July, 2030 – A new software unicorn has just been minted in Shoreditch, East London. 

The CEO logs in as usual, and gathers daily updates from the team. The Chief Technology Officer is suggesting a new feature to ship. The Chief Product Officer wants to redesign the ERP integration. The Chief Revenue Officer is showing off the new pipeline. The CCO is discussing the latest customer health and product feedback. The CISO has found a new privacy conflict, which they are addressing with a newly-revised infrastructure set-up.

This sounds like every software business you’ve ever heard of. But the difference is that the CEO’s teammates are entirely AI, not human:

• The CTO is Lovable
• The CPO is Cogna
• The CCO is Gradient Labs
• The CRO is 11x
• The CISO is Zylon

Back to 2024, and the hottest topic in software is AI agents. Founders are now rapidly standing up agentic applications that can solve specific needs in functions like sales and customer success – without a human in the loop. Meanwhile, software buyers seeing real opportunity to quickly improve their P&L are quickly building or purchasing agentic products. In recent months, investors have been pouring hundreds of millions of dollars in capital into startups in the space. 

Europe is home to many of these exciting companies, as you can see from our market map below. For example, H, a French AI agent startup, raised a $220 million seed round in May. 

AI agents mark a step-change from Robotic Process Automation (RPA) bots, which – as we explored last year – have several limitations due to their deterministic nature. Next-generation AI agents are non-deterministic, which means that instead of stopping at a “dead end” they can learn from mistakes and adjust their series of tasks. This makes them suited to complex and unstructured tasks, and means that they can transform the journey from intent to implementation in software development. They can deliver “pure work”, rather than acting only as a helpful co-pilot. 

We believe that the rise of AI agents is not only an opportunity to expand automation beyond what is possible with RPA, but to more broadly redefine how knowledge work is performed. Why? Because, given the right guardrails, next-generation AI agents have the potential to effectively and safely replace knowledge workers in many scenarios.  

These agents are about to revolutionise the world of work as we know it, and they are already getting started. For example, Klarna recently revealed that its AI agent system handled two-thirds of customer chats in its first month in operation.

In order to understand the massive impact that AI agents will have on the world, we decided to get some answers from those who know best. So, we interviewed leaders at some of the most exciting companies in this space: Ben Peters at Cogna, Anton Osika at Lovable, Dimitri Masin at Gradient Labs AI, Hasan Sukkar from 11x, and Iván Martínez Toro at Zylon. We asked these sector experts to explain the impact of AI agents on software development, and to outline to what extent they will replace knowledge workers in the future.

Ben (Cogna) and Anton (Lovable) are founders of agentic startups at the forefront of fundamentally changing how  software applications are built. 

Until now, the process of building software has been resource-heavy, expensive, and time consuming. Non-technical teams and companies have relied on excess internal engineering capacity (which is rare) to deliver their software projects, or on expensive and slow IT consultancies. This set-up has meant that app development has been inaccessible to vast parts of the economy.

With the rise of AI agents – including Cogna and Lovable – this is changing. 

Non-tech companies can now use these agents to build applications faster and at lower cost, because they reduce the knowledge bar to software building, enable rapid iteration cycles, and generally extend far beyond anything co-pilots or code generation engines have previously delivered. Therefore, these agents can directly improve productivity and company performance.

Cogna, for example, addresses the creation of complex internal applications that impact operations. It can replace mundane tasks that previously locked up time of knowledge workers, such as managing workforce training schedules — which Cogna already does for Cadent Gas — and overseeing procurement and supply chain areas. In Ben’s own words: “We use generative AI and agentic workflows to build solutions in days rather than months… We are tackling productivity as a critical issue, and unlocking it could free up resources for other essential areas to improve your operations.”

Lovable offers a similar overhaul for customer-facing applications. These applications are revenue generating, and previously required a developer background, deep product understanding, and multiple iterations including testing and customer feedback. As Anton says: “Our mission is to give everyone the engineering leverage to shorten the cycle of turning an idea into a product that users will love.”

Cogna and Lovable might make this look easy, but the reality is that agentic application builders are challenging to build for two reasons: 1) agents require necessary context and guardrails to develop the right products, and 2) translating user software needs from natural language into software requirements is both an art and science.

Anton solves the problem of context by enriching foundational models instead of competing with them. He explains: “We believe in building on top of the ever-improving [foundational] models by providing the missing context, domain knowledge and tooling. For us ‘open ended agents’ aren’t the answer, as errors and wrong paths compound and lead to bad results. But when you put sufficient guardrails around the agents they become increasingly powerful. This is why our approach focuses on teaching agents to use the best available tooling – think about us humans using internal developer environments (IDEs) and DevOps practices.”

Ben also explored how it is critical to crack the translation of natural language into application requirements. Cogna helps users ‘discover’ opportunities where software can help, ‘define’ solutions with natural language instead of tedious requirements engineering, and ‘deliver’ solutions to complex problems with precision, custom software. Its agents efficiently collect business requirements across stakeholders, before translating those intentions into the language of software engineers (functional and non-functional software requirements). This helps deliver shorter iteration cycles and cost savings, as well as ensuring greater stakeholder control.

With context and user understanding in hand, it’s no surprise that Anton believes that AI agents are your future CPTOs. His closing words when asked ‘what’s coming next?’ were: “Today Lovable is a full stack engineer, but in future it will become the CPTO of your company that manages a group of AI engineers.” 

We couldn’t agree more. 

When we spoke, Hasan (11x) quoted Brian Halligan’s summary of the shift AI agents will drive – that agents will turn traditional “software-as-a-service” applications into “service-as-a-software”.

Like Ben and Anton, both Hasan and Dimitri (Gradient Labs) agree that it will no longer be about “selling a productivity-improving co-pilot but, instead, selling the work itself directly”.  The founders all believe that AI agents will eventually replace knowledge workers, and that agentic startups will tap into budgets previously allocated to salaries as a result. As Hasan puts it: “This opportunity is infinitely more vast: we get to eat into budgets for services and employees, not for technology,”  

Both 11x and Gradient Labs are already developing your soon-to-be GTM AI colleagues.

11x has built your sales development representative: Your AI colleague even has a name – Alice. Alice is 11x’s virtual sales development representative, who will soon be able to conduct the first voice phone call with a potential customer. In time, the company expects Alice to own her own direct sales quota, and “report” progress in daily sales huddles. Hasan says: “This means codifying the SDR role into a series of discrete tasks, and using agents to pick them off one-by-one. She quickly graduates from our customers’ human guidance to fully autonomously sourcing, researching, and outbounding leads.”

Gradient Labs kicked off with a customer service agent: The company is enabling autonomous agentic work by codifying customer service workflows. Dimitri explains: “We started out with a customer service use case, providing E2E fully autonomous agentic work and there will be more interesting use cases coming shortly. We will leverage deep domain knowledge to drive this process that we gained from Monzo.”

With increasing availability of more powerful LLMs, and the continuous improvement of tooling for building LLM-powered products, we believe that the barrier to build agents will decline over time. When challenging the founders on that, Dimitri pointed out that the sector is “still observing a large gap in how to map business workflow logic into agentic workflows” which will only be solved by applying hyper domain-specific knowledge intricacies of functions, workflows, habits and patterns as well as integrating seamlessly into firms’ existing tech stacks.

Hasan agrees: “We believe that “off-the-shelf” agent builders will struggle in these use-cases without deep domain-specific reasoning informing their individual workflows.”

Agents offer companies an incredible way to deliver efficiencies and enhance productivity, and verticalised agents have a clear route to market. 

But, as data sits at the core of all AI agent development, agentic startups still have to contend with serious questions from businesses about trust and privacy. These questions are completely valid – and they could hamper the ability of agentic applications to go into production. Without the right guardrails in place, integrating agents into your business could go horribly wrong. Imagine what would happen if a company’s brand new customer success agent revealed a client’s personal, private data to another customer. It would be a disaster, both legally and reputationally. 

For Iván at Zylon, the solution is to bring AI into your secure environment and build there. Zylon has developed a production-grade, enterprise-ready, end-to-end solution that utilises PrivateGPT – an Open Source AI project that lets developers create data-protecting AI applications.  

He explains: “Data privacy is a legal requirement and there are significant privacy risks related to using AI… For contextualised interactions, private data sources like documents and databases could be connected to the AI platform which requires data to be read, analysed, and stored, increasing the risk of exposing sensitive information. Given the direct access to data required in these scenarios, the only way to ensure full privacy and minimise breach risks is to bring AI to your data, rather than exposing your data to AI.” 

While many companies lack the technical resources to build their own solutions – and as such, need Zylon’s off-the-shelf product, some businesses, especially larger enterprises, are keen to solve for privacy by building privacy-sensitive agentic applications themselves. To help these companies, Dataiku (one of Europe’s largest AI companies, and one of our portfolio companies) and TitanML, have partnered together to allow businesses to deploy LLMs privately by allowing them to self-host their models privately and securely use AI within their organisations.

We asked Meryem, TitanML’s CEO, what she thought about agents in enterprise, and for her, it’s a combination of security and ease of use: “At TitanML, we see a great opportunity in deploying agents within enterprises. We’re tackling two main challenges: ensuring data privacy with agents accessing sensitive business data through effortless self-hosting, and simplifying the deployment of complex, multi-model systems to be GPU-efficient and transparent. Our solution is the Titan Takeoff Inference Stack, designed specifically for enterprise deployment of Retrieval-Augmented Generation (RAG) and agent applications.”

We would like to extend a sincere thank you to all the founders who gave their input to this article. We are extremely excited by the rapid developments and progress they and others are making with AI agents, and we’re on the lookout for companies developing in this space. If that is you, please get in touch with nils@dawncapital.com, shamillah@dawncapital.com, zoe@dawncapital.com and owen@dawncapital.com.