[0:00] picture this. It's 2026. Right. And you're the CTO of a major Nordic telecommunications company. You've got 2.3 million subscribers spread across multiple countries. That's a massive footprint. Exactly. And you've just made the call to hand over your primary customer service operations, like the whole thing, to an artificial intelligence. Which, for most people, sounds terrifying. Right. I mean, if your media thought is at this ends in a spectacular PR disaster with thousands of furious customers, just mashing zero and screaming operator into their phones. [0:34] Yeah, we've all been there. We have. But the data says otherwise. Instead of a disaster, your customer satisfaction score is just skyrocketed from 76% to 87%. Wow. Your average handling times plummeted from over eight minutes to just three. You saved 8.4 million dollars in a single fiscal year. That's incredible. And here is the kicker across 1.2 billion interactions. You had zero compliance violations. Zero. See, that operational reality completely upends the conventional wisdom we've clung to for the last decade regarding automated support. [1:06] It really does. Because, you know, we've been conditioned to view automation as this necessary evil for cost deflection. Right. Not as a driver of actual customer satisfaction. Which brings us to the core question for you tuning in today. How did we leap from those rigid, infuriating decision tree chatbots that just trap you in an endless loop? I didn't quite get that. Exactly. How did we go from that to AI systems that literally run circles around traditional customer support models? Welcome to another deep dive on the AI insights by Aetherlink channel. Glad to be [1:42] here. Today's deep dive is based on a really fascinating article from Aetherlink. There are Dutch AI consulting firm. Yeah. And if you operate in the European text base, you definitely know them. Right. You know their product lines. There's Aetherbot for the AI agents, Aethermind for Enterprise Strategy, and Aetherday for custom deployments. Today, we are digging into their recent insights on multi-modal AI agents. It's a great piece of research. So, okay, let's unpack this because we aren't talking about a minor software patch here or like a slightly better natural language processor. No, not at all. We are looking at a fundamental architectural paradigm shift. [2:16] Right. For the European business leaders, developers, and CTOs listening, this is really the convergence of two massive, often opposing forces. Like unstoppable force meets immovable object. Kind of. Yeah. On one side, you have the rapid maturation of autonomous multi-modal AI agents. And on the other, you have the incredibly stringent governance of the EU AI Act. Right. Which is notorious for being strict. Exactly. And the Aetherlink research [2:46] illustrates why mastering this intersection matters right now. Because implementing this technology is no longer just a defensive play to trim the fat from operational budgets. It's more than just saving money. Much more. And reprises that architect these systems correctly, particularly those operating in innovation hubs like Helsinki, they're actually turning regulatory compliance into a massive competitive mode. That's a huge shift in perspective. But to understand the return on investment here, we first have to establish a baseline for what the technology is actually doing under the hood. Right. Because the word chatbot is practically a legacy term at this point. [3:20] I like to think of it like this. A traditional rule-based chatbot is basically like a train on a track. I like that analogy. Right. It can only go where the rails or the decision trees were laid. If a customer's query deviates even slightly from that track, the system derails and dumps you onto a human agent. Exactly. Please hold for the next available representative. Yes. But an agentic AI architecture is more like an off-road vehicle with a GPS. You give it a destination, and it autonomously figures out the best route to get there. That is the critical distinction. It [3:54] really comes down to the mechanism of action. Because traditional automation, it just retrieves static information. Like pulling up an FAQ page. Right. But agentic AI driven by large language models and frameworks like React, which stands for reasoning and acting, it takes concrete actions. It actually does the thing. Exactly. The LLM functions as a reasoning engine. So say a customer says their flight is canceled. The AI doesn't just pull up the cancellation policy. It does it do instead. It actually writes a JSON payload to query the airline's real-time inventory API. Then it [4:28] evaluates the available options against the customer's historical preferences. Oh wow. Yeah. And then it executes the rebooking API call. Processes a partial refund through the payment gateway and generates the new itinerary email all autonomously without a human touching it without a single human click. The McKinsey 2026 State of AI report actually notes that enterprises deploying these agentic systems complete tasks 2.3 times faster. That is why they see a 45% reduction in operational [4:58] costs specifically because they are taking actual actions. Because the model is actually authorized to change the state of the database rather than just reading from it. Precisely. And what's fascinating here is the multimodal aspect of that workflow. Right. Multimodality. Let's get into that. These modern agents aren't just processing a single stream of text input anymore. They are ingesting voice, images, video and structured data simultaneously. They're taking it all in at once. Yeah. Projecting all of those different data types into a shared vector embedding space. [5:29] So the system perceives the customer's context holistically. To ground this, the Aetherlink piece highlights a really practical scenario that shows how much friction this removes. The Telecom bill example. Yeah. Exactly. So you're calling your internet service provider about some bizarre unexplained charge on your monthly statement. Always a fun call. Now with a legacy system, you're either trying to explain a complex visual layout to a bot that only understands text or your painstakingly dictating a 16-digit account number over the phone. [6:02] Did you say B as in Bravo? Right. But with a multimodal agent while you're speaking to it in natural language on your phone, the AI has simultaneously ingested the visual layout of your PDF bill. It's literally looking at the same thing you are. Yes. It's queried your historical payment latency in the database and it's processed the sentiment of your past three support transcripts. It fuses the visual context with the acoustic context of your voice in milliseconds. It already knows which line item you're mad about before you even finish your sentence. Exactly. [6:34] And the source says that seamlessness alone cuts back and forth exchanges by 40%. Because it eradicates the cognitive load on the customer, you no longer have to like translate your real world problem into the machine's preferred format. And that visual understanding fundamentally changes the self-service funnel too. Oh absolutely. The research highlights hardware warranty claims. So you get a router in the mail and it has a smashed antenna. You don't want to type out a description of the crap. No, and you don't have to fill out a 10-field web form with serial [7:05] numbers. You just snap a photo from your phone and upload it. And the agent takes over. Right. The vision model isolates the damage, runs optical character recognition on the bar code to verify your warranty and just triggers the replacement workflow in the supply chain system. It's so clean. And the Gardner 2025 Enterprise AI survey backs this up. They found that 68% of enterprises using multimodal AI report massive improvements in first contact resolution. 68%. Yeah. Resulting in a 34% [7:36] average boost to customer satisfaction. Simply because the machine finally has eyes. Okay, but I'm going to push back a little here because I'm struggling to see how some of the downstream claims work in practice. Farron, what's the sticking point? Well, customer service is inherently reactive, right? Somebody has a problem. They contact you. You fix it. Right. But the Aether Link research specifically positions these AI agents as revenue generators driving proactive marketing automation. They do. How does an AI agent jump from fixing a broken router to making a [8:08] sale? I mean, if my fiber connection drops during a crucial meeting and I call in furious, the absolute last thing I want is an AI trying to upsell me a mesh network. That sounds like a recipe for catastrophic churn, honestly. Exactly. I would cancel on the spot. And you would if the system lacked nuance. But this is where real-time sentiment analysis paired with OmniChannel context completely changes the dynamic. Okay, walk me through that. The AI isn't just blindly reading a script. It's constantly evaluating the emotional state of the interaction. Let's break down the [8:39] actual mechanics of your dropped fiber scenario. Okay, I'm furiously calling it. Right. As you're speaking, the AI's acoustic model is measuring pitch variation, your speech ray, your volume. It's detecting that frustration. It knows I'm mad. It knows. The language model is also analyzing your syntactic choices. So the agent updates a hidden state variable that essentially flags you as high-turn risk, high frustration. Gotcha. And during this phase, the AI actively suppresses any logic related to sales. Yeah. It focuses entirely on leveraging its API access to run diagnostics, [9:13] reset the node and resolve your problem. So it's triaging the emotional state first? Precisely. Then it monitors for a state change. Once the router reboots and you verify the connection is back, the acoustic model detects the drop in speech tension. The sigh of relief. Exactly. The shift in the sentiment vector is the trigger. And because the agent maintains context persistence across your account, it knows you've exceeded your bandwidth allocation three times this quarter. Which is probably why you dropped in the first place. Right. So at that exact moment of resolved tension, [9:44] it smoothly suggests that for, you know, five euros more a month, you can avoid this specific bottleneck. It doesn't feel like a cold pitch at all. It feels like highly contextualized, proactive problem solving. It identifies the upsell, executes the tier change, and then routes that data to the marketing platform. So you stop seeing generic ads for that same upgrade. That is seamless. But wait, the system is also taking these millions of interactions and generating public facing content. Right. The source mentions it generates SEO optimized FAQs based on these very [10:18] interactions. It does. But knowing how large language models can, you know, hallucinate, yeah, letting an AI automatically write and publish your company's official help docs sounds incredibly risky. It is incredibly risky if you rely purely on the model's parametric memory. Like the stuff that learned during its initial training. Exactly. But enterprises circumvent this using strict retrieval augmented generation or our rag pipelines. Okay. So how does that solve the hallucination issue? Because the AI isn't daydreaming answers. It isolates clusters of successfully [10:52] resolved tickets where the sentiment ended positively. Okay. It extracts the precise step-by-step API actions that actually fix the problem and then translates that verified machine logic into a human readable FAQ. It's grounded entirely in internal validated enterprise data. Okay, here's where it gets really interesting. Everything we've just covered. Yeah. Streaming voice recordings, analyzing high-resolution photos of billing statements, building intricate sentiment profiles on [11:23] individual European citizens. Sounds like a lot of data. It sounds like an absolute gold mine for operational efficiency. But it also sounds like a massive GDPR and compliance minefield. Oh, without doubt. I mean, the regulatory environment in Europe isn't just strict. It's punitive. How are these companies legally pulling this off in Europe without facing ruin as fines? That is the defining architectural challenge for every CTO right now. But the companies scaling this successfully, especially out of Helsinki and the Nordics, they aren't trying to skirt the legislation. They're [11:53] leaning into it. They are utilizing it as a blueprint. They're actually thriving under the EUAI act. Wait, really? How? Well, if we connect this to the bigger picture, the EUAI act classifies AI systems based on a tiered risk framework. Right. The risk tiers. If an AI is making highly consequential decisions like denying a loan or filtering resumes, it's deemed high risk and needs exhaustive assessments. But customer service and informational agents generally fall into a lower tier. [12:24] But they still have rules, right? Oh, absolutely. They carry substantial transparency obligations, particularly under Article 13. Meaning you can't just have a black box making decisions about someone's bill and throw your hands up when regulators come knocking. Precisely the opposite. The prevailing methodology here is called AI lead architecture. AI lead architecture. Yes. Instead of bolting compliance modules onto a finished product, you design the systems reasoning engine to be intrinsically auditable from day one. Compliance isn't a burden. It actually [12:55] builds customer trust. How does that look in practice, though? Well, when AI agent decides to issue a 50-year credit, it's required to generate a chain of thought log. Like showing its math. Exactly. It records the specific policy document it referenced via the RXJ pipeline, the sentiment score that justified the appeasement and the exact API endpoints it triggered. So if a data protection officer audits the interaction. The system instantly produces a deterministic human readable trail of its logic. That structural transparency seems to be driving the vendor market, too. There's a [13:29] vital data point in the sources from URUSAT's 2025 digital economy survey. I know the one you're talking about. It states that 73% of EU enterprises choose their AI vendors based primarily on data residency guarantees. Yep. They are refusing to send sensitive European customer data to centralize US-based APIs. Which makes perfect sense. And this is exactly why solutions like Aetherbot, which allow for fully on-premises deployment or hosting strictly within sovereign EU data centers, [13:59] are winning out over those global APIs. The data never leaves the EU. Data sovereignty really dictates the entire model selection process now. You simply cannot risk data exfiltration. Where a customer's personal information accidentally becomes training data for someone else's model. Exactly. And because of this, European enterprises are aggressively adopting open source models for sensitive data. Like which ones? Look at Mistral, based in Paris. Their models, like Mistral 7B or Mistral, are highly capable reasoning engines. And their open source. Right. Because the weights are open, [14:33] a CTO can pull that model behind the corporate firewall, deploy it on their own servers, and fine-tune it on proprietary data without a single bite leaving the building. But here's a technical question. Smaller open source models, even highly optimized ones, sometimes struggle with the super complex multi-step logical reasoning that massive proprietary models like GBT4 or CLAW'd handle everlessly. That's true. They do have limits. So does keeping everything local limit the intelligence of the agent? It would if they only used one model. But they use a hybrid [15:04] routing architecture. Oh, okay. So they make some? Exactly. You deploy a fast local open source model as your frontline router. It handles all the PII, the names, addresses, account numbers, and executes the routine workflows. So the sensitive stuff stays locked down? Right. But if that local model encounters a highly complex reasoning task that doesn't involve sensitive data, say, analyzing a massive, anonymized log file of air codes to find a hardware failure. It sends that out. [15:35] Yes. Scrubs any remaining PII passes the anonymized data to a heavier proprietary model via an external API, gets the deduction, and brings the answer back in house. Wow. It optimizes for both privacy and cognitive power. It's the best of both worlds. So with the technology proven and the legal guardrails clearly in place, business leaders need to know the financial reality of deployment. The bottom line. Exactly. Can you lay out the hard numbers on ROI and infrastructure? Because building out an on-premise setup capable of running these models cannot be cheap. We're [16:09] talking about racks of high-end GPUs. You're right. The upfront expenditure is substantial. Moving from discovery to a hardened production environment typically takes about four to eight months. Okay. Forty eight months. And depending on how entangled your legacy systems are, you're looking at an initial implementation cost between 150,000 to 500,000. Half a million euros is a serious chunk of change. It is. But the return profile is driven by the marginal cost of inference. Once the system is live, the compute cost to process a complex multi-step customer [16:43] interaction drops to just eight to fifteen cents. Wow. From dollars per call with human agents down to literally pennies. Exactly. The compute cost is only one metric. What about latency? Ah, latency. Because if I'm speaking to a voice agent over the phone, the physics of network routing become a massive hurdle. If there's a four or five second delay, well, the LLM generates a response. It destroys the conversational illusion completely. People just start talking over the bot. The speed recognition gets confused and the whole thing falls apart. Latency really is the silent killer of voice agents. You have a very strict latency budget to make a conversation feel [17:17] natural. It's typically under 200 milliseconds of total round trip delay. The 200 milliseconds, that's practically instant. It has to be. And that includes the time for speech recognition to transcribe the audio, the LLM to generate the first token, and the text to speech engine to synthesize the audio back. You cannot achieve that if your data is making transatlantic hops. No. The speed of light literally prevents it, which is why edge deployment is critical. I'm guessing exactly. The Nordic companies in the study are leveraging regional data centers in places like Stockholm, [17:50] Frankfurt, or Amsterdam. Keeping the inference geographically close to the collar is the only way to meet that sub 200 millisecond threshold. So what does this all mean when you aggregate those efficiencies? The 8th-year-length data indicates the payback period for that initial half-million euro investment is incredibly fast. Very fast. Usually just 14 to 18 months. And the three-year cumulative return on investment regularly exceeds 300%. Which is the kind of metric that gets board approval instantly. Instantly. However, we do have to address the elephant in the room here. [18:22] When you show a board of 300% ROI based on software autonomously executing the workflows of human agents, the immediate cynical assumption is that this technology is just a precursor to mass layoffs. It's the standard macroeconomic fear, right? But the operational reality on the floor is a story of labor reallocation, not elimination. Explain that. Think about the volume of interactions at telecom handles. The vast majority are mundane, resetting a password, querying a [18:52] shipping status, disputing a standard overage fee. The repetitive grind. Right. When the AI absorbs that massive low complexity volume, you don't fire your human workforce. You upskill them. You move them up the value chain. Exactly. Those human agents are reallocated to the escalation tier. They handle the highly sensitive complex cases like a grieving family needing to transfer account ownership or a severe enterprise outage requiring deep empathetic negotiation. So the AI frees human workers to tackle complex cases? Yes. By removing that repetitive grind, [19:25] enterprises are measuring an effective human labor productivity boost of 25 to 40%. The humans are doing higher value strictly human work. And the system only works because the AI knows when to step back and let the human take over. Right. Exactly. I want to look closely at how that Nordic Telecom avoided disaster and achieved zero compliance violations. They didn't just build a smart model. They built explicit guardrails into the AI's access. The research calls them approval authorities. Yes. Without strict approval authorities, an agentic system is a massive liability. [19:58] You simply cannot give an LLM unfettered access to your core financial databases. Right. Picture an AI negotiating a billing dispute with a very persuasive customer. The AI analyzes the sentiment, calculates the lifetime value of the account, and determines a 600 credit is statistically the best move to prevent churn. It sounds logical for the AI. But the engineering team has implemented a hard-coded API limit. The agent physically cannot approve a credit over 500 euros without a human. So it hits a wall. It hits that digital wall. But instead of crashing or throwing a weird error code, the agent is programmed to gracefully pivot. [20:33] It generates a response like, I understand your frustration. And while I can immediately authorize a 500 euro credit, let me pull my manager into this chat to approve the remaining balance. That's so human. Right. And in milliseconds, it packages the entire summarized context and routes it to an upskilled human agent's dashboard. The human reviews it, clicks approve, and the AI finalizes it. Genuine autonomy requires smart boundaries. It's the principle of human in the loop by design. The AI executes, but ultimate accountability remains tethered to a human. [21:05] This has been an incredibly dense, actionable breakdown. As we wrap up this deep time, let's distill all of this down to our takeaways. Sounds good. For me, my number one takeaway from the A-fueling research is the revelation that multimodal AI isn't just a cost-cutting tool. We have to stop viewing it as just a deflection tactic. When architected correctly, it is a proactive revenue generator. It completely flips the script. It does. Transforming a tense, reactive troubleshooting call into a highly personalized proactive marketing [21:35] touchpoint without alienating the customer, that fundamentally changes the unit economics of an enterprise. I completely agree. And my number one takeaway. Well, this raises an important question for anyone evaluating their tech stack. Are you viewing the regulatory environment as a hurdle or a foundation? Because governance is no longer a speed bump. It's an enabler. Yes. The European companies treating the EU AI Act and GDPR as foundational architecture, rather than an afterthought, are the ones actually moving the fastest. Because they use local open-weight [22:08] models and enforced deterministic audit trails in the code, they aren't bogged down in retroactive legal reviews. They're built for speeding compliance from day one. Exactly. It redefines what enterprise grade software looks like. I want to leave you our listeners with a final thought to mull over. Something to consider as you map out your roadmaps for the next year. Oh, here we go. We've spent this entire deep dive analyzing how your company's AI agents will interact with human customers, right? Right. But if these AI agents are fully capable of executing complex workflows autonomously, what happens next year when a customer decides they don't want to wait [22:43] on hold either? What happens when a customer's personal AI assistant running locally on their smartphone calls? Your enterprise's AI agent to fiercely negotiate a warranty dispute on their behalf? Are we ready for bought to bought customer service? That shifts the battlefield entirely. It's totally different ballgame. Something to plan for. For more AI insights, visit etherlink.ai. Thanks for tuning into this deep dive.