Building an AI-first company: lessons from scaling from a small start-up to a Machine Learning powerhouse

In the rapidly evolving landscape of e-commerce and grocery delivery, Picnic has emerged as a standout success story. As the world’s fastest growing online supermarket with a unique business model and highly engaged customer base, Picnic’s journey from a small startup to an industry leader offers valuable insights into the transformative power of artificial intelligence and machine learning.

When Picnic started in 2015, most companies were still in the midst of their digital and mobile transformation. In our case, we were a fresh company that did not need to make any transformation. So we started digital-first, cloud-native, and mobile-only. That meant efficient scaling, and gathering world-class analytics. With that, our CTO Daniel Gebler already saw the big opportunity to make Picnic into (probably the first?) AI-first company. This wasn’t just a technological choice but a fundamental business strategy that would shape every aspect of the organization. In an era when AI was still viewed with skepticism by many traditional retailers, this forward-thinking approach positioned Picnic at the cutting edge of innovation.

So what did this really mean for us? What were the first steps to prepare us for the rising relevancy of AI? How could we make some meaningful progress without having year-long broad and deep training data? How could we prepare us for an upcoming breakthrough such as ChatGPT & LLMs?

This blog post explores the key lessons learned during Picnic’s transformation from an ambitious startup to a machine learning powerhouse. We’ll examine how our AI journey evolved from basic forecasting models that formed the backbone of our supply chain to highly sophisticated natural language processing systems enhancing customer interactions, advanced recommendation engines personalizing the shopping experience, and cutting-edge vision models powering our automated warehouses.

Lesson 1: Start with Business-Critical Problems (Forecasting)

When Picnic began its journey a decade ago, the company faced a fundamental challenge that would make or break its business model: how to accurately predict customer demand in a just-in-time supply chain selling perishable products. Unlike traditional retailers with physical stores and buffer inventory, Picnic’s app-only supermarket model required exceptional forecasting precision. Overestimate demand, and food would be wasted; underestimate it, and customers would face disappointment.

This challenge became the perfect starting point for Picnic’s AI journey, embodying our first critical lesson: start by applying machine learning to your most business-critical problems.

The Evolution of Forecasting at Picnic

Our forecasting capabilities didn’t emerge fully formed but evolved through continuous iteration and learning. In the early days, we relied on gradient-boosting trees for demand prediction. These models showed robust performance across various forecasting scenarios and allowed us to incorporate different types of features, from historical demand patterns to seasonal variations.

However, the limitations of this approach became starkly apparent during the COVID-19 pandemic. When everyone suddenly started buying toilet paper, our models couldn’t extrapolate to values they hadn’t seen during training. Customer behavior was changing radically, and we couldn’t accurately predict demand using our existing systems.

This crisis moment catalyzed a significant evolution in our approach. We began exploring more sophisticated models, first investigating Long Short-Term Memory Networks (LSTMs). While these offered improvements in handling sequential data, they didn’t provide the breakthrough we needed for our complex forecasting challenges.

Our exploration continued with DeepAR, but we discovered that its architecture couldn’t adequately consider the wide variety of inputs present in multi-horizon forecasting, particularly the mix of static and time-varying variables that characterized our business environment.

The Temporal Fusion Transformer: A Breakthrough

The real breakthrough came with our implementation of the Temporal Fusion Transformer (TFT), a transformer-based neural network specifically designed for time-series forecasting. This sophisticated architecture addressed the limitations of our previous approaches and provided several key advantages:

1. Feature Flexibility: TFT supports both temporal and static features, allowing us to incorporate data known only up to the present (like historical demand), temporal data known into the future (like weather forecasts or confirmed customer orders), and static variables (like city region or delivery period).
2. Intelligent Feature Selection: The model’s variable selection mechanism determines which variables are relevant at specific moments in time, enhancing its ability to generalize by focusing learning capacity on the most significant features rather than overfitting to irrelevant ones.
3. Multi-scale Temporal Processing: TFT processes both short-term interactions (daily demand fluctuations, special events) and long-term dependencies (seasonal trends, national holidays, salary patterns) through separate mechanisms optimized for each time scale.
4. Uncertainty Quantification: Instead of single-point forecasts, TFT predicts ranges of target values, reflecting the uncertainty at each timestamp and enabling more robust planning.

The technical implementation of TFT involved several sophisticated components, including Gated Residual Networks for adaptable transformations, Static Covariate Encoders for integrating metadata, Variable Selection Networks for feature importance, and specialized encoder-decoder architectures for processing sequential information.

Impact on Business Operations

The evolution of our forecasting capabilities has had profound effects on Picnic’s operations. With more accurate demand predictions, we’ve significantly reduced food waste while maintaining high product availability. Our supply chain efficiency has improved dramatically, allowing us to optimize inventory levels, reduce costs, and enhance customer satisfaction.

Moreover, the explainability features of our advanced models have provided valuable business insights, helping us understand demand patterns, identify emerging trends, and make more informed strategic decisions.

Opportunity: A self-optimising system through continuous feedback loops

A fundamental shift from rule-based, or heuristic algorithms to machine learning also comes with another benefit. As we fulfil deliveries every single day, this feeds our data warehouse with new data points. Combined with automated training of our ML models, this leads to automatic self-improvement.

For example: as Picnic grows, our density increases. That also means we have a better understanding of your neighbourhood, regional preferences, etc. This in turn allows us to better predict driving & drop time, which leads to more efficient deliveries, which lead to happier customers, which leads to more data, which leads to better predictions, …. A fly wheel!

The Lesson: Prioritize Core Business Problems

The journey of our forecasting systems exemplifies a crucial lesson for any company embarking on an AI transformation: begin by applying machine learning to your most business-critical problems. For Picnic, accurate demand forecasting was existential — it directly impacted our ability to deliver on our core promise to customers while maintaining a sustainable business model.

By focusing our initial AI investments on this fundamental challenge, we:

1. Ensured that our technical innovations directly supported business objectives
2. Created a clear path to demonstrating return on investment
3. Built organizational confidence in the value of machine learning
4. Established a foundation of data infrastructure and expertise that could be leveraged for future initiatives

Many companies make the mistake of starting their AI journey with flashy, but peripheral applications. We recognized early on that the real power of machine learning comes from addressing the core challenges that drive your business economics.

This lesson — prioritizing business-critical problems — set the stage for Picnic’s broader AI transformation and continues to guide our approach to innovation today.

Lesson 2: Enhance Customer Experience with NLP

As we established a solid foundation for our ML ambitions and demonstrated the first strong proof of value by means of our forecasting models that significantly optimized our supply chain, we turned our attention to another critical aspect of our business: customer experience. Natural Language Processing (NLP) emerged as a powerful tool to enhance how customers interact with our platform, embodying our second key lesson: utilize AI to directly enhance the customer experience.

The Role of NLP in Customer Interactions

For an app-only supermarket like Picnic, every digital interaction represents a crucial touchpoint with customers. Early on, we recognized that natural language understanding could transform these interactions from transactional to truly personalized and intuitive.

Our NLP journey began with fundamental applications focused on improving our customer service. By leveraging Machine Learning, we were able to get a much better understanding of our customer needs: prioritising urgent requests, or understanding when something is a really complex issue.

Evolving NLP Capabilities

As NLP technologies advanced, so did our applications. We progressively implemented more sophisticated capabilities:

1. Sentiment Analysis: By analyzing customer reviews and feedback, we gained deeper insights into satisfaction drivers and pain points, allowing for more targeted improvements.
2. Query Understanding: Our systems evolved to handle complex queries, recognize synonyms.
3. Multi-language support: By training our models across all languages, we’re both solving the cold-start problem for new markets as well as supporting many languages of our customers.

Technical Implementation Challenges

Implementing effective NLP systems presented unique challenges compared to our forecasting models. While forecasting dealt primarily with structured numerical data, NLP required processing unstructured text with all its ambiguities and contextual nuances.

One significant challenge was handling the domain-specific language of grocery shopping. Generic NLP models often struggled with product names, food-specific terminology, and the shorthand customers use when shopping for groceries.

Impact on Customer Experience

The implementation of advanced NLP capabilities transformed numerous aspects of the Picnic customer experience:

1. Smoother Customer Service: Automated responses to common questions reduced wait times, while more complex issues were enriched with contextual information before being routed to human agents.
2. Reduced Search Time: Customers found products faster, with fewer failed searches and less need to browse through irrelevant items leading to bigger baskets.
3. Personalized Recommendations: By understanding the nuances of customer queries, we could provide more relevant product suggestions tailored to individual preferences.
4. Improved Product Discovery: Customers discovered new products aligned with their preferences through more intelligent connections between search terms and our catalog.

The Lesson: AI Should Directly Improve Customer Touchpoints

Our experience with NLP underscores a crucial lesson for AI implementation: technology should directly enhance how customers experience your product or service. While back-end optimizations like forecasting are essential, customer-facing AI applications create tangible value that users immediately recognize and appreciate.

By focusing our NLP efforts on improving specific customer touchpoints, we:

1. Created visible differentiation from competitors
2. Generated immediate feedback loops for continuous improvement
3. Built customer loyalty through consistently better experiences
4. Gathered valuable data on customer preferences and behaviors

The most powerful AI applications are often those that customers don’t notice but where the experience is distinctively improved. Our AI systems work behind the scenes to make interactions more natural and intuitive, removing friction points that customers might not even realize existed.

This lesson — prioritizing customer experience enhancements — has guided our approach to NLP development and continues to inform how we evaluate and implement new AI capabilities across our platform.

Lesson 3: From Rules to Learning (Recommendation Systems)

As Picnic matured, we faced a new challenge: how to help customers navigate our growing product catalog efficiently while encouraging discovery of new items. This challenge led us to our third key lesson: embrace the transition from rule-based systems to learning-based approaches.

The Evolution of Recommendations at Picnic

Our recommendation journey began, as many companies do, with simple rule-based targeting. Early implementations relied on straightforward heuristics: customers who bought product A might be interested in complementary product B; those who regularly purchased certain categories would see more items from those categories; and seasonal items would be promoted during relevant time periods.

While these rule-based approaches provided value, they had clear limitations. Rules had to be manually created and maintained, couldn’t easily adapt to changing customer preferences, and failed to capture the nuanced patterns in shopping behavior that might not be obvious to human observers.

We quickly realized that static rules couldn’t capture the complexity of customer preferences or scale with our growing business. The shift to learning-based systems wasn’t just a technical upgrade — it represented a fundamental change in how we approached personalization.

This realization marked the beginning of a significant transformation in our approach to recommendations — a journey from simple rules to collaborative filtering and ultimately to sophisticated neural network architectures.

Collaborative Filtering: Our First Learning-Based Approach

Our first major step beyond rule-based systems was implementing collaborative filtering for personalized recipe recommendations. This approach leveraged user-item interaction data to shape a matrix for similarity calculations and generate recommendations.

The collaborative filtering system proved surprisingly effective for customers with extensive interaction histories. By identifying patterns in how customers engaged with recipes, we could suggest new options that aligned with their preferences without relying on manually crafted rules.

However, as we explored further, several limitations became apparent:

1. Scalability Challenges: The approach struggled to handle our growing dataset efficiently.
2. Cold Start Problem: New recipes or customers with limited history received poor recommendations.
3. Diversity Issues: Recommendations tended to become repetitive over time, reducing discovery.
4. Limited Feature Engineering: Relying solely on interaction data restricted our ability to incorporate valuable information about recipes and customer preferences.

These challenges pushed us to explore more sophisticated approaches that could overcome these limitations while maintaining the adaptability of learning-based systems.

The Two-Tower Architecture: A Sophisticated Evolution

The breakthrough came with our implementation of a two-tower neural network architecture for recommendations. This transition from collaborative filtering to a more sophisticated neural network approach represented a significant leap forward in our recommendation capabilities.

The two-tower architecture fundamentally reframed how we approached the recommendation problem. Rather than working with an interaction matrix, we shifted to treating interactions as independent data points enriched with contextual information. This tabular approach allowed us to incorporate a much richer set of features:

1. Customer Preferences: Detailed information about dietary preferences, cooking time preferences, and buying patterns.
2. Recipe Characteristics: Data on recipe types, cuisines, primary proteins, cooking times, and nutritional profiles.
3. Contextual Factors: Time of interaction, seasonality, recent searches, and app interactions.

The two-tower architecture processes these features through separate neural networks — one for query (customer) features and one for candidate (recipe) features — before combining them with a lightweight scoring function. This separation allows for efficient processing of both customer and recipe information while capturing the complex interactions between them.

Technical Implementation Details

The implementation of our two-tower recommendation system involved several sophisticated components:

1. Problem Reformulation: We shifted from an unsupervised matrix-based approach to a more flexible framework treating interactions as independent data points.
2. Feature Engineering: We developed comprehensive feature sets for customers, recipes, and contextual factors.
3. Neural Network Architecture: We designed and optimized the two-tower structure with appropriate layer configurations for each tower.
4. Training Strategy: We implemented effective sampling techniques to handle the imbalance between positive interactions and the vast space of potential recommendations.
5. Evaluation Framework: We developed metrics that balanced accuracy, diversity, novelty, and relevance.

Impact on Customer Experience

The evolution from rule-based targeting to sophisticated learning-based recommendation systems transformed the Picnic shopping experience:

1. Personalization: Customers received genuinely personalized recommendations that adapted to their changing preferences over time.
2. Discovery: The system successfully introduced customers to new products and recipes they might not have found otherwise.
3. Efficiency: Shopping became more efficient as relevant items were surfaced proactively rather than requiring extensive browsing.
4. Engagement: Customer engagement with recommendations increased significantly, driving both satisfaction and basket size.

The Lesson: Embrace the Transition from Rules to Learning

Our journey with recommendation systems embodies a crucial lesson for AI implementation: embrace the transition from rule-based systems to learning-based approaches. While rules provide a starting point, the real power of AI comes from systems that can learn complex patterns from data and adapt over time.

By evolving our approach from static rules to sophisticated learning systems, we:

1. Created more personalized experiences that continuously improved
2. Reduced the maintenance burden of manually updating rules
3. Discovered non-obvious patterns that human observers might miss
4. Built systems that could scale with our growing business and data

The shift from rules to learning represents one of the most fundamental transformations in how businesses operate. Rules encode what we already know; learning systems discover what we don’t yet know.

This lesson — embracing learning-based approaches — continues to guide how we develop and enhance AI systems across Picnic, pushing us to question where static rules might be limiting our potential and where learning systems could unlock new value.

Lesson 4: Building an AI-First Culture

While the technical evolution of Picnic’s AI capabilities — from forecasting to NLP to recommendations to vision — is impressive, perhaps the most important lesson we’ve learned is about organizational culture. This brings us to our fifth and final key lesson: AI-first requires both technical and cultural foundations.

The Vision

Ten years ago, when we articulated the vision for Picnic as an AI-first company, this wasn’t just making a technical decision but a cultural one. At a time when many organizations viewed AI as a specialized function isolated from core business operations, Gebler recognized that true transformation would require embedding AI thinking throughout the entire organization.

From day one, we didn’t want AI to be a separate department or a bolt-on capability, we wanted it to be part of Picnic’s DNA — a fundamental way of thinking about problems and opportunities that would influence every aspect of how we operate.

This vision required more than just hiring data scientists or investing in technology infrastructure. It demanded creating a culture where AI was understood, valued, and leveraged by employees across all functions and levels.

Building the Cultural Foundation

Creating an AI-first culture at Picnic involved several deliberate strategies:

1. Cross-Functional Literacy: We invested in building basic AI literacy across all departments, ensuring that team members from operations to marketing understood the fundamentals of machine learning and its potential applications.
2. Collaborative Problem Definition: Rather than having data scientists work in isolation, we established processes for collaborative problem definition, bringing together domain experts and technical specialists to identify the most valuable opportunities for AI application.
3. Shared Metrics and Goals: We aligned incentives by establishing shared metrics that measured both technical performance and business impact, ensuring that AI initiatives were evaluated holistically.
4. Democratized Access to Data: We built infrastructure that made relevant data accessible to employees across the organization, empowering them to explore patterns and generate insights.
5. Celebration of Learning: We created a culture that celebrated learning and experimentation, recognizing that not all AI initiatives would succeed immediately but that each attempt generated valuable knowledge.

Organizational Structure Supporting AI Innovation

Beyond cultural elements, we designed our organizational structure to support AI innovation:

1. Centralised AI excellence: Rather than having small pockets of knowledge across teams, we built a central team that powers use-cases across Picnic. This increases our focus on foundational models, a lean stack and cross-pollination.
2. Clear Ownership and Accountability: We defined clear ownership for AI systems, ensuring that both technical performance and business outcomes had accountable owners.
3. Investment in Infrastructure: We prioritized building robust data infrastructure and MLOps capabilities, recognizing that these foundational elements would enable faster innovation and more reliable deployment.
4. Talent Development Pathways: We created career development pathways that allowed technical specialists to grow while remaining close to their areas of expertise, rather than forcing them into management roles.

Overcoming Cultural Challenges

Building an AI-first culture wasn’t without challenges. We encountered and overcame several common obstacles:

1. Unrealistic Expectations: Early enthusiasm sometimes led to unrealistic expectations about what AI could achieve. We managed this by educating stakeholders about the capabilities and limitations of machine learning, and by setting appropriate timelines for results. Start small and iterate.
2. Balancing Exploration and Exploitation: We had to find the right balance between exploring new AI capabilities and exploiting proven applications. We addressed this through portfolio management approaches that allocated resources across different time horizons.

Impact on Innovation and Growth

The AI-first culture we’ve built has had profound effects on Picnic’s ability to innovate and grow:

1. Accelerated Innovation: By embedding AI thinking throughout the organization, we’ve identified and implemented valuable applications more quickly than others.
2. Improved Decision-Making: Data-driven decision-making has become the norm, leading to better strategic choices and more efficient operations.
3. Talent Attraction and Retention: Our reputation as an AI-first company has helped us attract and retain top technical talent who want to work on meaningful problems at scale.
4. Organizational Agility: The combination of technical capabilities and cultural readiness has made us more adaptable to changing market conditions and emerging opportunities.

The Lesson: AI-First Requires Cultural Transformation

Our experience building Picnic’s AI-first culture embodies a crucial lesson: successful AI implementation requires both technical and cultural foundations. While algorithms and infrastructure are essential, the human and organizational elements ultimately determine whether AI creates sustainable competitive advantage.

By investing in cultural transformation alongside technical capabilities, we:

1. Ensured that AI initiatives addressed the most valuable business problems
2. Accelerated adoption of AI-driven insights and recommendations
3. Built organizational resilience to navigate the inevitable challenges of AI implementation
4. Created a continuous cycle where initial successes fueled further innovation

The technology of AI is powerful, but it’s the human systems around that technology that determine its impact. Building an AI-first culture isn’t just about what tools you use — it’s about how you think, collaborate, and make decisions as an organization.

This lesson — prioritizing cultural transformation — has been perhaps the most important factor in Picnic’s journey from an ambitious startup to a machine learning powerhouse, and it continues to guide our approach to innovation today.

Looking Forward

As we look to the future, Picnic remains committed to our identity as an AI-first company. The technological landscape continues to evolve rapidly, with new capabilities emerging that will undoubtedly create opportunities for further innovation and transformation.

We’re particularly excited about several frontiers:

1. Multimodal AI: Systems that can seamlessly integrate understanding across text, images, and other data types promise to unlock new applications across our operations. The convergence of vision and language models is already transforming how we approach quality control and customer interactions.
2. Reinforcement Learning: As these techniques mature, they offer potential for more sophisticated optimization of complex operational decisions. We’re exploring applications in dynamic routing, inventory management, and personalized pricing strategies.
3. Causal Inference: Moving beyond correlation to better understand causation will enable more robust decision-making and intervention design. This frontier is particularly promising for understanding customer behavior and optimizing marketing strategies.
4. Foundation Models: The rise of large foundation models presents opportunities to consolidate and enhance many of our existing AI capabilities while enabling entirely new applications we haven’t yet imagined.
5. AI on the Edge: The trend towards performant smaller LLMs/VLMs already signal the possibility of running sophisticated AI applications on the edge, e.g. in our warehouses (optimised for latency), or even customer’s devices (optimised for privacy).

However, as we explore these new frontiers, we’ll remain grounded in the lessons of our first decade. We’ll continue to prioritize business-critical problems, enhance customer experiences, embrace learning-based approaches, adopt emerging technologies strategically, and invest in our AI-first culture.

Final Thoughts

Our vision of an AI-first company has proven prescient. As AI has moved from the periphery to the center of business strategy across industries, Picnic’s early commitment has positioned us at the forefront of this transformation.

Yet the most important insight from our journey may be that becoming an AI-first company is not a destination but a continuous process of learning, adaptation, and growth. The specific technologies will evolve, but the fundamental approach — applying intelligence systematically to solve meaningful problems — remains constant.

For organizations at earlier stages of their AI journey, we hope these lessons provide both inspiration and practical guidance. The path from startup to machine learning powerhouse is neither short nor straight. However, with strategic focus, technical excellence, and cultural commitment, it’s a journey that can transform not just how a company operates, but what it can achieve for its customers, employees, and stakeholders.

As Picnic enters our second decade as an AI-first company, we remain as excited about the possibilities ahead as we were when we began. The revolution in artificial intelligence is still in its early stages, and we’re honored to be part of writing its next chapters.

Building an AI-first company: lessons from scaling from a small start-up to a Machine Learning… was originally published in Picnic Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.