I spent the last two days at Web Summit in Lisbon, sitting through talks from companies building AI agents today, not in five years. From Replit's autonomous coding agents to Intercom's warning about "the death of SaaS," here's what's actually happening in the AI agent space.
Customer Experience is Being Completely Reimagined
The most consistent message across multiple talks: AI agents are becoming the primary interface between brands and customers.
Jesse Zhang from Decagon put it bluntly: conversational AI isn't just a support channel anymore, it's becoming "the new UI for the brand." Think personalized concierge, not chatbot.
The numbers back this up:
- Chime: 60% cost savings with 2 points of NPS improvement
- Oura: 70% of customer problems solved by AI, with a 1:20 people-to-agent ratio
- Hertz: Made customer experience their main priority before even talking to vendors
Malte Kosub from Parloa (they just raised a €130M Series C) made a prediction that stuck with me: in 2-3 years, most customers will prefer talking to AI agents over humans. Not because humans are bad, but because agents will be faster, always available, and actually remember your previous interactions.
The bigger shift? In about 3 years, apps won't have complex UIs anymore. You'll just have a search bar where you ask what you want to do. "Book me a flight to Paris next Tuesday" and it's done. No menus, no forms, no clicking through seven screens.
The SaaS Business Model is Under Threat
Des Traynor from Intercom gave one of the most provocative talks: "The Death of SaaS, The Rise of Agents." His mantra? Agentify, agentify, agentify.
His main point: AI is a convergent force. It collapses multiple tools into one. SaaS companies are essentially just UIs for relational databases, and agents are going to penetrate the data layer directly. That's where the stack compresses.
Timothy Young expanded on this: SaaS companies are facing serious pressure: lack of growth, shrinking margins, changing business models. Big SaaS companies will struggle. Smaller ones might adapt faster because they're more nimble.
The only real advantage SaaS companies have? They own the data. But that won't be enough if they don't figure out how to integrate agents into their essence.
The Technical Challenges That Actually Matter
Context management is one of the biggest problems right now. Michele Catasta from Replit talked about their approach: sub-agents that protect the core agent's context without compressing it. They also have supervision agents that test the app, the code, and the APIs in real-time.
Hallucinations are the other major issue. Graham Sills from Luminous (they work on understanding legal contracts) laid out what actually works:
- Fine-tuning with specific data
- Multiple models doing the same thing
- Citations (though I kept wondering: how do you know the citations aren't hallucinated too?)
- LLM as judge - or actually, a jury of models
They're model agnostic, and they still keep lawyers in the loop. Full automation isn't the goal. (For now?)
Traynor emphasized this too: 80% good isn't enough. You need to compare AI reliability to human reliability (humans are also not 100% reliable). But reliability is everything. It's why most previous AI implementations failed.
Daniel Hulme added an important caveat: AI is great at showing you the world, but terrible at making complex decisions. It's good at using algorithms, eventually at creating algorithms, but decision-making is still opaque compared to traditional software.
And honestly? Just automation can solve many problems. You don't always need fancy generative AI.
Stephen Bye, speaking about network operations, pointed out something that applies across every industry: most companies are working with incomplete datasets. AI can help extrapolate and work around bad data, but it's not magic.
Here's what emerged as critical data challenges:
- Organizations don't know what they have: Most companies have no idea how bad their data actually is until they try to use it.
- Incomplete datasets are the norm: Legacy systems, siloed databases, inconsistent formats - the data infrastructure at most companies is messier than anyone wants to admit
- It's a foundational issue: The agents can be sophisticated, the models can be state-of-the-art, but if the underlying data is messy or incomplete, none of it works
This kept coming up as one of the main signals that a company will fail at AI deployment. You can have the best AI team and unlimited budget, but if your data isn't well organized, you're building on sand.
Why We Don't Need Superintelligence
Max Tegmark gave what he called an "optimistic vision of AI" - and it's not what you'd expect.
His main argument: people don't want AGI, big tech companies want AGI.
The problem with superintelligence is that it combines three superpowers simultaneously: autonomy, generality, and intelligence. That combination creates unpredictable threats and raises really hard questions we're not ready to answer.
But here's his key point: we can achieve all the big promises of AI without superintelligence. Curing diseases, creating wealth, removing humans from boring tasks. All of this is possible with really good generative AI and AI agents, without pursuing superintelligence.
Instead of chasing AGI, we should focus on solving the problems we already have with current AI: prompt injection, hallucinations, bias detection, etc. He mentioned some fascinating applications like "vibe coding with verification" using formal specifications, and AI for truth-finding (like Verity.news detecting bias automatically).
The message: we're pursuing the wrong goal. Tool AI can give us everything we actually want, without the existential risks.
The People Problem is Bigger Than the Tech Problem
This came up in almost every talk: organizational transformation is the hardest part.
Timothy Young laid it out clearly:
- Employees start managing agents instead of doing their actual work
- Companies need to rethink their organization structure
- KPIs need to be redesigned
- What even counts as "human work" needs to be redefined
And here's the thing: companies have no idea about their own workflows or edge cases. Traynor was emphatic about this - if you don't understand your processes, don't try to automate them.
New roles are emerging:
- Agent Managers (managing AI performance, not people)
- CX Architects (designing agent experiences)
- Engineers with different responsibilities (making sure systems run, not building everything from scratch)
Companies that will fail at AI deployment show these signals:
- Never built software before
- Constrained budget for the AI team
- Data isn't well organized
- Chasing quick wins instead of real differentiation
- Siloed organizations (AI creates more problems here, not fewer)
Traynor called this "the most chaotic moment in his life." Organizations need massive amounts of education and guidance to make this transition work.
Big Companies Are Moving Faster (And That's Unusual)
Here's something that surprised me: contrary to normal tech adoption patterns, big companies are investing way more in AI than small and medium businesses.
Usually, SMBs adopt faster, then enterprises follow. With AI, it's inverted. Small and medium businesses are struggling with day-to-day integration and process changes, while enterprises are moving aggressively.
This creates a real competitive gap. The companies that typically move fast are getting left behind while enterprises double down on AI infrastructure.
What This Actually Means
A provocative statement I heard: "If you're not deploying agents right now, your company is probably dead soon."
I don't think this is a death sentence, it's a warning. You have time, but you need to start now. Not just for customer-facing applications, but for internal processes too.
The good news? AI won't take your job in the next 5 years. What it will do is take away boring tasks.
You have time to train. You have time to prepare. But the clock is running.
The right question isn't "what can AI do?" It's "how can AI solve problems for my customers? How can it solve problems for my team?"
Web Summit made one thing clear: the trend isn't coming, it's the norm. The question is whether you're moving with it or getting left behind.
Ready to Start? Assess Your AI Readiness
Remember those failure signals that kept coming up throughout the conference? Companies that don't understand their workflows, have messy data, or chase quick wins instead of real differentiation?
The first step is understanding where you actually stand.
We've built an AI Automation Readiness Assessment based on the patterns we see in companies successfully deploying AI agents. It takes 5 minutes and gives you a clear picture of:
- Where AI agents can create the most value in your organization
- What your biggest blockers are (hint: it's probably not the technology)
- Specific next steps based on your current state
You have time to prepare. But as every speaker at Web Summit emphasized: you need to start now.