According to Gartner, 40% of today’s agentic AI projects will be canceled by 2027. But does that mean businesses should pause AI adoption? Absolutely not. What it does mean is that businesses need to become more strategic. Success with AI isn’t about chasing hype but rather about solving real problems.
We sat down with Felix Rothmund, Director of Artificial Intelligence & Data, to explore what sustainable AI adoption really looks like, and how companies can avoid becoming just another statistic.
Q: With Gartner predicting that 40% of agentic AI projects will be canceled by 2027, do you think we’re in danger of an AI “overhyped expectations”? What’s causing this wave of under-delivery?
Felix: In fact, I would say that the actual percentage of the cancelation could be even higher. Many of the solutions we’re developing today will not stand the test of time, and that’s okay. These efforts lay the groundwork for more robust, effective AI agents in the future.
After all, early-stage project failures shouldn't automatically be seen as evidence of underperformance. They’re a natural part of the innovation cycle. And yes, expectations around AI are undeniably inflated. But more concerning than the hype itself is the risk of inaction. Organizations that hesitate or wait too long will simply be left behind.
When it comes to failures, I would say there are a few common reasons. One big one is jumping into solutions way too quickly. It’s totally understandable. Digging into the root of a problem can feel slow and a bit tedious, whereas building something feels exciting and productive. But if you don’t really take the time to understand what you’re solving for, the solution often misses the mark or doesn’t hold up in the real world. Felix Rothmund,Director of Artificial Intelligence & Data at intive
Q: That makes sense. Anything else you think contributes?
Felix: Definitely. People often underestimate just how much foundational work AI needs. There’s this misconception that AI can magically make sense of messy, outdated, or scattered data. But in reality, if the underlying data infrastructure – things like availability, structuring, cleaning and governance – isn’t solid, the AI just can’t perform well.
And even when the tech is solid, projects can still fail if there isn’t strong alignment from stakeholders. It’s one thing to build a prototype, but scaling it? That takes long-term vision, real buy-in from leadership, and a clear plan. You need to be able to tell a compelling story about what the AI is doing, why it matters, and how it fits into the bigger picture, otherwise, it tends to stall out after the proof-of-concept phase.
Q: The rate of successful task completion for AI agents, as measured by researchers at Carnegie Mellon University (CMU) and at Salesforce, is only about 30 to 35 % for multi-step tasks. What are the biggest technical or design challenges preventing higher completion rates, and how close are we to overcoming them?
Felix: Multi-step problems require more than just surface-level understanding. They have to be broken down into smaller sub-tasks, and that means the system needs to develop a coherent plan. That involves a level of reasoning that current models still struggle with. Large language models aren’t reasoning in the way humans do; they’re generating outputs based on patterns from their training data, not from a true understanding of cause and effect.
Q: So even if the response sounds reasonable, it might not actually be correct?
Felix: Yes. The model can generate a sequence of steps that seems plausible, but those steps might include errors. For example, if the model tries to use an external tool incorrectly, that mistake can derail the entire process. Sometimes those errors are recoverable, but more often, they lead to cascading failures that prevent the task from being completed.
Q: What role does memory play in this?
Felix: Memory is an important piece of the puzzle. Multi-step tasks require persistent working memory, the ability to track progress, remember past decisions, and adjust plans on the fly. AI agents try to simulate this by saving notes or logs and then feeding those back into the model as context. But unlike humans, who use notes to jog their memory, these notes are the model’s memory. Every model has a finite context window, and once you hit that limit, older information has to be dropped or summarized. That can lead to inconsistencies in planning, where earlier decisions get lost or overwritten. So even if the agent starts with a solid plan, it can lose track of it as the task progresses.
Q: So, in short, the low success rate reflects a combination of limited reasoning, fragile tool use, and constrained memory?
Felix: Each of those factors adds friction, and together they make it very difficult for AI agents to reliably complete complex, multi-step tasks. There’s been impressive progress, but we still have a way to go before we see truly robust task execution at scale.
Q: How is it possible to understand if we have an AI use case that’s really transformative or one that’s maintaining the status quo with a more expensive toolset?
Felix: The first step is simple: fully understand the problem you’re trying to solve. Too often, generative AI is applied to processes because it delivers quick, visible results. But speed alone doesn’t equate to value. The real test is whether the AI-enhanced process is more efficient, more accurate, more cost-effective, or more satisfying for users. To make that evaluation, you need to clearly define what success looks like. Are you aiming to reduce costs? Improve employee satisfaction? Increase throughput or output? The metrics must be concrete and aligned with business goals.
Consider this example: imagine integrating a generative AI chatbot with access to your Power BI dashboards. You might ask, “What’s the forecasted revenue for next quarter?” and get a response. But is that actually the best way to access and use that data?
Without a clear understanding of the process, you're enhancing, and the metrics that matter, it's easy to mistake novelty for innovation.
Q: What frameworks do you use to decide whether a use case is good for AI augmentation, full automation, or no AI at all?
Felix: Too often, traditional innovation sprints spend most of the time on solutioning and leaving minimal room to define success or assess real impact. intive’s GenAI Sprint flips that model.
Over four weeks, we dedicate two weeks to exploring the problem space and identifying the true levers for change. Then, one week to prototyping and testing feasibility with GenAI tools and one week to structuring findings and building a clear path to production, including stakeholder-ready pitch material
This process ensures you're not just applying AI for the sake of it, but where it can truly make a difference.
Q: Why do you think so many organizations fail to define ROI for AI initiatives? And how do you avoid that?
Felix: Because it’s genuinely hard. Defining ROI for AI starts with understanding the problem you're solving, and that takes time and discipline. If you don’t know how many hours your team spends on a specific task, how can you judge whether automating it is worth the investment?
Take AI coding assistants, for example. Measuring productivity by lines of AI-generated code is misleading. True value comes from a broader view: Are your developers shipping faster? Are users more satisfied? Has the product quality improved? Are your developers happier?
To avoid the ROI trap, you need to start with meaningful success criteria and a holistic understanding of value, then work backward to assign a business case and price tag.
Q: One of the promises of AI is that it will empower humans and not replace them. How can companies design AI systems that enhance human work, not deskill or displace it?
Felix: AI will inevitably take over some human tasks, ideally, the ones we don’t enjoy or that are repetitive. The key, at least in the foreseeable future, is balance.
Organizations need to design AI systems that support and empower employees rather than replace them. That means focusing on augmentation: giving people smarter tools, not sidelining them. Involving employees in the design and rollout of AI tools helps ensure the systems enhance human judgment, creativity, and productivity.
Q: What AI trends will redefine process augmentation in the coming years?
Felix: In the coming years, large language models will continue to improve, though progress will likely be more incremental and specialized compared to the rapid advancements of recent years. At the same time, we’ll see a shift toward agent-centered design, where applications are built not just for humans but also for AI agents capable of accessing tools and knowledge through emerging protocols like MCP.
As AI becomes more accessible, no-code and AI-assisted coding tools will continue to empower non-technical users to build their own workflows and automations. Meanwhile, robotics is poised for significant breakthroughs, as research expands beyond language into comprehensive world models, enabling machines to navigate and operate in human-designed environments more effectively.
Overall, the coming years will be about balancing the strengths and limitations of AI. It’ll be a critical skill to decide when to trust AI judgement, when to train and deploy specialized ML models and when to rely on traditional software engineering.
.png)
.png)
.png)
.png)
.png)
.png)