# In the Loop 01: the gap between a demo and production Canonical: https://brew.new/templates/rasa/in-the-loop-01-the-gap-between-a-demo-and-production Brand: rasa.com Category: newsletter ![Preview of In the Loop 01: the gap between a demo and production](https://cdn.brew.new/email-preview-d5d1dcf6c306a91d-tracking_r57p50chdqppc1cr0h99t4p34x8dvzb3-1790176945868.png) ## Email content Issue 01 | August 2026 Rasa In the Loop Everything worth knowing in AI this month, minus the hype Welcome to the first issue of In the Loop, Rasa’s monthly roundup that cuts through the AI news cycle and flags what actually matters if you’re building agents for a real enterprise. This month, OpenAI shipped GPT-5.6 after Washington asked it to wait, a new challenger undercut everyone on price, and the leaderboard reshuffled yet again. But the story we keep coming back to is about the widening gap between what AI can do in a demo and what it can safely do in production. Gartner now predicts more than 40% of agentic AI projects will be scrapped by 2027. Nearly eight in ten enterprises have already had an AI security incident, and the pattern is hard to miss: the hard part is the control. That’s the lens we’re bringing to everything below. – The Rasa Team What we’re reading The latest AI developments worth your attention Phone showing ChatGPT, Claude and Gemini app icons OpenAI released GPT-5.6, but only after clearing a federal security review On July 9, OpenAI released GPT-5.6, its most capable AI model to date. That came after a delay reportedly prompted by U.S. government concerns about the national-security risks of increasingly powerful AI systems. The capability race continues to accelerate, but it’s notable that a frontier launch now comes with the potential for a government-issued pause. Rasa’s take: When even the model makers are being told to slow down and show their work, blind trust begins moving off the table as a launch strategy — and visibility into what an AI system actually does becomes table stakes. Read the full story from Reuters Illustration of AI moving to private cloud Enterprises are quietly moving AI back behind their own walls Broadcom’s Private Cloud Outlook 2026 found that most IT leaders (56%) are now running or planning to run production AI on private cloud. Public-cloud use for related workloads fell 15 points YoY, with security, cost, and data sovereignty driving the shift. Half of enterprises have already moved tasks away from public cloud, with intent to increase private-cloud investments jumping from 51% to 72%. Rasa’s take: This is the market voting with its infrastructure. When AI touches sensitive, regulated data, control over where it runs stops being optional. Owning your stack and your data is fast becoming the default rather than the exception. Read more from Broadcom European Union flag Brussels blinked, and the EU AI Act just got its first rewrite On June 16, the European Parliament voted 423-57 to amend the EU AI Act for the first time, pushing several high-risk compliance deadlines later (high-risk employment obligations now land December 2, 2027). At the same time, they added new prohibitions, including a ban on AI-generated non-consensual and abusive imagery. Council sign-off is expected before August 2. Rasa’s take: The deadlines moved, but the direction didn’t. Regulated buyers still have to prove how their AI system makes decisions, but a later date means more time to change course and build on an architecture that can show its work. More thoughts on that at Ogletree (below) and on the Rasa blog. Read more from Ogletree Edge case Bizarre, unexplained dispatches from the AI world Portrait of a fictional character Why do so many AI models keep inventing the same guy? Researchers recently noticed that an oddly specific character, “Elias Thorne,” turns up in a startling number of the short stories generated by chatbots — around a quarter of them, by one analysis. He often emerges alongside a similar set of recurring words, like “lighthouse” and “keeper.” The leading theory is that, as models increasingly train on other models’ output, they drift toward the same safe, derivative center in a slow-motion “model collapse.” It’s funny until you remember that your own production system might be quietly converging, too… What we’re thinking about Expert analysis and points of view worth sitting with Gartner’s warning: 40% of agentic AI projects may not survive into 2027 Gartner projects that more than 40% of agentic AI projects will be canceled by 2027, pointing at governance gaps, unclear business value, and weak risk controls rather than model quality. “Agent washing,” where a plain chatbot gets rebranded as an autonomous agent, also plays a role. This is the clearest external validation we’ve seen that deployment discipline is the real bottleneck to AI success. Only 7% of organizations have actually scaled agentic AI Teradata’s survey of 1,000 senior tech leaders found that under 1 in 10 have operationalized agentic AI at scale. Most (63%) report little or no return so far, and an even greater number (77%) say too much of their enterprise data lacks the context agents need to act on it. This comes even as 90% plan to spend more on their agentic AI initiatives. The data makes it clear: the roadblock isn’t necessarily ambition or budget, but whether the underlying system is grounded enough to trust in production. What if you could really watch the model think? Anthropic published new interpretability research describing an internal “workspace” where its model appears to represent and reason about concepts before putting them into words. It’s still early and exploratory, but it’s also a serious attempt to tackle the hardest problem in AI, which is understanding what these systems are doing on the inside. It’s essential reading for anyone who cares about explainability (and if you’re working with AI, you probably should). What Rasa is up to Updates worth a look Ai4 2026, August 4-6, The Venetian, Las Vegas Meet us at Ai4 2026 in Las Vegas, August 4–6 We’re closing out Rasa on the Road 2026 at Ai4 in Las Vegas, where we’ll be sharing how regulated industries can ship enterprise AI agents without giving up compliance, control, or security. Meetings are filling up, so reach out early if you’d like to set up some time with our team. Book some time with us Introducing the new Rasa Developer pathway and certification course Introducing Rasa University: a free, hands-on learning path for agent engineers We just launched Rasa University, a simple way to learn how to build enterprise AI agents on the Rasa Framework. The first course, Foundations, runs about 3.5 hours across six chapters. It takes you from a basic FAQ agent to a fully orchestrated one — covering memory, enterprise search, and agent autonomy — and ends with a Rasa Developer Certificate. Learn more about the course Two categories, one platform: Rasa in CMP's 2026 Prism Rasa recognized by multiple analysts It’s been a strong season for analyst validation. Rasa was named an Honorable Mention in the 2026 Gartner® Magic Quadrant™ for Conversational AI Platforms, with Gartner highlighting customer ownership and control. We were also recognized in the CMP Prism as an Up & Coming provider for Chatbot/Virtual Agent and a Core Performer for Voicebot/Conversational IVR. Read all about it Build your next AI agent with Rasa Power every conversation with enterprise-grade tools that keep your teams in control. Get a demo Rasa Build trustworthy AI agents for real-world use. GitHub X LinkedIn YouTube Community Not interested? Unsubscribe here. © Rasa Technologies Inc.. All rights reserved. Rasa Technologies Inc., 1 Embarcadero Center Suite 1200, San Francisco CA 94111 View in browser [Open and remix this design](https://brew.new/templates/rasa/in-the-loop-01-the-gap-between-a-demo-and-production) [Browse email designs](https://brew.new/browse/templates)