By Hugo Casteret, Gabrielle de Massol and Zeevi Michel. Edited by Anya Brochier and Marc Rougier.
Over the past few decades, our digital world has been shaken up by countless so-called “revolutions”, some have quickly vanished, others have reshaped entire industries. We’ve ridden the waves of the early dot-com boom, Web2.0, mobile, IoT, data, cloud, SaaS, and more recently the breakthroughs in AI. We’re familiar with what a real “digital revolution” looks like. Agentic automation looks poised to spark one of the most profound transformations yet, one that will upend not only how software is built and deployed but also how it’s delivered, monetized and integrated into enterprise workflows.
We need to sit up and pay attention. This convergence of GenAI with next-generation agents is more than a neat upgrade, it’s a fundamental shift that promises to rewrite business models and strategic roadmaps across all industries. While there’s plenty of unknown in the details, from data sovereignty to R&D implications and business models, we’re convinced that agentic automation stands out as a true revolution, not just another evolution.
In this new paradigm, we are seeing both the emergence of native AI agent companies born from the latest innovations in LLMs and GenAI, and the transformation of existing SaaS companies who are adopting these new agents to internally transform and expedite their product development. As investors, it’s our role to look at both of these players although it might be tempting to only focus on the new joiners, there is a lot of opportunity for existing SaaS companies who could accelerate their velocity and efficiency by embracing the agentic revolution.
The numbers are undeniably in its favour. Since July 2024, we’ve observed an astonishing 14x increase in the number of AI companies classified as agents in the AI agents directory. However, the landscape isn’t just about new entrants: major players like Microsoft and Salesforce have launched their own iterations of agents, while Anthropic recently enhanced Claude with advanced agent-like features, OpenAI introduced its Operator to facilitate multi-step tasks and Mistral is expected to follow suit with multi-step agents and data connectors. The number of startups is growing, especially in the U.S., but we anticipate swift growth and adoption across EMEA. Globally, for context, over 60,000 AI companies are currently active, with 25% based in the U.S.
We’re going to cover how we got to this revolution, what agents are, the enterprise shift and finally how we are seeing the field emerge and the business models we see agentic startups adopting. Let’s dive in!
Part 1: From Evolution(s) to Revolution
Abstracting away complexity
At the heart of technological evolution lies a push for simplicity: from virtualization to cloud computing and NoCode platforms, the evolution was towards keeping complexity away from end-users with increasingly simplified tech landscapes. AI now takes this further, enabling unprecedented abstraction and efficiency. Large Language Models (LLMs) have ushered in an era of “seamless complexity,” allowing users to interact with high-level instructions while the underlying technology handles complex tasks.
Yet, this abstraction needs to be trustworthy for the end-user. For example, Elaia portfolio company, .txt, is developing tools to make LLMs more reliable and efficient for software development, implementing safeguards to prevent errors before they occur to bring trust to those technological evolution. Indeed, users are more likely to adopt a tool that is not only intuitive and easy to navigate but also demonstrably reliable in its performance and integrity. Trust is the cornerstone of technological adoption, and as AI-driven abstraction deepens, ensuring transparency and dependability becomes paramount.
The evolution of Human Machine Interfaces (HMIs)
HMIs that previously relied on screens, buttons and tactile interfaces with multiple clicks and steps are evolving. We are beginning to see the emergence of low interface technology through LAMs (Large Action Models: AI models designed to understand human intentions and translate them into actions within a given environment or system) and voice interaction. Voice is the first area we see shaping the future of HMIs as in our portfolio company GetVocal AI (Elaia portfolio company), whose voice AI agents enhance customer interactions, boosting leads and opportunities. Similarly, Rabbit R1 operates on a LAM, performing actions based on user touch and voice input which aim to lower cognitive load and accelerate the path from question to answer.
HMIs can also be boosted by hyper-personalization and contextual awareness. At scale they can disrupt interactions from traditional LLMs and retrieval-augmented generation (RAG). Next-gen platforms like Koios leverage continuous user profiling and real-time behavioural signals to deliver results tailored precisely to the user’s role, history, and current objectives. AI software can go even further by predicting the user intention : paving the way for novel approaches to software creation, interaction and operations, revolutionizing how we approach digital tools and solutions.
From infinite knowledge to infinite know-how
If the internet brought access to infinite knowledge to the people, AI has unlocked infinite know-how. Specialised skills like accounting (briefcase), logistics (cube), architecture (ark) or software development are now accessible via a wide range of applications and AI services. By simplifying complex tasks like server provisioning, memory management, and API integrations, engineers can focus on more complex tasks and problem-solving while increasing technology accessibility for a wider range of users. In the case of software development, tools developed by companies like Tabnine (Elaia portfolio company) and Cursor enable non-technical users to pick up coding practices more easily while seasoned developers can increase their productivity by 30%. This action-based knowledge paradigm shift is focused on creating greater efficiency, at all levels of the workforce and automating menial tasks.
Part 2: Task-driven agents are changing the paradigm
From thought to action: what is an agent?
Agents are autonomous, contextually aware entities capable of executing tasks end-to-end. They combine advanced understanding with a structured decision framework and a set of integrated tools, pushing beyond the static workflows of Robotic Process Automation (RPA).
These AI agents go beyond passive, predetermined instructions: they reason about context, learn from novel data, and autonomously adapt their actions. Drawing on capabilities that blend natural language understanding, reinforcement learning, and tool integration, agents can now manage long-tail use cases with highly variable inputs and outputs that were previously too complex or costly to automate.
An agent can be thought of as an AI-model “wrapped” in operational logic and empowered with action capabilities:
According to emerging industry standards and leading AI research (such as OpenAI’s policy guidelines or work by Stanford’s Human-Centered AI Institute), these factors differentiate agents from a pure GenAI product by embedding enterprise-specific logic and risk mitigation strategies into its operations, showcasing orchestrated operational intelligence unlike RPAs. Studies from IDC and Gartner highlight that organizations using AI agents with integrated tools are significantly reducing manual effort and error rates in complex workflows, surpassing the limited capabilities of RPA bots that merely replay recorded actions without adaptive reasoning.
From Robotic Process Automation (RPA) to Agent Process Automation (APA)
LAMs combine LLM capabilities with autonomous, decision-making architectures. These agents move beyond simple text generation: as OpenAI’s Operator demonstrates, it can directly interact with websites that lack a formal API, executing API calls, triggering workflows across SaaS platforms, and even learning to incorporate new tools on the fly, dramatically expanding the boundaries of what can be achieved autonomously or Mistral, after Le Chat agentic capabilities, is expected to release multi-step agents with data connector soon.
For enterprises, the potential implications of this shift are vast. AI agents diffuse decision-making with context and adaptability, transforming how businesses operate. H Company (Elaia portfolio company) released their first AI agent, Runner H, that can execute nearly any task with human-like precision from a simple prompt, this agent uses pixel-perfect navigation to operate web interfaces as smoothly as a skilled human, interpreting pixels and text to “see” and understand screens. This is the first step on their roadmap to reach Artificial General Intelligence (AGI), an autonomous human-level cognitive AI.
We are transitioning from traditional RPA and static PaaS to autonomous AI agents that understand language, orchestrate complex operations, and dynamically optimize enterprise automation with minimal oversight.
Verticalized agents
As the AI agent ecosystem matures, we’re witnessing a diverse and rapidly evolving landscape where solutions range from broad, general-purpose AI assistants to highly specialized, verticalized agents.
Our AI Agent Market Map
We’ve built a (non-exhaustive) map to cover the initial players we’re seeing emerge in the agentic AI field.
Digging into this AI agent landscape, a few trends have begun to stand out. First off, when it comes to industry-specific solutions, think sectors like healthcare, we are still seeing very few pure AI-native players, meaning that this wave has yet to fully hit the sector.
On the language model side, it’s interesting to note that only a handful of projects have been funded exclusively to build LAMs. Instead, the field is largely dominated by emerging champions like OpenAI and Anthropic, who are backed by deep-pocketed investors with a strong focus on building foundational AI.
And then there’s the evolution of voice agents. These aren’t just run-of-the-mill digital assistants anymore, they’re rapidly transforming into sophisticated, context-aware interfaces that are redefining human-machine interaction.
As we’ve seen through startups, deep specialization is already revolutionizing core business processes:
- 11x and Artisan leverage domain-specific insights to streamline sales pipelines, from lead generation to deal closure, boosting accuracy and operational efficiency.
- DeepOpinion focuses on enterprise AI to automate complex tasks across multiple industries, handling everything from insurance claims to loan processing.
Why verticalization drives ROI
Verticalization maximizes ROI by aligning solutions with industry-specific needs and fosters differentiation. This approach offers numerous advantages that contribute to improved financial performance and operational efficiency:
- Specialized terminology and workflow minimize errors and friction, producing more reliable, real-world outcomes.
- Reduced training cycles and fewer AI “hallucinations” enable quick wins, accelerating product launches and cutting costs.
- In regulated sectors like finance or healthcare, verticalized agents often come preloaded with compliance safeguards, reducing legal risks and protecting data privacy.
It’s likely that there will be a growing wave of “digital specialists” across legal, finance, marketing, IT and beyond. While some verticalized agents still require human oversight before running end-to-end processes, there is huge long-term potential for fully autonomous operation. Success, however, depends heavily on rigorous data management: companies with robust data practices will unlock these agents’ full power, while those lacking a strong data foundation will struggle with scattered, siloed information.
Part 3: Orchestrating agents and enterprise shift
Horizontalized agentic platforms
Agentic automation platforms are rapidly transforming enterprise workflows by unifying disparate tools, streamlining data orchestration, and intelligently delegating tasks.
While vertical agents offer laser-focused capabilities, horizontal agentic automation platforms aim to unify an enterprise’s entire AI workflow under one cohesive framework. Think of them as “AI operating systems” that orchestrate specialized agents across diverse business functions and data sources in real time. In centralizing these moving parts, they eliminate silos, streamline complex integrations, and drive data-informed decisions at each organizational level.
Agentic automation platforms are already reshaping enterprise workflows by bringing together disparate tools, simplifying data orchestration, and delegating tasks among AI modules. Beam is one example of this horizontal AI operating system. It acts as an enterprise orchestrator, enabling AI models to communicate seamlessly amongst themselves and with various internal databases, lowering overhead and accelerating processes through a unified interface.
What could a horizontalized platform look like for organizations?
In an ideal world, horizontalized platforms would offer enterprises the following:
- Unified data lake: Consolidates data from departments like Marketing, Finance, and HR, creating a single source of truth that reduces duplicate efforts and simplifies cross-functional reporting.
- Integrated platform: Integrates key AI modules into one platform, cutting down on maintenance, onboarding, and the need for multiple specialized tools.
- Enterprise Agility: Makes it easier to launch new AI initiatives without reinventing the wheel, leveraging existing data pipelines and tools for faster time-to-market.
Ultimately, these platforms replace “patchwork AI” with a streamlined intelligence layer, making organizations more adaptive, resilient, and productive. Over time, we may see horizontal and vertical approaches converge, seamlessly integrating specialized agents into enterprise-wide orchestration layers, creating multi-agent systems (MAS).
Agent-to-Agent Interaction & Multi-Agent Systems (MAS)
One of the most compelling aspects of AI-driven systems is agent-to-agent interaction, where specialized agents collaborate or critique each other’s outputs. When structured properly, these exchanges significantly boost accuracy. For instance, a financial forecasting agent might pass a proposed budget to a compliance agent for regulatory validation, resulting in more robust decisions than any single agent could achieve alone.
At the core of this approach lies the concept of a Multi-Agent System. In MAS, each agent operates autonomously with specialized capabilities and objectives, yet coordinates or competes with others when needed. Early examples include:
- Microsoft AutoGen: Facilitates agent “conversations,” enabling brainstorming and vetting among multiple models within a higher-level framework.
- Relevance AI: Provides a no-code platform to rapidly deploy multi-agent solutions, assigning specific tasks to each agent for adaptive, scalable workflows.
By embracing multi-agent architectures, enterprises move beyond simple task automation, unlocking semi-autonomous or even fully autonomous capabilities. The result is a level of operational synergy that single-agent setups can’t match.
Enterprise Cross-Function Interactions
In many companies, core functions such as finance or HR are siloed organisationally and separate software platforms and databases that can lead to misaligned incentives, duplicated effort and unnecessary overhead.
AI agents can bridge these gaps by pulling data from multiple sources, validating it through inter-agent communication, and alerting human stakeholders only when anomalies arise. This approach streamlines routine data and status updates, boosting cohesion and productivity and can turn employees into managers more than task-doers.
A Multi-Agent System in Action
Picture an enterprise where finance, HR, R&D and marketing teams operate in near-real-time coordination under a MAS, unifying previously siloed data into a shared “lake” accessible by all relevant agents. The benefits are threefold:
- Comprehensive overview: Centralizing data and enabling transparent agent-to-agent communication gives executives and teams a 360° view of operations, sharply reducing blind spots.
- Optimisation: Predictive and prescriptive insights help agents flag inefficiencies or emerging issues, then propose or even automate solutions.
- Adaptability: Agents continuously interact, reallocating resources, adjusting timelines, and refining targets, all with minimal downtime.
Shifting toward this “AI-Managed” workflow could significantly improve how decisions are made, how resources are deployed and how teams collaborate, assuming the right agentic architectures are in place. It will also clearly change how enterprises are managed and organised, and what new skills employees will need to adopt.
Key Challenges
MAS offer the promise of a more efficient central system, but they’re not without their risks. Here are the three main challenges we foresee facing the widespread adoption of MAS (it’s also where there is opportunity for innovation by young companies):
- Security & Compliance: As agents share sensitive information, strong access controls, encryption, and real-time audits are essential to ensure regulatory compliance (GDPR, HIPAA for eg.) and guard against cyber threats.
- Workforce Acceptance: Highly autonomous systems can spark skepticism. Employee trust must be earned through transparency, clear roles, and thorough training. Nvidia CEO Jensen Huang encapsulates this shift, stating that “the IT department of every company is going to be the HR department of AI agents in the future” (Fortune). This means organizations must rethink how they onboard, train and manage these digital workers.
- Interoperability: For effective multi-agent orchestration, standardized protocols, robust APIs and consistent data schemas will be vital.
By proactively addressing these issues, organizations can capitalize on AI-driven cross-functional orchestration while preserving control, compliance, and trust. Just as software is disrupting traditional systems of record, AI-powered agents could dismantle departmental silos, enabling real-time, cross-functional decision-making that transcends conventional organizational boundaries.
4: Our Take at Elaia
AI is carving out distinct lanes in the business ecosystem, clearly differentiating between AI-native startups and legacy companies pivoting to AI-first strategies. AI-native companies, built from scratch around AI tech, are leveraging this tech to push boundaries in efficiency and innovation but are only emerging just now. Don’t discount legacy companies yet. Traditional firms aren’t just surviving; they’re adapting, using AI to overhaul their distribution systems and operational frameworks. This isn’t just a trend, but a fundamental shift in how businesses operate, with everyone from startups to stalwarts racing to harness AI’s transformative power. From the investor perspective, both company profiles offer areas for growth as it’s not just about integrating AI, but how can AI make product delivery faster and cheaper, accelerating growth.
In the case of AI agent adoption, there are lessons to be learnt from the SaaS GTM playbook as utility will be dependent on proving consistent, strong ROI. While the transition from wrapped LLMs to full LAM systems is still in its infancy, the increasing rise of vertical plays in agentic AI will be the first places we see it develop in its full breadth.
Adapting the SaaS Business Model Playbook
As AI agents run continuously, learning and interacting at every hour, the typical SaaS model of “access fees” is quickly losing their appeal. Instead, usage-based pricing is emerging as the new standard: customers pay only for what they use, and vendors share in the gains when outcomes improve. This dynamic organically drives deeper collaboration: When the agent performs, everyone benefits. As Satya Nadella warned, static SaaS pricing may soon be a relic.
Early AI companies datas showed that they reach on average 50–60% gross margin mainly due to computational costs, human involvement, R&D and scale challenges, much less than the 80–90% median gross margin for SaaS. Variable costs could rise even more with agentic companies. When usage spikes, LLM API-related costs can scale greatly, putting pressure on these startups’ gross margins. Providers may pass some costs along, negotiate volume discounts, or opt for more specialized (and sometimes cheaper) LLMs if top-tier accuracy isn’t required. Balancing usage costs and accurate tools with the right ROI will be critical for sustainable growth. AI agent companies with usage-based pricing now have interest to maximise utility to maximize topline. Customer success teams will play a crucial role here as they monitor agent performance and ensure immediate, measurable impact, and might become a hiring focus for these companies as they look to scale usage. This can foster a virtuous cycle: stronger user outcomes drive higher usage, which in turn generates more revenue.
At Elaia, we view this as a major shift in the ways that traditional SaaS revenue structures have been laid out, challenging how investors and enterprises alike forecast and plan. It will be more important than ever for startups to have investors that can navigate these shifts and we will actively seek entrepreneurs and companies that can navigate the trade-offs between rising costs and real user value.
Still, much like the early SaaS days, the perfect KPI playbook for AI agents hasn’t yet emerged, and, deciding which metrics truly matter for agentic AI enterprise success is still ongoing.
Agents must prove their ROI to avoid high churn pitfall
AI agents bear a much heavier burden for ROI than traditional SaaS ever did. After seeing demos that promise near-magical outcomes, enterprise buyers now expect tangible impact within days or weeks, far quicker than the rollout timelines SaaS could once get away with. If those results don’t materialize swiftly, clients move on faster than they would from a conventional subscription.
This elevated bar for performance needs from the dramatic claims around cost savings, operational speed, or revenue boosts that AI agents commonly make. Founders are therefore under intense pressure to deliver verifiable results almost instantly.
From ‘wrapped’ LLMs to a LAM-based system
Beneath these ongoing shifts lies a crucial question: How can AI agents achieve the best possible performance and accuracy? We see the rise of LAMs as a genuine game-changer. Today’s AI agents often rely on a suite of LLMs and integrative tools, but once LAMs become widely available, the leap from basic text outputs to complex task execution could and should grow exponentially.
Agents built on LAMs will learn to orchestrate diverse workflows, adapt nimbly to shifting conditions, and integrate new capabilities without lengthy set-up. Enterprises will soon spot the difference between an LLM ‘wrapped’ in a few guardrails and a LAM-based system, one capable of pixel-by-pixel UI interactions, autopiloting financial transactions, or dynamically reconfiguring supply chains. That’s why we believe the toolbox and parameters around these models will be essential for success that will need further computing costs and investments.
More and more specific plays
Most live AI agents remain horizontal by nature, designed for broad, cross-functional applications as operating systems for organizational workflows. However, we believe a strong wave of verticalization is on the horizon, particularly by industry, by profession and specific use-cases.
From Elaia’s perspective, we’re excited and clear-eyed. Large incumbents are moving fast, but there’s room for new entrants, especially in Europe, where domain expertise, abundant talent and data sovereignty concerns fuel unique opportunities. It’s not impossible that a proliferation of ‘agent factories’ and specialised ‘agent enablers’ could emerge in the next few years, however AI usage has not been fully adopted by all companies, let alone agents. Enterprise and SMB adoption of AI and agentic AI is still in its infancy, like the adoption of the cloud twenty years ago. While competition is fierce for agentic AI dominance, there is still room for innovation, especially across Europe.
Changes in infrastructure, usability and deployability: traditional SaaS, AI and agent-first companies
Traditional software companies are increasingly incorporating AI agents into existing products to automate tasks, improve decision-making, and enhance user experiences. This transition often involves moving from monolithic architectures to microservices or containerized deployments that can handle specialized AI workloads. Companies typically upgrade their tooling — from version control and CI/CD to MLOps pipelines — to ensure seamless development, testing, and deployment of agent capabilities. While market excitement fuels exploration of novel use cases, the real-world adoption of AI agents remains moderate, underscoring the need for careful evaluation of ROI, data security, and operational maturity before scaling.
Agent-first startups architect their technology stacks around advanced AI models from the outset, which entails a strong emphasis on GPU/TPU resources, robust orchestration (e.g., Kubernetes) for dynamic scaling, and integrated observability solutions to monitor performance and reliability. MLOps practices become foundational: data collection, training, and continuous updating of models must be automated and closely tracked, Hopsworks, addresses these needs with its platform for realtime AI lakehouse for example. Delivering these agent services at scale requires a mature infrastructure play — beyond just model hosting, it involves building or leveraging cloud platforms that offer streamlined deployment, security, and compliance. Founders must take into account a realistic view of cost and complexity, alongside measured adoption.
If we assume that “intelligence is the ability to adapt” (so says Stephen Hawking), we are not yet facing full multi -agent systems, but the possibilities of agentic automation represent a fundamental shift towards systems that are not just automated but deeply integrative, adaptive, and capable of learning in real-time.
The possible implications for business are profound. Traditional operational models will be challenged and redefined as AI agents streamline decision-making processes, optimize operations and drive strategic initiatives autonomously. As media, fellow investors and tech giants have been discussing, the integration of these intelligent systems into everyday business practices will enhance productivity and ideally foster innovation by freeing up human capital to focus on creative and strategic pursuits rather than routine tasks.
Ultimately, some of the winners will be those organizations that can effectively harness these capabilities to create new value propositions and competitive advantages. Companies that can adapt to this new paradigm quickly will find themselves at the forefront of their industries, while those that lag may find it increasingly difficult to compete.
As investors, our role is to identify and support those visionaries who are building the future of agentic automation. If you’re building in the space, we’re excited to hear about what your vision might be.
