[0:00] Ik heb het een momentje voor een seconde. Je hebt gewoon een product op de lijn, hè. Dat is iets dat je echt voor een trip opgevallen moet. Oh ja, we hebben er allemaal. Exactly. Dus je open de box en je hart gewoon zinkt. Het is vergeten. Oh, het is worst. Het is brudel. En instantaneously, je feel dat... Dat familiegevond van de dreden, die in het gevoel is. Je denkt dat het een geweldigde naam moet zijn. [1:55] En Ather-Dv voor custom bespoke AI-developen. En we hebben de laatste findings in de afgelopen voor de discussie. En de centrale thesis van de onderzoek dat de jaren van 2016, dat is dus op de koren, is de definitief tipping point voor agentische AI in Enterprise-customer-service. Ja, en we gaan kijken specifically aan de lenzingen van Europian, business leaders, CTOs en developers. Because they're the ones trying to balance this massive, unprecedented ROI of these systems [2:28] with the incredibly strict code-level requirements of the EU AI Act. Okay, let's unpack this, because to understand why 2026 is like flashing-in neon lights for every tech executive out there, we have to clearly define what is actually changing. Right. The shift from reactive to proactive. Exactly. I mean the conversational AI we interact with today, De standard chatbot dat popstelt wanneer je het weer terug te pieren is. Het is fundamentally reactief. Ja, het wist voor je. Het wist voor je dat het een keerje wordt. Het scan is het database en het spits uit een preprogrammdancer. [2:59] Het is dus een glorified interactief FAQ page, hè? 100%. Maar agentic AI is een hele andere beest. Het is ook een auto. Het is niet gewoon respond. Het initiatie is actie, het navigatie is complex workflows across different software platforms. En het is echt een leer van de uitkomen. Dus het is optimerlijk hetzelfde proces als het gaat? Exact. Het is een agentie. Ik denk dat het een humanterme term is. Traditionaal AI is een intern die alleen speelt als het spoken is. De intern die handen je de company handbook [3:29] als je een complex kan. Ja, exact. Maar agentic AI is een verantwoord van een vorm. Het zie je een problem, het is een verantwoord van interactief met de inventorie en de billen systems met een verantwoord voor de probleem. En dan, en dit is het geën, het update de interne logik, dus het probleem is handeld meer officieally de volgende keer. Wow. En ik denk dat de markt is validatief dat geën agressief. De eakerlinkreerde de McKinsey AI-report voor 2025 die verantwoord van 72% van businessleaders nu vuue AI is superior [4:00] te humanonly supporten. 72% dat is een massive shift in sentiment. Het is echt. Ja, het is niet echt om het geëngeverdigde alternatie te worden. Ze zijn dat het actiebleer is beter. En de financiele projectie is tijdens dat het is staggering. Oh ja. Bij 2020-6 agentiek systems zijn projecten te handelen tussen 65 tot 80% van alle routine-kustomers servicequestions entierally autonomistisch. Globally, dat drijzen expecteensavies van 80 billen dollar's annually in call center operations. 80 billen, [4:31] het is al heel hard te wrapen om je geëngeverdigde. Ja, en in zoeming en on Europe specifically we zijn er 60 tot 70% kostsavings in Tier 1 support operations. Wat's fascin Rodriguez is hier Не me отправsen in деньги verspreikt dit 마무�iem is niet collega van onzipaald verwachtig zuidelijk. технологievis regist kampels niet beter die volum en euro voor socialdem 악enschesi. Eoi eencial Mileage toeverekel Giving TE being would only recite single czym benchmarking wat [5:01] de agenda politie brrok 24-7 om alle variantie ook Exact als jullie title call center data De vast majority of interactions are high frequency low complexity, things like where is my order, or how do I reset my password? Or can I change by billing address? Yeah. Exactly. So when an agentic system can process millions of those simultaneously, without any geographic staffing limits, shift changes, or holiday pay, that 65 to 80% autonomy rate just becomes a mathematical certainty. [5:33] But I actually want to push back on that 80% number for a second. Okay. Sooms the AI actually understands the customer seamlessly. Yeah. I mean, as a consumer, I actively avoid jetbots right now. Oh, I don't blame you. Right. Because typing out the nuanced context of a problem usually ends with the bot giving me some generic response that doesn't help at all. They haven't quite catch that. Yes. Exactly that. And if the interface is full of friction, the customer will just bypass it and demand a human anyway, which totally kills that ROI, you mentioned. [6:03] That's a very fair point. So how are these 2026 systems processing information in a way that actually resolves that friction? Well, the friction you're describing is exactly what multimodal capabilities are engineered to solve. See, the AI of 2023 and 2024 was heavily constrained by text. And text is an incredibly low bandwidth way to convey context. You have to type a whole paragraph just to explain a physical issue. Which nobody wants to do. Right. But agentic AI in 2026 utilizes multimodal architecture, meaning it natively integrates [6:34] text, voice and image recognition and processes all of those inputs simultaneously in real time. Okay, let's dig into the mechanics of that because it's not just about like taking a picture and attaching it to a text file, right? How does the system actually correlate a crack in a JPEG with my frustrated voice note? It comes down to how neural networks map data. So in a multimodal system, audio, visual and text data are all projected into a shared semantic space using vector embeddings. Okay, so they all get translated into the same kind of map? [7:05] Precisely. So when the AI is computer vision model detects a shattered phone screen in a photo, it encodes that visual data into the same mathematical neighborhood as the spoken words shattered screen from your voice note. Oh, wow. Yeah. The system isn't just looking at a picture and reading a transcript separately. It is fusing them together to gain immediate holistic context. It mimics human sensory perception, but processes the data at machine speed, which totally explains the efficiency metric highlighted in the source because this multimodal approach [7:39] reduces average resolution times by 40% compared to legacy single channel systems. 40% is huge when you're dealing with millions of calls. Absolutely. Yeah. I mean, going back to our opening scenario, the customer submits a photo with a damaged product alongside a verbal complaint. The AI processes the physical damage visually, cross references it with a spoken context, checks the inventory database and authorizes the replacement. Right. With no human triage required at all. Exactly. The etherbot platform handles that entire multisensory intake instantly. [8:11] But, and this is a big, but giving a machine the power to issue refunds, move inventory, and altered databases based on a voice note introduces a massive legal liability. Oh, for sure. If the AI gets the context wrong or hallucinates a policy who's actually responsible, in Europe, the answer to that coin is currently forcing a massive architectural reckoning. Here's where it gets really interesting, because making autonomous decisions is great for the bottom line, right? But if a neural network is automatically authorizing replacements, it is basically making a contractual [8:46] decision on behalf of the company. Yes. And that brings us straight to the EU AI Act. The regulatory landscape in Europe is shifting from these theoretical guidelines to hard punitive laws. Under the EU AI Act, customer service AI isn't just viewed as a harmless conversational tool anymore. It's way more serious now. Much more. If the agentic AI has the power to process a refund, escalate a serious legal complaint or deny service to a customer, it is classified as high risk. It's legally treated as an autonomous, financial, and contractual agent. [9:17] And the penalties for non-compliance are... Well, they're the kind of members that keep executives awake at night. Oh, definitely. We're talking about fines of up to 30 million euros, or 6% of a company's global revenue. Which is astronomical. Yeah. For a multinational corporation, 6% of global revenue isn't just a slap on the wrist. It is a devastating financial blow. It's essentially the regulatory body saying, you must build the brakes before you upgrade the engine. If we connect this to the bigger picture, Aetherlink's research highlights a fascinating [9:50] paradigm shift regarding these regulations. Because most companies naturally view compliance as a burden, right? Oh, yeah. Like a legal hurdle that just slows down deployment. Exactly. But looking toward 2026, compliance is actually a hidden ROI driver. It's a massive competitive moot. Wait, okay. I love that analogy. It sounds like the EU AI Act is forcing companies to eat their vegetables. But Aetherlink is arguing that eating your vegetables actually gives you superpowers. Explain the mechanics of that moot. How does complying with strict regulations actually make a company more profitable? [10:23] It fundamentally comes down to system, architecture, and trust. So first, consider explainability. If an AI denies a customer a refund, the EU AI Act demands that the system can explain exactly how it arrived at that decision. You can't just say the computer said no. Right, you cannot just rely on a black box LLM. You need a transparent architecture like the frameworks Aetherbot uses that maintains instantaneous audit trails. And that level of transparency builds immense customer trust. [10:54] Which is the cornerstone of retention in an Omni Channel environment, right? Exactly. And second, built in bias detection prevents the AI from inadvertently discriminating against certain demographics. That saves companies from highly publicized brand-destroing lawsuits. And there is a pure market access advantage as well, right? The source explicitly states that compliant systems unlock lucrative EU enterprise contracts that non-compliant competitors simply cannot touch. Yes. You essentially get a VIP pass to operate in the European market while your competitors are locked outside. [11:25] And the financial trap here is retrofitting. Early adopters gain an almost instrumentable market advantage because the research shows that retrofitting a non-compliant older AI system to meet new EU standards is three to five times more expensive than building it right the first time. Three to five times. Yeah. Yeah. You can't just bolt an audit trail onto an opaque model after the fact. You have to tear the entire architecture down to the studs to track exactly which parameter triggered a specific refund denial. That makes sense. [11:56] So companies either invest in compliant architecture now to lead the market, or they pay five times as much later just to survive in it. Exactly. It's a stark reality for developers. It really is. Yeah. But let's look at the technical vulnerabilities of the models themselves for a second. Because taming the machine is just as critical as navigating the law. Oh, definitely. I mean, autonomy introduces very specific risks that simply didn't exist with the old rule-based decision tree chatbots. Yeah. And the primary barrier to enterprise adoption remains the phenomenon of hallucinations. [12:28] Because an autonomous system is only as valuable as its accuracy. Right. And the Aetherlink piece gives a very specific, dangerous example of this. Imagine a customer service AI agent confidently citing a warranty clause to a customer, a clause that absolutely does not exist. Just completely made up. Right. The generative model simply hallucinated it. But it delivers the information with absolute confidence, the customer acts on it, and instantly you have a situation that triggers regulatory complaints, voids your compliance, and destroys brand trust. [13:00] It's a nightmare scenario. Yeah. So how do you stop a generative model, which is inherently designed to predict the next most likely word creatively from generating fiction? Well, the technical mitigation for this is RG, which stands for retrieval augmented generation. RG, right? Yeah. In a standard setup, an AI relies on its vast, generalized training data to generate answers, which is where the hallucinations occur. R places strict architectural guardrails on the LLM. OK, but how does it actually look under the hood? Like how does R physically constrain the model? [13:32] It utilizes vector databases and semantic search. So before the LLM is allowed to generate a single word, the RG system intercepts the user's query and searches a closed verified knowledge base. Like the company's actual legally approved warranty documents. Exactly. The database turns paragraphs of that warranty into mathematical coordinates. So when a customer asks about a return policy, the system only retrieves coordinates that mathematically map to that specific verified policy. OK, if I'm following me. And then the LLM is forced to formulate its answer based only on the text retrieved [14:06] from those specific coordinates. Ah, so if the answer isn't in the verified database. The AI physically cannot invent a new clause, because those coordinates don't exist, it anchors the agent in verified reality. That makes perfect sense. It transforms the AI from like a creative writer into a highly restricted synthesizer of the pro facts. Precisely. But beyond factual accuracy, there is a deeply human element to customer service that machines struggle with. The source highlights the critical need for emotional AI. [14:37] Yes, very important. The system has to be able to detect when a customer is getting frustrated, angry or confused. But how does a machine actually know I'm angry? Is it just looking for all caps and exclamation marks? Well, in legacy systems, yes, it was rudimentary keyword spotting. But true agentic workflows in 2026 integrate deep, nuanced sentiment analysis. Because they're multi-modal, right? Exactly. Because these platforms are multi-modal, they aren't just reading text. They're analyzing the acoustic properties of a voice. [15:08] They measure pitch, speech rate, and volume fluctuations. Wow. They can even read micro expressions of video is involved. They understand the fundamental difference between someone typing quickly because they are in a rush and someone typing quickly because they are furious. And that ties directly into the concept of escalation logic. Precisely. Having a rock solid escalation logic based on that deep sentiment reading is non-negotiable. If the AI sentiment analysis detects genuine distress or anger, it must seamlessly hand the interaction over to a human agent. And I imagine it provides that human with the entire context of the conversation instantly, [15:43] so the customer doesn't have to repeat themselves. Exactly. The human steps in fully informed. So what does this all mean? For the CTO or business leader listening right now, we have incredible multi-modal technology, we have steep regulatory hurdles with the EU AI Act, and we have technical risks like hallucinations to manage. Let's talk about the actual friction of implementation. How does an organization deploy this safely based on AetherLinks methodology? It requires a highly disciplined, phased approach to engineering. [16:14] AetherLinks outlines four core best practices for implementation that serve as a practical roadmap. Okay, lay them out for us. Step one is audit and grounding. This is where you implement the Arge architectures we just discussed. You must perform a comprehensive audit of your internal data and anchor your agents strictly in verified company documentation. Got it. And step two is escalation logic. You need crystal clear code level rules for exactly when a human takes over the conversation. The AI needs hard boundaries defining its own limitations. Step three is bias testing, which is incredibly complex in practice. [16:49] This isn't a one and done audit. Right, how does an engineering team actually build bias testing into their sprints? They have to establish continuous testing pipelines. They must regularly evaluate their models against edge cases and demographic variations to ensure the AI isn't developing unfair hidden biases over time. Especially for those high risk contractual use cases that fall under the EU AI Act. Exactly. And the final step number four is customer communication. Total transparency. [17:19] Yes. You must clearly and immediately disclose the user when they are interacting with an AI system. You don't try to trick the customer into thinking they are talking to a human named Sarah. Because as the article emphasizes, transparency is what actually builds trust. Customers don't mind talking to an AI if the AI is competent and honest about what it is. It raises an important question for anyone leading a technical organization today. Look objectively at your current infrastructure. Are your support systems actually ready for this level of auditability, vector based grounding [17:50] and multimodal complicity? Or are you just bolting on API calls to an LLM as an afterthought to an outdated legacy architecture? Because if it's the latter, adapting to the 2026 landscape is going to be incredibly painful and expensive. And that is the multi-million euro question right there. As we wrap up our analysis of this Acerlink research, I think we should distill this down to our main takeaways. For me, it is the practical everyday marvel of multimodality. Yeah, the user experience side. [18:20] Exactly. We get so caught up in the high level business stats, the $80 billion savings, the EU compliance codes. But just think about the sheer relief of a customer being able to show a photo and speak of voice note instantly bypassing 20 minutes of agonizing typing and repetitive back and forth. What's changing for the consumer? It really is. It fundamentally changes the customer experience from an interrogation into a simple, frictionalist transaction. The technology is finally matching the speed of human thought. That reduction in friction is absolutely going to redefine brand loyalty. [18:52] For me, though, the standout insight is the strategic view of regulation. The mode concept. Yes. We are so conditioned to view legislation specifically things like the EU AI Act as an inhibitor of innovation. But the reality is that regulation is creating a powerful competitive moat. The companies that are agile enough to embrace compliance early to build transparent, auditable, biased, tested systems from the ground up, they are going to lock in market [19:22] share that their slower, non-compliant competitors simply won't be legally permitted to touch. Wow. Compliance isn't a tax. It's an investment in market access. It's a gold rush, but only for the organizations that took the time to build the right safety gear. Exactly. And I want to leave you with one final provocative thought to mull over as you look at your own roadmaps. That's here. If a gentick AI successfully handles 80% of all routine customer service tasks autonomously by 2026, think about what remains. That 20%. Right. [19:53] The remaining 20% of tasks that human workers will still have to handle won't be simple password resets or shipping updates or basic returns. Now the AI handle all of that. Exactly. They will be the absolute most complex, the most emotionally charged and the most critical edge case interactions accompany faces. It fundamentally changes the job description of a human customer service representative. Oh wow. I hadn't thought of that. How will your company need to completely retrain, compensate and emotionally support its human workforce when their entire day consists of handling only the hardest, most stressful [20:27] problems? That is a profound operational challenge. We are leveraging AI to solve one massive problem of scale. But in doing so, we are creating an entirely new dynamic for the human workers left in the loop. Yeah, it's going to require just as much innovation in human resources and team management as it does in artificial intelligence and machine learning. Absolutely. You can find more AI insights at etherlink.ai. Thanks for joining us on this deep dive.