Why most organizations invest in AI-
but fail to reach production
Lotan Lebkovitz
As a founder, one of your biggest challenges isn’t building AI – it’s getting people to actually use it. Adoption is often the biggest barrier to success – not technical capabilities. Even the most advanced AI application will struggle if users don’t trust it, don’t understand how to integrate it into their work, or feel it over-complicates existing workflows.
This playbook focuses on Enterprise AI solutions – those that integrate seamlessly into existing tools and real-world workflows, keeping users in control while improving efficiency. These companies focus on assisting users and enhancing decision-making, rather than fully automating processes. We learned those lessons from the experience of embedding AI into hundreds of workflows through our portfolio companies in recent years.
Recently, AI agents – autonomous systems designed to handle entire workflows end-to-end – are on the rise. While this trend is evolving, it’s too early to conclude what works for startups succeeding in this area.
The following lessons focus on how founders can successfully integrate AI solutions into enterprise workflows:
As a founder building a product designed to integrate into existing workflows, your goal is to build an AI-driven solution that enhances how people work, rather than making them feel replaced. Ultimately, you want users to trust and even love the product, seeing it as a helpful partner, not a threat to their role.
Users don’t adopt AI simply because it’s advanced; they adopt it when it helps them make better decisions with more confidence. Therefore, AI should act as an assistant, supporting and improving human work rather than taking over. Trust, transparency, and real value are key to AI adoption. Users are more likely to embrace AI when they understand its role, see clear benefits in their daily work, and feel confident in its outputs.
That said, AI adoption doesn’t happen instantly. AI startups that push full automation too soon may face resistance, as users need time to adjust and trust the system. As a founder, introducing automation gradually – rather than all at once – can help build trust and improve user adoption over time.
So far, the best AI-driven companies have taken an incremental approach, ensuring that human oversight is part of the process until users trust the system and feel in control.
What works: AI that starts with decision-support tools, where users remain involved but benefit from automation. Over time, as trust builds, automation can increase
Common Pitfalls: AI that forces users to give up control too quickly, leading to distrust and resistance.
Example
When ActiveFence started, its AI-driven moderation primarily flagged harmful content for human review, allowing content moderation teams to make final enforcement decisions. This initial approach ensured that AI recommendations were validated by human experts, helping build trust in the system.
As ActiveFence scaled and demonstrated its effectiveness, more customers opted for automated moderation through an API, reducing the need for manual review.
For founders, ActiveFence demonstrates the power of gradual AI adoption – starting with human oversight, proving value, and only increasing automation as users develop confidence in the system. This balance between efficiency and human validation helped customers transition from manual review to full automation without sacrificing trust or accuracy.
FOUNDER TAKEAWAY: If AI helps people do their jobs better, they will adopt it. If it disrupts their role, they will resist it. Rushing automation before users trust the system can slow adoption rather than accelerate it.
As a founder developing a product aimed to be integrated in existing workflows, your Enterprise AI product better complements how people work, rather than requiring them to adjust to it. Founders who design AI that integrates smoothly into familiar processes tend to see higher adoption rates.
To get this right, you need to understand your users: how open they are to changing habits and workflows, how likely they are to adopt new tools early, and how much disruption they’re willing to accept. The product should match that level of openness – aligning the degree of workflow change with their comfort zone.
It’s worth considering that effective AI solutions not only fit into existing workflows – they also improve them by reducing friction, speeding up tasks, and making day-to-day work feel smoother.
What works: Align the degree of workflow disruption with your users’ readiness — their openness to change, digital maturity, and willingness to adopt new tools.
Common Pitfalls: AI that requires users to relearn everything from scratch, adds unnecessary complexity, or slows them down instead of making their work easier.
Example
Muvan.AI, which operates in real estate, focused on automating routine KPI report preparation – a repetitive but time-consuming task for real estate teams. By starting with a workflow-enhancing feature that users already needed, Muvan.AI established an immediate value proposition while building trust in its AI system.
As adoption grew, the company gradually expanded into more complex aspects of real estate automation, ensuring that AI integrated naturally into existing workflows rather than forcing users to adapt to an entirely new process.
For founders, this highlights how Enterprise AI adoption increases when the technology blends seamlessly into familiar workflows. Users shouldn’t feel like they are “using AI” – they should simply experience a better, more efficient way of working.
FOUNDER TAKEAWAY: A successful pattern for Enterprise AI adoption, particularly among users with lower digital readiness, begins by delivering early value through features that align with existing workflows. Addressing known needs with minimal disruption builds trust and lays the groundwork for deeper integration over time.
Not all AI problems require the same approach. Founders sometimes default to using the latest AI techniques and models, even when they aren’t the best fit for the problem at hand. Providing value in the decision-making processes can be achieved with various technologies and companies should choose the technology that meets their needs in their current stage.
In addition, while off-the-shelf LLMs often work well, there are cases where working with domain-specific or proprietary data calls for some level of model customization. Choosing the right model for the task can improve accuracy, shorten time-to-value, and over time, thoughtful customization can become a source of real differentiation.
What works: Selecting AI methods based on accuracy, reliability, and usability for the given problem.
Common Pitfalls: Using GenAI or LLMs in cases where structured, rule-based models perform better.
Example
Limitless.CNC, an AI-powered solution for programming manufacturing processes (CAM programming), illustrates how different AI challenges require different AI techniques, especially in physical, industrial contexts. Limitless recognized that traditional Generative AI models fall short when interacting with real-world constraints like machine dynamics, tool wear, and CAD/CAM-specific variations.
To address these challenges, Limitless built a custom AI base model fine-tuned with customer data and further refined through reinforcement learning (RL) in detailed digital simulations. This lets Limitless adapt continuously to shifting production realities and offer reliable, real-time suggestions right in CAM workflows. By automating repetitive tasks and embedding tailored AI agents, even junior programmers can deliver expert-level results going well beyond what template-based automation can achieve.
FOUNDER TAKEAWAY: AI should be designed for real-world usability, not just technological sophistication.
As AI systems become more complex, the involvement of domain experts is increasingly important. Many successful startups we work with bring in domain experts early to assist with solution strategy and product design, as well as data management tasks such as cleaning, preprocessing, and accurate tagging. This helps ensure the AI is aligned with real-world workflows and can validate that the solution meets the intended outcomes.
Given AI’s unpredictability, ongoing oversight from domain experts is crucial. Their insights guide the product roadmap and development process and help monitor AI’s performance throughout its lifecycle, ensuring it remains reliable and effective.
What works: Integrating domain expertise at every stage of development – whether for product strategy, data handling, model validation, or ensuring the AI aligns with real-world needs.
Common Pitfalls: Building a product for a specific industry without input from domain experts during development.
Example
The founders of Navina, an AI copilot for primary care, brought with them years of experience developing cutting-edge AI solutions for decision-makers that achieved high adoption by end-users. They understood that creating engagement in the medical field required a nuanced understanding of primary care. From day one, the solution design was led by in-house clinical leadership and a large team of in-house doctors who worked closely with design partners to develop a product that met the specific needs of family medicine in the US healthcare market. This medical expertise continues to impact all phases of product development, from complex clinical AI model building and training to data integrity.
FOUNDER TAKEAWAY: Deep domain expertise is key to creating strong product-market fit, and ensures the Enterprise-AI solution is practical, effective, and aligned with nuanced market needs.
