Decagon. The AI concierge for every customer.

It is 7:14 on a Tuesday morning and a Hertz customer somewhere outside Phoenix is asking why her rental was charged twice. She is not on hold. She is not being escalated. The agent on the other end of the chat has read her booking, pulled the duplicate charge, processed the refund, and recommended she check her credit card app in roughly the time it took to read this sentence. The agent is not a person. The agent is Decagon.

For decades the customer-support industry sold patience as a feature. Hold music. Ticket numbers. The phrase "please bear with us while we transfer your call," which is corporate English for "we have not solved your problem and we are not sure who has." Decagon, which raised $250 million in January at a valuation that tripled in six months, is selling something else: an end to the ticket as a unit of suffering.

01 The problem they saw

Customer support, as a business problem, is interesting only because it is so universally hated. Every consumer brand has it. Almost none of them are proud of it. The standard playbook involves a 200-page standard operating procedure, a contact center in a low-cost country, and a metric called Average Handle Time that quietly rewards rushing customers off the phone. The result is the experience everyone reading this has had: the loop where a chatbot asks you to rephrase, then routes you to a person who asks you to repeat what you just typed.

Jesse Zhang and Ashwin Sreenivas, the two founders, both came out of companies they had sold - Zhang's no-code analytics startup Lowkey went to Niantic, Sreenivas' computer-vision company Helia went to Scale AI - and started Decagon in 2023 with a thesis the rest of the AI industry was still warming up to: large language models were finally good enough to actually resolve a customer's issue, not just route it.

"Good enough to resolve" turns out to be a very specific technical bar. It requires the model to read the conversation, find the relevant policy, take an action in another system, write the customer back in the brand's voice, and remember what it did so the next person picking up the thread isn't starting over. Most chatbots do step one. Decagon's agents do all five.

02 The founders' bet

In 2023 the conventional wisdom was that AI customer support meant a thin wrapper over a foundation model. Decagon disagreed, loudly and in code. Their bet was that the interesting part of the problem was not the model. The model was a commodity. The interesting part was the layer above the model: how an enterprise tells the agent what to do, in language the legal team can read and the engineering team can audit.

They called that layer Agent Operating Procedures. AOPs are written in natural language, but they behave like code. A support operations manager - the same person who used to maintain the 200-page SOP - can edit them without filing an engineering ticket. The result is one of the more counterintuitive features in modern AI: a system where the non-engineers are the ones writing the logic, and the engineers have promoted themselves to platform builders.

03 The product, briefly described

There are four pieces. The AI Concierge is the customer-facing agent itself, multichannel by default. Agent Operating Procedures are the configuration layer that everyone writes about because it is the part competitors keep trying and failing to copy. Agent Assist is the version that sits next to a human, drafting and recommending instead of replacing. Voice Agents are the newest piece - real-time speech, branded caller IDs, designed to plug into whatever telephony stack a Fortune 500 has already standardized on.

In the Spring 2026 product release, the company added Proactive Agents: outbound conversations seeded with user memory, so the AI calls the customer rather than waiting for the customer to call. The cynical reading is that Decagon now does outbound sales calls. The accurate reading is that the same plumbing handles the call your bank should have made you three days ago to ask whether the unusual charge was really yours.

04 The proof, in money and logos

The most boring thing about Decagon is also the most damning to its competitors: it kept signing customers. Hertz. Chime. Notion. Duolingo. Rippling. Bilt. Oura. Affirm. Eventbrite. Substack. The list reads like the contents of a young professional's phone home screen. F100 airlines, banks and telecoms quietly sit alongside the consumer brands, less likely to be named in a blog post and more likely to be the actual reason the revenue chart bends upward.

Investors followed customers, which is how it is supposed to work and almost never does. Accel and Andreessen Horowitz anchored the early rounds. Bain Capital Ventures, BOND, Forerunner, Ribbit and Avra came in next. Coatue and Index led the Series D, with Elad Gil and a long bench of returning funds writing larger checks than they had six months prior. Total raised stands at $481 million across six rounds, which would be eye-watering if the revenue numbers were not, apparently, also moving.

05 The mission, said plainly

If you ask Zhang what Decagon is for, the answer is short enough to fit on a luggage tag: build the AI concierge for every customer. The word "concierge" is doing a lot of work in that sentence. A concierge resolves things. A concierge remembers your name. A concierge does not transfer you to a different concierge. The bar Decagon is aiming at is not "better than a chatbot." It is "better than the best human you ever talked to."

Whether that bar is reachable is a fair question. Whether it is worth aiming at is not. Customer support touches roughly every adult in a developed economy, sometimes daily, and it remains - by most surveys - the worst routine interaction those adults have with the companies they pay money to. The economic value of fixing it is large. The social value, in the form of not making people cry on a Tuesday afternoon, is harder to put a number on but easier to feel.

06 Why it matters tomorrow

The thing to watch is not whether Decagon's revenue grows. It will. The thing to watch is whether AOPs become the way enterprises talk to their AI agents - not just for support, but for sales, for operations, for the next ten categories of corporate work that look suspiciously like customer support did in 2022: large, expensive, owned by ops teams, and waiting for a layer that lets non-engineers reshape the logic without breaking the system.

If that happens, Decagon will have done something rarer than build a successful AI company. It will have built a new kind of programming interface, one written in English, edited by ops managers, and audited by lawyers. That is a quietly enormous idea. It would also explain why investors keep tripling the price.

07 Back to that Tuesday morning

The Hertz customer outside Phoenix has closed her laptop. The duplicate charge is gone. There is no follow-up survey, no "did this resolve your issue" pop-up, no need to retell the story to anyone. The interaction was, in the dullest possible sense of the word, fine. Which, if you think about how customer support normally goes, is the most radical adjective you could apply to it.

Decagon is two and a half years old. It has 350 people, 100-plus enterprise customers, and a valuation that suggests several large investors believe support is one of the largest software categories nobody has yet bothered to build properly. The bet is straightforward. The execution is not. The morning, for one customer in Arizona, was uneventful. That is the entire point.