baranowski, the company’s CEO, explains that “predictability matters more than novelty, because false positives and inconsistent outputs carry direct financial and operational risk.”
“AI algorithms and technology, which go along with massive computation and energy consumption, are a misguided path,” Baranowski adds. His critique extends beyond hardware, questioning the industry’s embrace of nondeterministic systems that produce different outputs for the same input. “Training must result in 100% deterministic responses. Otherwise, something is wrong.”
That stance runs counter to the excitement around large language models, which treat randomness as a feature. For Baranowski, the lesson of 2025 was simple: AI systems only earn trust when they behave consistently and can be relied on in real operating conditions.
Eini also shares that view. At RavenDB, the goal wasn’t building the smartest AI. It was building predictable AI that could handle routine tasks without drama. “We don’t necessarily want ‘smart’ AI,” Eini says. “We want predictable AI.”
As compute costs remain high and energy consumption becomes a public concern, 2026 will favor companies that figured out how to do more with less over those still chasing the biggest possible models.
TRUST DEMANDS BOUNDARIES
In 2024, Air Canada’s chatbot promised a customer a bereavement fare discount that didn’t exist. The airline was held liable. The case crystallized a problem that became unavoidable in 2025: AI agents can’t be trusted the way employees can.
Eini frames it bluntly. A bank teller is bound by policies and consequences. An AI agent isn’t. “I like to think about them as employees who I know are susceptible to bribes,” he says. “It’s crucial to consciously set boundaries for their actions and actively implement protective measures.”
Those boundaries took practical form. At AutoDS, Pozin created a dedicated team to verify AI outputs and ensure the system received accurate source data. At RavenDB, the team developed and implemented chain-of-approval processes and clear limits on what AI agents could access or promise.
The lesson extends beyond technical safeguards. AI agents exist in a gray zone between tool and actor. They respond to instructions but lack judgment. They execute tasks, but can’t weigh the consequences. That reality requires new frameworks for accountability that don’t assume good training guarantees good behavior.
Organizations thriving in 2026 will treat AI deployment as a trust problem first. That means transparency about capabilities and limits, clear expectations for users, and systems designed to fail safely when things go wrong.
SMALL FIXES BEAT MOONSHOTS
The year’s biggest AI narratives centered on autonomous vehicles, artificial general intelligence (AGI)—which AI scientist Yann LeCun thinks is an illusion—and models replacing entire professions. But companies making actual progress focused elsewhere: solving small, annoying problems at scale.
“The biggest changes will come from fixing many small problems, not from one big, all-knowing AI,” Eini says. “Quantity has a quality of its own, and removing many small frictions leads to a much faster pace overall.”
RavenDB empowered regular team members to build AI features in days rather than waiting for top engineers to approve and execute. AutoDS measured success by whether AI made employees faster and more efficient, not by how many AI projects were running. The results were individually modest but collectively transformative.
A year earlier, companies chased AI for its own sake, deploying pilots that looked impressive in demos but never scaled. In 2025, the focus shifted to measurable impact. Eini compares it to how we today make water potable for drinking, a practice so ordinary now that no one thinks about it. “In the same sense that ATMs or self-checkout services haven’t fundamentally changed the entire world, but have made our lives measurably better, I think we’ll see a lot of that,” he tells me. “The sheer quantity of changes will have a transformative effect.”
PREPARATION MATTERS MORE THAN REACTION
Steve Brierley wasn’t building AI in 2025. As CEO of quantum computing company Riverlane, he was watching how unprepared industries were when ChatGPT arrived. “The AI boom exposed how unready many industries were when tools like ChatGPT suddenly entered the mainstream, forcing companies to scramble around regulation, scalability, data readiness, and consolidation, and a widening workforce and skills gap,” Brierley says.
His takeaway: understand emerging technologies early enough to anticipate challenges rather than react to crises. Quantum computing will arrive sooner than many expect, and it won’t be a marginal improvement. “AI excels at analyzing and generating insights from data, while quantum computing will enable the creation of new kinds of data altogether,” Brierley says. “Together, they will unlock far greater exploration, discovery, and innovation than technology could achieve on its own.”
Gilles Thonet, deputy secretary-general at the International Electrotechnical Commission, saw the same dynamic in regulation. As AI laws took effect in 2025, companies struggled to translate legal requirements into operational reality. “International standards are essential to fostering trust in this transformative technology,” Thonet says.
WHAT COMES NEXT
The lessons from 2025 point toward an AI future grounded in operational reality rather than hype. Companies leading that shift built infrastructure, set boundaries, and solved real problems instead of chasing headlines.
But new challenges are emerging. Sheetal Mehta, global head of cybersecurity services at NTT Data, warns that AI capabilities driving productivity gains are being weaponized. “Agentic AI’s speed and ability to learn and make decisions autonomously can also be used by cybercriminals, exposing enterprises to new attack surfaces and unexpected security vulnerabilities,” Mehta says.
That means 2026 will require better safeguards, not just better systems. Organizations will need to treat AI security, governance, and ethics as foundational, not optional.
Pozin captures that shift rather poignantly. “The next phase of AI is AI that lives with us, learns us daily, and delivers exactly what we need, just in time. It won’t feel like a tool anymore. It’ll feel like a teammate that truly gets you,” he says.
Eini puts it even more simply: “Moving beyond the initial awe to become a transparent tool that simply gets things done.”
Not AGI. Not full automation. Just AI that works reliably, scales predictably, and solves problems without creating new ones. For an industry that spent years chasing moonshots, that might be the most ambitious goal of all.
