What We’ve Learned From Building AI-Powered Products at Gong
Introduction
Over the past few years, artificial intelligence has made tremendous strides in its capabilities, offering companies the potential to revolutionize industries. But integrating AI into commercial products requires a unique approach, as the models are computationally large and largely black-box in nature, and research projects have a high likelihood of failure.
Dr. Omri Allouche, head of the research department at Gong, shares five lessons the company and his team have learned in building successful products based on AI.
- Focus on creating customer value, not on technology
As technology advances, it’s easy to become enamored with the latest and greatest AI models that can produce spectacular images from sentences, transcribe conversations in multiple languages, and even translate them. While these advances are certainly a blessing, it is important to remember that the ultimate goal of AI is to create value for the customer.
Sun Tzu wrote in his seminal book The Art of War that a skilled commander knows how to win a battle without going to war. This is equally true with AI models — they require a lot of effort to build and may not even solve the right problem or be necessary. The first step of AI research is to understand if using AI is a must.
One way to test this and make sure we’re focusing on the right thing is the “Wizard of Oz” method, in which we meet with customers and show them the results of a “smart AI model” we built. In reality, there’s no model — instead, team members are working behind the scenes to create an output that looks like one that a model would generate. For example, we can describe to them a model summarizing the role of a contact in a sales opportunity. We’ll ask where in our product they would want to see the model output, and understand what they think constitute a useful summary of a person, using real examples that were actually manually labeled. This way, we can quickly understand if the solution has value for the customer, without having to build an actual AI model. If not, AI might not be the necessary approach.
When building AI models, we always start by mapping the user’s pains and challenges, progress through rapid research and development iterations, and get feedback from the user as early and as often as possible. Throughout the process, we remind ourselves not to get too theoretical — there’s no substitute for the user’s feedback on the results of a model on their data.
Contrary to popular belief, it’s actually best to start with a complex model. Neural networks with billions of neurons that have been pre-trained can be adapted to a specific problem with the help of a small number of labeled examples, and can achieve impressive performance in a short development time. This is known as transfer learning and can prove value to the customer. Once the value has been established, then the costs and complexity of the model can be reduced.
2. Appreciate the power of data
Data has the power to unlock insights that can be invaluable to businesses. At Gong, we recognized this from the very beginning and therefore opted to publish a weekly post analyzing millions of sales calls. This gave sales teams the ability, for the first time, to verify whether so-called best practices in the field actually proved true. For example, the old adage of “Listen more than you speak” did end up being true in the world of sales. These insights established Gong as an opinion leader in revenue intelligence. The Gong blog writers were named to LinkedIn’s list of the 10 Top Voices in sales, and Harvard Business School even created a case on Gong to demonstrate its power. Companies, eager to learn whether the models held true for them, purchased the Gong platform to find out. Indeed, the Gong examples showed just how powerful data can be, when actions are taken according to the insights it delivers.
3. Find your oasis
Research projects are often accompanied by the fear of failure. In artificial intelligence research, this fear of failure can be rooted in issues such as missing, partial or inaccurate data, incorrect labeling, errors in the information processing process, difficulties in the learning process, and more. To combat this uncertainty, we have developed the “Oasis” method, which seeks to reduce uncertainty as much as possible.
The Oasis method starts by identifying a part of the world where conditions are optimal for success. For example, when developing our Action Items extraction system, we started with accurate manual transcriptions, using only sentences that actually had an Action item. This allowed us to quickly understand whether the problem could be solved at all and established a sense of commitment among the development and product teams.
Once this task was successful, we began to relax the lenient constraints and filters and understand where to invest our efforts. For action items, for example, we’ll check the effect of transcription errors on performance, and adapt the model accordingly. This allowed us to improve performance and make the feature accessible to users.
With the Oasis method, we can reduce uncertainty and quickly determine whether a problem can be solved, so we can decide how to invest our resources and time.
4. Get ready for the real world
It’s often said that an ML model is like a new car: once you drive it off the lot, it’s already worth 20% less. That’s because ML models are trained on only subsets of the data, which don’t adequately represent the real world. This can lead to bugs and malfunctions in the model’s processing of information. To address this, it’s important to monitor the model’s predictions in production to identify anomalies, trends, and subgroups with suspicious characteristics.
At Gong, we recently trained a neural network to detect languages and accents in speech segments. The model performed excellently in the lab, but when we ran it on actual sales calls, we noticed that a large percentage of short conversations (less than 5 minutes long) appeared to be in French. After listening to them, we found that hold music was being detected as French speech. To fix this, we built a preliminary model to detect music in conversations and make sure the language recognition model only runs on speech.
Monitoring is also essential for detecting changes in the data distribution over time. For example, Gong has seen rapid growth in sales, which has resulted in English-speaking customers with a variety of accents, such as Scottish, Australian, and Indian. To maintain a high quality product, we’re constantly monitoring for decreases in performance and making adjustments.
5. Dream big
Even though AI is prone to hype, the field itself is often shrouded in mystery, making it difficult for product managers to determine which technologies are ready to be implemented in a mature product. It’s up to research teams to help advise on this.
For example, Gong developed a tool for writing emails to customers, which allows a salesperson to view any previous correspondence, action items from the last conversation, and so on before sending a new email. As a “smarter” product, the Gong Product team asked for the development of email templates and a smart auto-complete feature that would complete a few words ahead of time, saving a seller time and improving accuracy. Not only is the value to the customer clear, but these additions are relatively simple from a research perspective.
The research team went a step further and created a presentation about the current state and potential of generative AI, to inspire product people to dream, in which they presented emails that were automatically generated by a neural network given various instructions. For example, we asked the model to write a short and humorous sales email about the analogy between presenting at a conference and working in a circus:
Hi there!
If you’ve ever given a presentation at a conference, you know that it can be a bit like working in a circus. There’s the juggling act of trying to keep everyone’s attention, the tightrope walk of staying on topic, and the circus-like atmosphere of being in a big room full of people.
But just like the circus, giving a presentation can be a lot of fun. It’s a chance to show off your skills and knowledge, and to connect with other people who are passionate about the same things you are.
So if you’re feeling like a circus performer, remember that it’s all part of the fun. And if you’re looking for some tips on how to make your presentation a success, we’ve got you covered.
Thanks for reading, and we hope to see you at a conference soon!
With machines able to produce content like this on cue, there’s no doubt: we’re at a transformative time, with people trusting AI more with major parts of their work, and AI quickly rising up to the challenge.
If you also have a passion for data, AI, and making customers happy, we’d love to hear from you!
About Gong
Gong unlocks reality to help people and companies reach their full potential. The Gong Reality Platform™ autonomously empowers customer-facing teams to take advantage of their most valuable assets — customer interactions, which the Gong platform captures and analyzes. Gong then delivers insights at scale, enabling revenue and go-to-market teams to determine the best actions for repeatable winning outcomes. More than 3,000 innovative companies like Morningstar Inc., Paychex, LinkedIn, Shopify, Slack, SproutSocial, Twilio, and Zillow trust Gong to power their business reality. For more information, please visit www.gong.io.
