Self-Improving AI: What It Means for B2B SMEs

The timeline: from Salesforce AI to $650 million for an AI that builds itself

Richard Socher is no newcomer to artificial intelligence. He previously led Salesforce's AI division and founded you.com, a search engine built on language models. In May 2026, however, he made a significant leap in scale. His new startup raised $650 million with a stated goal: to build an AI system capable of researching and improving itself indefinitely.

According to TechCrunch, Socher stresses one crucial point. The project is not purely academic. The company therefore intends to release concrete commercial products, setting itself apart from the foundational research of OpenAI or DeepMind in their early stages.

This detail matters. The gap between frontier research and business applications is closing fast. As a result, B2B SMEs can no longer afford to watch these developments from the sidelines.

Who wins and who is at risk: a no-nonsense view

Self-improving AI systems — also known as recursive self-improvement systems — are a distinct category from today's generative models. Instead of being trained once and then deployed, these systems modify their own parameters or architecture in response to new data and goals.

The main beneficiaries of this technology include, first of all, large enterprises with mature data infrastructure. They can use self-optimizing systems to keep cutting operating costs on an ongoing basis. In addition, tech vendors that build these capabilities into their SaaS products will gain a structural competitive advantage.

By contrast, SMEs that rely on a single supplier risk ending up technologically dependent. If the system running their processes evolves on its own, operational control gradually shifts to the vendor. This is a real risk, not a hypothetical one.

Compliance is a risk too. Systems that modify themselves are hard to audit. In regulated sectors — such as finance, healthcare or logistics — they could therefore clash with regulations like the EU AI Act.

SHM Studio's take: between enthusiasm and pragmatism

At SHM Studio, we have been following the evolution of AI systems for SMEs for several years. Our view on this news is nuanced.

On one hand, Socher's announcement confirms a direction that was already visible. Today's generative models show increasingly sophisticated in-context learning capabilities. So the move toward systems capable of structural self-improvement is a logical progression, not a sudden quantum leap.

On the other hand, the distance between a $650 million announcement and a product that a manufacturing SME in Brescia or a B2B distributor in Turin can actually use remains considerable. So our practical advice is not to wait for this technology before starting your digital transformation. Quite the opposite.

Companies that build a solid data foundation, digitized processes and in-house AI skills today will be best placed to adopt self-improving systems once they're actually available on the market. That's why investing in AI solutions you can apply today doesn't conflict with preparing for the future — it's its prerequisite.

Three real-world scenarios for Italian B2B SMEs

It helps to translate this scenario into realistic use cases. Here are three contexts where self-improving systems could affect Italian B2B SMEs over the next 24-36 months.

• Adaptive sales automation: CRM systems that autonomously optimize outreach sequences based on response rates, with no human intervention. This continuously lowers the cost per qualified lead. LinkedIn campaigns could benefit from AI layers that calibrate themselves.
• Self-optimizing SEO and content: publishing platforms that rewrite content on their own based on ranking signals. As a result, the role of strategic copywriting shifts toward editorial oversight and goal setting.
• Paid campaign management: systems that reallocate budget across channels in real time, optimizing not just individual campaigns but the entire media architecture. This directly affects how Google Ads campaigns and channel mixes are managed.

In all three cases, the value isn't in automation itself. The real value lies in a company's ability to set clear goals, monitor results and course-correct. Human skills don't disappear — they move to a more strategic level.

Work still in progress: what's missing before mass adoption

Despite investor enthusiasm, there are significant technical and regulatory obstacles. First of all, self-improving systems require enormous computing infrastructure. At least at first, that makes them accessible only to large organizations or through APIs from specialized vendors.

Second, the question of interpretability remains unsolved. As MIT Technology Review has documented, today's models are already hard to audit. A system that modifies itself adds another layer of opacity. Compliance with the EU AI Act — which requires transparency and traceability — thus becomes a technical challenge even before a legal one.

Finally, there's the matter of organizational trust. Italian SMEs have historically adopted new technologies with caution. So even when these systems are technically available, the adoption cycle will be longer than in Anglo-Saxon markets. That's not necessarily a disadvantage: it lets you watch others' failures before you invest.