Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Michael Hunger, VP of Product Innovation at Neo4j
After the first practical years of AI, one central question is emerging for many companies: Where is the ROI? AI agents are not autonomous enough, model usage suffers from “context rot” or quality degradation, and traditional database concepts are reaching their limits. The graph experts at Neo4j highlight which trends in 2026 will significantly advance the practical use of AI.
1. AI Reality Check: Scaling in Focus
Feedback from business is clear: Most AI projects are still not delivering what was originally expected. According to an MIT study, 95% of pilot projects do not generate measurable outcomes. Gartner expects that 40% of agentic AI projects will fail by 2027 – held back by costs, unclear ROI, and unresolved risks. This reveals the GenAI paradox: General AI tools and assistants can be rolled out quickly, but the ROI is hard to measure. The truly value-adding, vertically integrated AI systems, on the other hand, are struggling to gain a foothold in companies.
However, it is premature to speak of AI frustration. The end-user hype surrounding ChatGPT, Copilot & Co. -and not least the enormous investment sums of the tech giants – has created expectations that collide with the reality inside organizations. Here, AI systems must be deeply and securely integrated into existing processes, data structures, and IT landscapes. This integration requires time and adaptation, but even artificial intelligence can accelerate that process only to a limited extent. Moreover, AI is experimental. Many prototypical projects have to fail in order to reveal which approaches will work in the long term. True scaling is only just beginning.
2. AI Agents: The New Trainees
AI agents illustrate this paradox particularly clearly. Autonomous “armies of agents” replacing entire departments are scarcely seen in practice. Most systems operate behind the scenes, taking on time-consuming research tasks in areas such as law, compliance, or medicine. According to McKinsey, while the majority of companies are experimenting with AI agents, only 23% bring them into any productive environment at all, and in no function does the share of scaled agents currently exceed roughly 10%. This suggests that their usefulness is highly context-dependent, and their practical application remains narrowly limited. Companies must first tune out the AI hype and soberly determine where agents actually have an impact.
This iterative approach is justified because AI remains unreliable. The issue is usually not a “lack of intelligence” in the models. Rather, context and instructions were not conveyed clearly enough to guarantee relevant and reliable results. For functional integration, AI agents therefore need – similar to new employees – a kind of onboarding: They must be trained, informed, monitored, and regularly corrected. Because they operate probabilistically, they do not necessarily deliver the same results even with identical inputs. Validation requires testing, feedback, and review processes – effort that scales only to a limited degree.
This raises another question beyond technical implementation: How can AI agents be integrated into existing workflows, teams, and organizational culture? Companies must invest not only in training and setting up their agents, but also in training their employees. They will be the ones to validate the results of their AI colleagues and understand the underlying limitations of the models. This requires rethinking work models, including clear governance, new roles, flatter structures, and, most importantly, unambiguous responsibilities.
3. Context Engineering: Information Architecture for AI
AI – even in agentic, iterative architectures – is only as good as the context it receives. Yet it often receives too little, too much, or overly imprecise input. Many people think of prompting primarily as giving direct instructions. In real applications, however, the system’s main task is to dynamically shape the context so that the LLM receives exactly the information it needs for its next step.
LLMs function in some ways like human memory: They retain the beginning and the end, but lose the thread in the middle, as shown in a study by Stanford University. Long context leads to errors, friction losses, and a decline in attention (context rot). Models become confused when too many or very similar tools are used (context confusion). Or they stumble over contradictory work steps (context clash). Although models could theoretically process huge amounts of context, practice shows that the more you load into the context window, the less reliable the results.
As with data processing in general, the same applies to AI context: quality over quantity. Models have only a limited attention span – a concept Anthropic refers to as the “attention budget.” Each additional piece of context consumes part of it and dilutes what matters. Well-curated context, or context engineering, thus becomes a fundamental requirement for reliable AI.
4. Push vs. Pull: Data on Demand Instead of Data in Advance
With the rise of agent systems, the way AI accesses information is changing. While earlier approaches, such as Retrieval-Augmented Generation (RAG), operated according to the push principle, a pull principle is now taking hold. Instead of gathering information in advance and pushing it to the model, the AI decides for itself which information it lacks and retrieves it specifically using a tool. Instead of an information avalanche, a true information selection emerges.
As a result, AI is increasingly taking on an organizational role: It analyzes tasks, identifies work steps and required information, and selects tools or data sources that close these gaps. It becomes the coordinator of information acquisition, which plays to the strengths of the language model. For companies, this means thinking like an information architect. What matters is not the quantity but the correct dosage: the principle of “minimum viable context” (MVC). The AI should receive exactly the information it needs for the next step-no more and no less.
5. Graphs: A Navigation System for AI Agents
The information AI needs next depends heavily on the use case: sometimes deep, linear context chains, sometimes broad, branching knowledge structures, clusters of relevant information, or just a single precise snippet. This is exactly where traditional data structures begin to falter. Graph databases offer a structurally different approach – described by Forrester as the backbone for LLMs – to map, capture, and re-provide context.
Especially in combination with AI agents, graphs will move more into focus in 2026. As AI systems increasingly coordinate decisions, tools, and processes independently, they require robust and transparent context models. Graphs link knowledge, actions, and interactions in real time, making agents navigable, reviewable, and scalable. This creates a semantic information layer – the Knowledge Layer – that not only enables more precise answers but, above all, provides one thing: agents that understand where they are, what they are doing, why they are doing it, and what consequences their next step will have.
6. The Database of the Future: Adaptive
Databases and data infrastructures are becoming a central pivot point for AI success. After four years of AI hype, it is becoming increasingly clear that while hardware and models are advancing into new dimensions, the databases beneath them are still stuck in the thinking of the 1970s. AI systems are expected to deliver peak performance, yet operate on architectures that were never built for them. The key question is no longer how to improve existing databases, but what a database designed for AI actually looks like.
The next generation of AI databases might function similarly to “live code.” Queries are rewritten and optimized iteratively during execution, borrowing from modern compiler designs, such as Just-in-Time (JIT) techniques. The execution plan continually adapts to data distributions, load patterns, and the available hardware. The result is a permanent feedback loop in which the database becomes more efficient with every iteration, even as complexity and data volumes grow. This dynamic architecture forms the foundation for the knowledge layer that AI agents will need in the future.
##
ABOUT THE AUTHOR
As Vice President of Product Innovation at Neo4j, Michael Hunger leads product innovation and developer strategy, bringing over 25 years of software development expertise. For the past 13 years, Michael has held various leadership roles within the company, spanning open source initiatives, the full scope of Neo4j’s graph database and analytics, and the broader graph community of users and contributors. His work also includes generative AI (GenAI), cloud integrations, and enhancing the developer experience.
A passionate developer, Michael excels in programming languages, continuously learning and contributing to exciting and ambitious open-source projects. He has authored and contributed to several software-related books and articles and is an active speaker and organizer at numerous conferences, including those for the Java Champions program.
Beyond his professional endeavors, Michael is dedicated to mentoring the next generation of programmers. He runs weekly girls-only coding classes at local schools.
He is based in Dresden, Germany.
