With a new fundraising of $250 million in SerieD, Decagon aims to make AI the pillar of tomorrow’s customer service. Jesse Zhang, its CEO and co-founder, details to JDN how its autonomous conversational concierges are already attracting businesses.
JDN. Decagon today announces a new round of $250 million in Series D, led by Coatue Management and Index Ventures, with the participation of historic investors such as Andreessen Horowitz and Accel. For what objectives?
Jesse Zhang. This funding round triples Decagon’s valuation in just six months, bringing it to $4.5 billion. After its first full year on the market, Decagon has signed more than 100 new corporate clients, in sectors such as travel, financial services, and retail. Among them, I can name Duolingo, Oura or Ritual Cosmetics. We are growing extremely fast, driven by high demand and a huge market. Instead of a simple support agent, our desire is to enable our clients to deploy real AI concierges, who can act proactively on behalf of a brand. These will be used for both retention and acquisition, able to proactively contact prospects but also recommend new products, etc. We will continue our international development, including in Europe.
Decagon launched in 2024 with the ambition to transform customer support with generative AI. How does your solution differentiate itself from competing platforms?
Decagon builds specialized AI agents around customer service tasks for large enterprises that handle high volumes of requests. Concretely, our agents can converse directly with end users by telephone, chat and email. The idea is not new but our difference is to have developed a solution integrating generative AI from the start, making it possible to obtain truly autonomous conversational agents available 24/7, unlike old rigid systems based on deterministic paths. Headquartered in San Francisco, with offices in NYC and London, Decagon now employs more than 200 people.
What steps are necessary to deploy Decagon agents in a company?
Traditionally, in our sector, it was necessary to call on technical profiles to be able to configure and program AI. Our bet lies in accessibility. We want non-technical business teams to be able to configure, test and improve agents themselves using our AOP format (Agent Operating Procedure, editor’s note)a configuration mode in natural language, which acts as a sort of operational guide executable by the agent. AOPs function similarly to standard operating procedures used by human agents. Simply put, you write structured instructions in natural language, and the AI interprets these instructions to manage complex situations.
Do you have any examples to illustrate how AOPs work?
Firstly, you must identify the priority use cases that the agent will have to deal with as well as the sources to connect, such as your helpdesk or CRM for example. It is then enough to write down all the steps that a human agent would follow in a classic procedure. The AOP can trigger actions, use your existing tools, or enable other workflows. When a company contacts us, we typically help them identify their priority use cases for customer support, and deploy the first AOPs together. Then their teams can add or modify them independently. This is the fundamental difference with older solutions which systematically required an engineer. However, certain aspects, such as connecting to specific systems, to manage reimbursements for example, sometimes require technical expertise. We also created Decagon University, a training program that gives employees the skills necessary to properly build and optimize their AI agents.
How far does the autonomy of these agents go today? Can they handle truly complex cases?
Today, the models are already powerful enough to handle almost all use cases. The most complex cases are those where the agent must perform around ten actions. For example, for the banks we work with, a common request concerns losing your bank card and ordering a new one. The AI must manage a whole process including understanding the problem, determining whether to order a new card, blocking the old one, confirming the address, identifying which card it is, checking if the others work, etc. AI already handles this type of situation very well.
However, is AI capable of replicating certain human behaviors that make an experience pleasant, such as being offered an upgrade for a birthday?
She can if you teach her, via AOPs. You can therefore give them instructions such as: “In certain cases, you have the possibility of waiving such fees, according to such and such criteria.” You can also configure your agents so that every such decision is subject to human validation. You can also set up a set of safeguards and rules for the AI to follow, such as not giving financial advice, not answering questions about other businesses, not saying anything vulgar, etc.
Do you really think that within three years the majority of customer support tasks will be handled by AI?
Our company has only been around for over two years and we have seen a massive improvement in AI capabilities. Within three years, this will undoubtedly be the case, especially if we observe the current rapid progress. Even today, the models are actually already very good at handling the majority of support cases, which are ultimately not that complex. The challenge is to teach AI how to do things well, and that’s what Decagon is all about. The vast majority of incoming support requests will be handled by AI. We want to enable employees carrying out these tasks to devote themselves to activities with higher added value, focused on creating relationships with customers or sales.
How much does your solution cost on average?
We charge either per resolution or per conversation. Prices vary with a fairly wide range, as clients have different levels of complexity. Typically, this works out to less than a dollar per conversation for our customers. Our goal is for them to achieve at least an ROI of 3 to 5 times their investment.
Jesse Zhang is the co-founder and CEO of Decagon, a platform that aims to transform business-customer relationships through the deployment of concierge agents powered by generative AI. Before founding Decagon in San Francisco in 2023, Jesse Zhang founded Lowkey, a start-up acquired by Niantic in 2021. He is also an investor in several companies, including Lovable and Cursor. He graduated from Harvard in computer science.
