At FINN, we’ve always believed in moving fast and automating even faster. From the beginning, we’ve seen automation not just as a technical tool, but as a mindset that drives scale and impact. That’s why we’ve embraced no-code tools like Make and Retool across all our mission-based teams, building a culture where quick automation is second nature.

From this foundation, a clear pattern emerged: solve a problem once, then automate it for good. Or as we like to say: Do it. Do it. Automate it. This mindset has helped us scale operations efficiently, reduce repetitive tasks, and empower non-technical teams to create their own solutions.

To take this even further, our CTO, Andi, introduced the role of Business Automation Manager (BAM) three years ago. Recently, we expanded to Business Automation & AI Manager (BAAM). People in this role are typically digital natives who sit at the intersection of business and tech. They rapidly build solutions grounded in domain expertise and powered by cutting-edge low-code tools. The BAAM role is a direct expression of our core belief that the closer business and tech work together, the stronger the outcomes.

However, we realized even our best workflow automation shared one crucial limitation: they relied on very pre-defined (a.k.a. deterministic) patterns. This is both the core strength and main weakness of Robotic Process Automation (RPA). As long as a process follows a predictable path, it works flawlessly. But introduce one exception, and it breaks until it is manually fixed. This meant we sometimes found ourselves building fragile solutions dependent on rigid, step-by-step instructions rather than fully automated systems. Fortunately, a new hope emerged: AI Agents.

Our journey with Agentic AI began in early 2025, when we first started exploring the concept in depth. What drew us in was a compelling promise: AI Agents acting as our future digital coworkers. Instead of breaking when faced with unexpected behavior, they learn to navigate ambiguity and adapt through reasoning and continuous improvement. In short, automation with a brain. Recognizing our ambitious subscription growth targets, we identified AI Agents as a strategic opportunity to significantly enhance company value by scaling top-line revenues disproportionately to bottom-line personnel expenses. Of course, we weren’t alone in embracing this trend; it seemed every other tech startup suddenly claimed to have the ultimate AI agent solution. Most, however, fell short of their promises. This led us on an intensive search across the AI landscape to discover the true game-changer for FINN.

How do we effectively build AI Agents and roll them out across the organization? This was the guiding question we had in mind lacing up to fight the noise around AI. Reflecting upon the progress we made over the last months, the following factors stood out to make a difference for us:

  1. Follow the hype selectively We started where many do: scrolling through LinkedIn. And while a lot of the hype was just that, it turned out that some of the most frequently mentioned tools actually earned their place in the spotlight. We tested over 50 tools and more often than not, the ones promoted by credible voices in our network delivered real value.

Pro tip: Tailor your LinkedIn feed. People like Armand Ruiz, Alexandre Kantjas, Allie K. Miller, and Rakesh Gohel helped us shortcut hours of research by sharing what actually worked. Also, make use of aggregators such as aiagentslist.com or aiagentsdirectory.com.

  1. Know yourself Even the best AI solution will fall flat if your team doesn’t have the mindset, skills, or structure to use it. At FINN, we knew exactly what we were looking for: something our BAAMs could jump right into and create business value with. That ruled out anything overly technical; we didn’t want to build AI Agents in code-heavy frameworks. We needed tools that were intuitive, low-code, fast to experiment with, and provided the base for a steep learning curve in Agentic AI.

  2. Build a pipeline Navigating the AI buzz meant juggling dozens of tools and demo/sale calls. The only way to stay sane? Building a pipeline. We used Airtable to track the status of every tool we tested, what stage it was in, what problem it aimed to solve, and why it did or didn’t make the cut. This became our internal source of truth and a way to align across teams.

  3. Get hands-on How do you know if a provider is actually capable of solving your problem? You have to test them. Often, this process involved signing up for numerous waiting lists and utilizing our Google SSO for any domain that seemed promising. We observed a significant difference in product readiness between self-service platforms offering free trials and scenarios where a salesperson was pushing for a multi-month trial. We recommend opting for self-service options; if a product is robust, providers aren’t afraid to allow users to cancel at any time.

  4. Involve those affected The market tells you AI can solve everything, but where should you start? We decided to ask the experts: our 350+ FINNies. We sent out a simple form and received over 100 remarkable use case submissions. Some were straightforward traditional automation candidates, but many aligned perfectly with our Agentic vision. The bonus? Involving the whole company built bottom-up excitement and momentum around AI at FINN by addressing their biggest pain points.

  5. Cluster problems Our submissions quickly revealed that there’s no single AI tool that solves everything. However, we noticed patterns: some problems were deeply specific to FINN (such as car appraisals), while others were universal across industries (known as verticals, for example, customer care). This clustering helped us prioritize and sequence our efforts more effectively.

While we had successfully implemented solid solutions for vertical use cases, a larger question emerged: How could we build Agentic solutions specifically for FINN’s unique challenges? No off-the-shelf solution existed to automate a car subscription company with our custom processes. We needed an AI Agent builder that matched the speed and usability of our preferred tools, like Make and Retool.

This realization set the stage for our next chapter, moving beyond implementing existing AI agents to building our own custom solutions that could handle the unique complexities of our business model.

Hacking our way forward

We ran a company-wide Agentic Hackathon in March 2025-the biggest one in FINN history. Over 50 FINNies from across Tech, Operations, Fleet, Sales, and Customer Care participated. Everyone curious about building with AI Agents came to the table to try things out.

Based on our preliminary internal research and use case collection, we prioritized taking on the following seven real-world challenges. Here’s what we learned about each:

  1. Analyzing customer feedback for actionable insights
    We found out verticals like Enterpret work much better.

  2. Responding to delivery drivers’ questions
    Worked extremely well with Relevance.

  3. Running financial risk analyses on B2B prospects
    Worked extremely well with Relevance as well.

  4. Leverage reasoning to turn car configuration PDFs into website-ready data
    We found PDF quality varying per brand, raising the question how to scale.

  5. Approach new B2B leads we can partner with
    We found Lindy to be very powerful for dialogue-driven personal agents.

  6. Onboarding new joiners faster with prompt libraries and aggregated knowledge
    Worked well, but Sana required a long-term contract.

  7. Calling small local workshops to receive updates on cars’ repair status
    Worked surprisingly well with LeapingAI and ultimately won the hackathon with a stunning live demo: Congrats!

Beyond the tooling question, we also obtained learnings worth their weight in gold:

  1. AI Agents are not the catch-all solution

Analyzing large data, like thousands of customer reviews and tickets is better suited for high-control machine learning or data analytics solutions built with code. This is not where Agentic AI shines as of today. Not every problem requires an AI solution and not every AI problem requires an Agentic solution.

  1. AI cannot read our minds

AI Agents can do a lot and are insanely powerful, if provided with the right and enough context. This is where effective prompting comes into play, which means precisely translating and expressing what we want. It is an emerging skill that we must be willing to learn and shape. We consistently underestimate the amount of implicit human context in our brains.

  1. Start small and iterate fast

While it’s tempting to solve everything at once, the real power of AI Agents lies in focused, incremental progress. Smaller, testable iterations reduce complexity, surface insights early, and lay the groundwork for scalable solutions. Setting realistic expectations is key to long-term success.

  1. Build trust through rigorous, real-world testing

Trust is paramount. People need to believe in what the agent produces and understand how to intervene if needed. Since LLMs are probabilistic rather than deterministic, consistency cannot be assumed; it must be earned. At FINN, we build this trust by testing extensively on real or production-like use cases, tracking outputs to identify regressions, and validating reliability. We emphasize collaboration throughout the process as two pairs of eyes are always better than one. This ensures that human-in-the-loop oversight and robust evaluation pipelines make our agents dependable.

  1. Tooling is evolving-but not perfect

From shady credit limits over flaky integrations to exhausted context windows, we’ve learned to expect (and debug) a lot. As of today, there isn’t one perfect low-code AI Agent Builder in the market, but some come very close by offering a nice blend of ease-to-use plus a large variety of integrations–and we’re actively developing strategic partnerships with these front runners. The market continues to move fast and what was a bottleneck yesterday could already be solved by tomorrow-we are ready for it.

Even more important than our use cases, tools, or hands-on learnings is that the Hackathon changed the way we think about AI across the company. It proved that Agentic AI isn’t science fiction, but it’s here to stay.

Building our Agentic agenda

Following the momentum of our hackathon, we continued to aim high and move fast. We didn’t just talk about AI; we built with it, tested it, and created a strategy around it. We developed an ambitious Agentic Agenda after our Hackathon: we hired an extra set of interns to amplify our BAAMs, embedded Agentic projects into our team-level OKRs, and restructured our BAAM team to maximize knowledge sharing and centralize innovation.

This enabled us to already hire an entire squad of Full-Time Agents (FTAs) across various business functions. FTAs highlight our ambition to mimic partial or complete work of human roles’ job descriptions that we would usually hire FTEs (Full-Time Employees) for. Each Agent is designed to autonomously execute repetitive, high-leverage tasks and to escalate to a human when needed. The results speak for themselves: together, these agents demonstrate the transformative potential of Agentic automation at scale. Here are some examples of the more than 13 powerful agents we already built at FINN:

  • Hades, the Underwriter-Financial Risk Analysis
    Saved over 80 hours of manual research by autonomously assessing B2B risk profiles.

  • Lenny, the Lead Launcher-Sales Development
    Boosted lead-to-meeting conversion to 16% via personalized outreach.

  • Conny, the Configurator-Content Creation
    Achieved 100% automated image extraction for Hyundai and BYD configurators.

  • Enrico, the Enricher-Sales Development
    Qualified over 2,500 B2B prospects by injecting custom data into our CRM.

  • Dom, the Driver-Delivery Coordination
    Resolved 50% of all driver-related requests automatically via multi-turn context handling.

Not every experiment was a success story and that’s part of the process. Ingo, the Invoice Classifier, started as a Relevance-based agent for labeling, forwarding, and sorting documents. But in Accounting, where errors need to be avoided at all costs, even small inconsistencies became unacceptable. We learned that reliability beats complexity and replaced the original version with a single Gemini prompt governed by deterministic rules-simpler, cleaner, and much more consistent.

What’s next?

Agentic AI is not a moonshot-it’s a shift in how we automate, collaborate, and build. At FINN, we don’t believe in waiting for technology to be perfect but in learning by doing. Logically, we continue to dedicate resources, run follow-up focus days, hackathons, and invest in rolling out company-wide education around Agentic AI.

We’re sharing this journey not because we’ve figured it all out, but because we believe in collective progress. The Agentic shift won’t be built by one company alone-it will come from teams sharing what works (and what doesn’t). So if you’re experimenting with Agentic AI too, we’d love to hear your stories. Let’s keep learning, testing, and building-one reliable agent at a time.

This article was written by Jonathan Hippe and Maximilian Gebhard.