Choco: Becoming AI-first

In March 2023, we decided to go AI-first. At this point we were several hundred people across US, UK and EU, billions of dollars in GMV, tens of thousands of users, and millions of transactions every month.

• By September 2024 100% of our new revenue came from AI.
• Our Sales costs are down 50% - automated by AI.
• Customer service cost is down 80% - automated by AI.
• Revenues are growing 3x faster than pre-AI.
• User value is up 5x.

What We Do:
Our AI technology automates the order processing workflow for food distributors. Instead of having a human operator manually process orders that come in via phone, SMS, WhatsApp, or email, Choco’s AI steps in to handle everything automatically.

How It Works:
When a customer (e.g., a restaurant) calls or messages, saying something like, “I need 5 tomatoes and 10 kg of beef,” our AI identifies exactly what they mean —such as understanding whether they want 5 kg of cherry tomatoes or 5 boxes of Roma tomatoes—and then enters this data directly into the distributor’s system.It can even answer the phone and it will be happy to consult you on your menu choices. We achieve this using our proprietary neural network combined with large language models (LLMs) for tasks like transcription and translation.

Benefits:
This automated approach not only saves time and reduces errors but also helps distributors grow their revenue by guiding sales and marketing strategies. Moreover, our predictive capabilities help cut down on food waste through improved demand forecasting.

Our Transformation:
We evolved from a Unicorn SaaS company into an AI-first organization with an AI-first team. This is how we did it:

Why AI-first is required to win

When a new technology provides value, is available, cheap, and reliable, we are responsible as leaders to assess this technology critically.

AI will make our products 100x better and automate most internal processes, leading to more revenue with higher cash efficiency. AI will create way better companies than we could have imagined even two years ago. Tech companies without AI at their core are obsolete and will be outcompeted by AI-first companies. Companies 100% committed to AI will thrive; all others will die.

When a technology is as impactful as AI, it is also our responsibility as leaders to ensure our team has the skills to build with that technology — without AI skills, your skillset will quickly become obsolete. So, we wanted to make Choco a place where people can learn the most relevant skill of this century: how to build with AI.

The biggest AI fallacy is to build AI for AI’s sake - AI window dressing is everywhere.  We wanted to use AI to get to our vision faster, solve real-life user problems, build an actual AI business model, and, even better, disrupt our own business model. 

Hence, we defined what “AI-first” means:

A) Only 50%+ revenue makes you AI-first.

Only one metric to truly be AI first: Most of your (new) revenue must come from AI.

Any product that makes 50%+ of your revenue becomes naturally the heart of your company. If that product is an AI product, then AI is at your core, and you become AI-first.

B) Using AI internally for productivity just makes you survive

AI in your internal processes to increase productivity is like having a corporate social media account. It doesn't make you a social media company, but it’s a requirement to be competitive. We could call this AI-enabled, but in the end, we also don't call companies “Internet-enabled”. AI for productivity is a must for survival.

This is our playbook on how we became an AI-first company, and how any founder, even non-technical ones, can make their company AI-first, too.

1. An AI-first company needs an AI-first Founder/CEO

Change starts with the leader. You can only expect 20% less of others than you expect of yourself. So the first step it to become knowledgeable about AI yourself. AI products are built, marketed, priced, sold, and scaled differently than non-AI products, there is a lot to learn. Every part of the organization will change, and as with everything, change is hard. 

Of course you will need a team that is generally open to AI. We have been lucky with great engineeringing, product and data leadership to jump on the train quickly.

• Read AI newsletters; it's the easiest way to stay up to date. There are many: The Neuron, Alpha Signal, and Rundown AI.
• Follow the most important AI thinkers on X; start with folks from foundational layer companies like OpenAI.
• There is a 100% chance you can use an LLM in your daily work. Working with AI has become second nature to you. Using ChatGPT as a sparring partner, to summarise long documents to synthesize large amounts of qualitative (user) feedback into the most essential points. Of course, I wrote this playbook with the help of OpenAI’s Canva.
• Build a tiny AI tool yourself. It can be as simple as a custom GPT summarising your weekly business performance report.

Essentially, make yourself informed and curious about AI. Once you are up to date, you can expect the same from your team, but not before.

Today, I fundamentally believe that AI will change economies, society, and the planet. This belief is helpful, if not necessary, to build an AI-first company. A clear vision of AI helps as well.

2. Create an environment of curiosity

Once the founder / CEO is AI-first, the team is next. 

Your job is to create the environment which enables your team to come up with game-changing AI ideas, and learn the required skills to execute them. Your team will be able to execute if you let them.

Building with AI might be the biggest change in company building since electricity. Many people's tasks will change completely, so there is a lot to learn. Naturally, some people will be scared and reluctant to work with AI. And it is hard to teach someone if they don't want to be taught. But someone curious doesn't need to be taught; they will learn it themselves. So create an environment of curiosity and the rest will take care of itself.

1. Allow & Encourage AI: After ChatGPT was released, I learned that some people were uncertain whether they could use AI at work. Officially allowing AI is the first step:

2. It doesn’t matter if the ideas you receive are good or bad - the only thing that matters is that people get curious

3. Recognition is the simplest tool on your journey to AI-first because it can even convince AI pessimists. I shared my personal "top 10" ideas with the team. People became motivated to come up with better ideas and started to experiment more. With more experiments, the ideas improved, we recognized these ideas, and our people became more skilled in using AI. 

4. Don't make big internal (or external) announcements: The lower your voice, the more curious people become.

5. Be ready to invest and pre-approve the budget: Becoming AI-first is the most important investment for any company to win. You want to encourage AI experiments, so make sure there is absolutely no friction. Even just asking for a budget is a hurdle that some won’t take, so just declare that people will get the necessary budget as long as it's for AI experiments. You don't need much anyway (think $20-200 pP).

6. Go 1:1 for massive budget requests: It's a good thing if people are bold with AI. If the requests become outrageous, speak 1:1. Don't introduce process. 

7. Encouragement all the way. Failed experiments are just as valuable to drive skills and curiosity. Make sure failures are public, for example share them as case studies on your AI slack channel.

8. Create room for conversation. Some people will naturally be more curious than others and seek to discuss their ideas. I asked one of our most AI-curious people to set up an open Slack channel, "GPT-wizards," and added some other curious people. He started to write his ideas there, and people chimed in. Only invite AI champions for the beginning so that engagement is high. Some of the first joiners were from brand design, marketing, product, engineering, and data science. Curiosity can come from anywhere. Soon, many more people joined.

9. It’s a movement. It's important not to open the channel yourself; the team has to do it. A grass-roots approach creates more excitement and curiosity— it feels like a movement rather than a top-down thing. But be active in the channel yourself and post your ideas, thoughts, and news (remember, you must be AI-first yourself). Our AI conversations were initially shallow, but our discussions became more technical and cutting-edge as we stayed curious and learned.

Little recognition for our best AI hackathon projects

3. Generate Skills

AI is new (at least on the application layer), and no one has much experience building with it. So, we are in a race to upskill our teams, integrate AI into our core, and become AI-first. 

Learning how the APIs of foundational models work, fine-tuning, prompt engineering, which model works best for each use case, and other skills takes some time, so you must invest in developing these skills in your team.

Hackathons are a great way to generate AI skills. It's hard to find time to upskill in the busy day-to-day. And it's boring and theoretical to do online courses. So, we ran three different week-long AI hackathons to ensure everyone had the time to experience building with the latest and greatest AI models. 

The first hackathon started the moment we had API access to OpenAI. We branded it “Battle of the Bots - GPT Hack 2023.” The whole company split into 2-8-person teams of their choice, working on one problem for the entire week with one goal: build something with AI.

The hackathon produced useful and not-so-useful ideas. We were among the first to work with agents, where engineers effectively tried to replace themselves by asking agents to build a feature. It worked, but the result was not replicable.

Our game plan for the "Battle of the Bots" Hackathon:

• One clear owner/organizer. (not the CEO / founders, it's a movement)

• Ask people to submit ideas before and to gather a team of 2-8 people.

• AI only

• Aligned with the company vision

• Launch event Monday morning explaining the goal: live demo by Thursday

• In-person only, it speeds things up and increases intensity.

• Cater breakfast, lunch, dinner

• Live demo (!) of the new products on Thursday night

• Everyone votes on the coolest projects; the best 3 get a fun price, ideally something that fosters team spirit (not Amazon vouchers)

• Give the hackathon a brand, and make merch that people will be proud to wear

• Only give this merch to participants (exclusivity drives curiosity)

• Recognize the winners in a big way, on your allhands, on the Slack team channel, etc. 

You don’t have to put these projects on the roadmap. The sole purpose of hackathons is to provide time for learning AI skills. Every time we did another hackathon, our goal was to have better quality AI demos than at the last the hackathon.

The results of the first hackathon could have been better, but by the third, our team had done real magic. I remember sitting on the jury of our third hackathon and thinking, "I want to put all of this on the roadmap." We will continue to hold three hackathons per year. AI is evolving fast, and we need to stay on the cutting edge.

Our first AI hackathon: Battle of the Bots - GPT Hack 2023

4. Bring AI into your day-to-day & reward AI-first people

Hackathons are a great way to build initial excitement. However, to sustain momentum, AI must become embedded in daily workflows across the organization. Here are small hacks that made all the difference:

Explain why AI is necessary for the individual. I told the team, "For every individual in this company, if you were ever to leave this company and had no AI skills, then your skillset would be obsolete. At our company, we want to create a place for you to learn relevant skills so you can succeed here and once you leave Choco. But it is your responsibility to be curious and learn. We can just give you all the room you need." People remembered that because it was personally affecting them.

• Adapt your 1:1s. Ask, "How are you working with AI? How do you think it can help us?" The first time people hear these words in a serious meeting, they will likely not have answers. But keep asking these questions. People will soon have answers. 
• Set simple targets, e.g., every engineer should use GitHub Copilot within two months. (A side effect of becoming AI-first was that our engineering productivity rose 120% since we encouraged copilot use.)
• Asking serious questions casually to the team at the water dispenser will help people understand that AI is important. "What do you think of GPT 4 mini?" "Have you seen the latest Gemini release?" " Did the voice latency get better?" " How do you use AI in your daily life?" It's a simple technique that does wonders. 
• Introduce questions on AI to every recruitment process: When the OpenAI API was new, we asked every PM and Engineering candidate if they had already experimented with it. We also asked everyone else, from Finance to Design, how they were using AI. If there were no real answers, we would reject the candidate. People have to bring AI curiosity. This contributed to our team’s understanding of the importance of AI.
• Continue to recognize AI-first thinkers.
• Be consistent and trust the process: The AI-first mindset will trickle down in the organization within weeks. Each individual's evolution will lead to the company's evolution. People will soon understand the criticality of AI, learn how to work with AI, and become literate in the most critical technologies. It will feel exciting, too.

Your whole team now understands the importance of AI and is building AI skills. You have created an exciting environment where you experiment with the latest technologies. This is every tech worker’s dream. At this point, people wanted to put AI-enabled features on the roadmap, which is when I rejected AI ideas for the first time. AI has to be aligned with our vision and, ideally, gets us there faster. We did not want to become defocused because of the AI euphoria we created. 

Don't use AI for AI's sake. Use it to reach your vision faster.

Keep your eyes and ears wide open. Keep doing what you are doing; the idea will come. You can't force it, so focus on creating the best environment. Your idea might come from hackathons, meetings, or having a drink with the team. For us, it happened on March 22nd, 2023, 8 days after OpenAI launched GPT-4: One of our PMs had a conversation on our kitchen floor and pitched an idea for how we could leverage AI technology to, in his own words, "10x our product." 

4. An AI speedboat is the fastest way to a POC

Two days after the idea, a talented engineer and I took a plane to Boston, where we sat in a customer’s office for a month to crank out the first POC. What followed was one month of night shifts and many sleepless nights. As a founder I continued doing my most essential 1:1s and team meetings, but used all my working time for the speedboat.

In retrospect, the speedboat approach was one of the best decisions we could have made as it focused us on the one thing that mattered: finding product market fit. This is what we have learned:

• Get a proof of concept asap. Don't aim for perfection; aim for validation. You might have to rewrite all the code at some point; it doesn’t matter. Get a helpful product out as soon as possible. Deploy it to a user as quickly as possible.
• "In-house" the team: Product market fit is not found - but forged. Sit in your user’s office. You will discover countless edge cases that would have taken months to discover otherwise. Our team would spend weeks in users’ offices, hearing feedback directly and building features or fixing bugs on the spot. It is essential to create a setup where you can move fast and respond to user feedback quickly, ideally within 24 hours from feedback to improvement shipped. Sometimes, we built new functionality on the spot after getting feedback, which made our first users extremely excited and loyal, as they experienced us taking their feedback seriously.
• Choice of team is decisive. There are people who are better at 0 to 1 and people who are better at 1 to 10. You need the 0 to 1 folks to move fast.
• Be part of the speedboat until POC is done: It’s like starting a new company; by definition, founders have 0 to 1 skills. It will also enable the speedboat team, and after you hand it over to people who will drive operational excellence long-term, it allows you to run an AI-first company. Once you are gone, the speedboat should keep reporting to you for a whole to ensure any roadblock is gone as soon as possible. 
• Complete independence. The speedboat should 100% focus on the AI product. There should be no service ownership over legacy code, no maintenance of existing services, no on-call for your existing product, etc. Also, make it clear to the rest of the company not to bother the speedboat with absolutely anything. 
• Grow the team only when necessary. More people make you slow. Keep it under four people with one clear owner.
• No process. Processes exist to produce consistent quality in a relatively predictable environment. This is the exact opposite of what you need now—no quarterly roadmap, no PRDs, no design reviews, etc. Just sit with the user and build what they need. Nothing else. We didn’t even work in sprints—those were too slow. Our product manager would create tickets, and the engineers grabbed them and shipped. 
• Let the speedboat team maintain its freedom as long as possible. Make it part of your organizational structure only once the product is ready for wider market adoption. 

Shipping an improvement on our way home from our user

It was very challenging to build this with only one engineer in the beginning; we should have started with two. Additionally, we should have added a data scientist from the beginning to avoid engineers covering up for them. But in the end, we shipped a POC and saved our users hundreds of thousands of dollars annually. It was not perfect, but it worked, so we continued. 

Launch, Fail, and Breakthrough: Learning is the most important feature

People take it for granted that AI should learn and improve from its mistakes. And while AI can do that, you must build this ability first.

Our POC did not have learning functionality, so our early users were clearly frustrated. However, user frustration is a blessing in disguise. If users get frustrated, then they really care about the problem you are solving. So, from that moment on, we focused only on one thing: Making it learn. 

Learning ability is the most essential feature of your AI product. 

To this day, most of our work is making it easy for users to “teach” the AI. It is not about the day-one experience; it is the rate of learning that makes an AI product successful. The interface for users to train the AI was the breakthrough moment for us, allowing users to “correct” the AI and customize it to their needs. Users truly loved the product from the moment they could teach the AI to behave precisely as they wanted it to. From that point onwards: all gas, no brakes. 

5. From POC to main revenue driver: Becoming AI-first

Meanwhile, sales and marketing started aligning their compensation structures, branding, and press strategies to support the new product. It's a good thing if you haven’t thought about go-to-market, monetization, economics, and marketing at this point because it means you have been focused on the only thing that matters: finding product-market-fit. 

How do you now get this little speedboat project to become your main revenue driver?

• Take the most AI-motivated seller. Naturally, someone from Sales will want to sell this product more than others. Wherever this seller is, make this the test team/market. Speak daily with your seller, be a sounding board, and enable them. Then, get at least five commercially successful showcases to show that your AI tool is not a one-hit-wonder (and to provide sufficient user feedback to the speedboat team).
• Daily AI success updates on Slack, weekly user stories on our all hands, and photos on our WhatsApp group. Ensure everyone knows, even if you have had a tiny success with your new AI product. You want the other sales teams to think: wow this baby sells!!
• Symbolism works. Once we had these success cases, we organized a global offsite to mark the start of our AI transition. When you hold a global offsite, everyone understands: this is important. We shared the learnings of the first users, explained the product in detail, presented our v1 AI sales guide, and held learning sessions on all these things. Our pioneers in our test market held the sales sessions, so it felt first-hand and authentic.
• Every team will work on AI soon: Some people who are not directly working on your new AI product might feel left out. Make it clear to the company, that soon we will have AI in every product and process across the company.
• Go all-in. Don’t roll it out until you roll it out. You never want to be in a limbo state of “trying things, so don’t allow any other market (or team) to “try” selling the product—no need to do the same learnings twice. One team is pioneering, and once you are confident, you go all-in or don’t do it at all. 
• Reflect it in the bonus. Even when all your sellers see the value of AI, they will still have to learn how to sell it, and build pipe for the new product. For a smooth transition we kept the bonus scheme the same, and added a 1.2x kickerfor selling AI, while the old product remained at 1x. Essentially we made it 20% easier for our sellers to reach 100% bonus when selling AI, but simultaneously they were at no disadvantage selling the old product. After 3 quarters everyone made use of this accelerator and was selling only AI. Then we removed the accelerator, and the incentive for the old product alltogether, having people only incentivized on AI. AI became 100% of our new revenue after this transition.
• It's a good thing if AI cannibalizes your existing product. Otherwise, someone else will. Don’t be scared, and tell the company not to be scared either.

​​

Our pioneering sales team & the AI speedboat doing a fireside chat for the rest of the company

In parallel, we beefed up the speedboat with dedicated product, engineering, and design leadership. We brought in additional backend and client engineers to add firepower to the rate of development. We also grew our already great data science team into a world-class group of machine learning and data engineers to research and develop the learning systems our customers expected. But we stayed close to the user. Our enlarged speedboat team onboarded the first 50 users themselves to quickly iterate on first-hand user feedback and teach sales how to do it.

We slowly integrated the AI product into our core organization, adding scalability, automated testing, security protocols, and design reviews one step at a time.

Today, we have rolled out Choco AI across the US, UK, and Europe. Our users love the product, and we process hundreds of thousands of transactions every month. Every single one of our product teams has incorporated AI into their roadmaps. Today, AI powers most of our internal tooling, processes, and infrastructure. All this goes to the credit of our Choco AI and our sales team who figured out how to scale an AI product from 0 to market leadership.

After 1.5 years, AI is indeed 10x-ing our product. It just produces more value for our users than pre-AI. Our sales, customer success, and marketing teams saw the deal win rate increase and customer happiness, engagement, and retention increase. We grow 3x faster than before.

Final Thoughts: AI-First Is the only way forward

We don’t know what companies will, or rather ‘should’, look like three years from now. But we do know that AI-first companies will lead and that AI is enhancing nearly weekly. We have to stay cutting edge.

The single best thing we can do to achieve that: Building a skilled and curious AI-first team. The team will figure out the rest. Teams who can learn as fast as AI is evolving, will win. 

Create curiosity, build AI skills. Don’t use AI for AI’s sake—use AI to reach your vision faster.