Decagon Valuation and Revenue: How It Grew to $4.5B

Decagon builds AI agents that answer customer support conversations for large companies. They handle refunds, cancellations, disputed charges, rebookings and phone calls without a human picking up.

As of October 2026, Decagon's latest valuation is $4.5 billion. That price was set by a $250 million Series D led by Coatue and Index Ventures in January 2026. Decagon's employee tender offer in March 2026 used the same price, and the company has not announced a newer round.

On revenue, Decagon told Newcomer in August 2026 that it had crossed $100 million in annualized revenue. That is a company disclosure to a reporter, not an audited figure. It compares with roughly $10 million at the end of 2024, which The Information reported in November 2025.

Adding up its five announced priced rounds, Decagon has raised about $481 million in funding, plus an undisclosed strategic investment from Deutsche Telekom's venture arm.

The one-line reason it grew: Decagon won customers in head-to-head bake-offs. Then it made its agents easy enough for customer support teams to change themselves, so each deployment could grow without a vendor engineer on every change.

The company is only about three years old. Jesse Zhang and Ashwin Sreenivas started it in August 2023.

Decagon's growth in one table

Date	Milestone
August 2023	Jesse Zhang and Ashwin Sreenivas start Decagon after meeting at an a16z founders' retreat in Utah
2023	a16z writes the first check and leads a $5M seed round
June 2024	Decagon emerges from stealth with $35M (seed plus a $30M Series A led by Accel) and customers including Eventbrite, Bilt, Webflow, Substack and Rippling
October 2024	Raises a $65M Series B led by Bain Capital Ventures; Forbes reports a $650M valuation; total funding reaches $100M
December 2024	Publishes its per-conversation and per-resolution pricing model
End of 2024	ARR passes eight figures, about $10M according to The Information
February 2025	Launches Decagon Voice with ElevenLabs
June 2025	Raises a $131M Series C at a $1.5B valuation, co-led by Accel and a16z; total funding reaches $231M
September 2025	Launches Voice 2.0 at its first Decagon Dialogues conference
November 2025	Starts a Deutsche Telekom pilot and takes a strategic investment from T.Capital; The Information reports talks at a $4B–$5B valuation and "significantly more" than $30M ARR
January 2026	Raises a $250M Series D at a $4.5B valuation; says it signed 100+ new enterprise customers in 2025
March 2026	Closes its first employee tender at $4.5B for its 300+ employees; says 80%+ of model traffic runs on its own models
August 2026	Tells Newcomer it has crossed $100M in annualized revenue
September 2026	Opens in São Paulo and expands into Benelux
October 2026	Launches Voice 3 and Personal Agent Gateway; becomes a launch partner on OpenAI's B2B marketplace

1. Decagon started by asking companies what they would pay for

Decagon did not start with a product idea.

Zhang and Sreenivas met at a founders' retreat in Utah hosted by Andreessen Horowitz. Both were repeat founders deciding what to build next. Andreessen Horowitz wrote that Zhang had sold his social gaming app Lowkey to Niantic in 2021, and Sreenivas had sold his computer vision startup Helia to Scale AI in 2020.

Both had grown up competing in math and science Olympiads, according to Index Ventures.

In August 2023, the pair surveyed dozens of startups to find a product companies needed and would pay for, Forbes reported. Bloomberg reported that they asked companies which part of their operations AI could help most.

The answer kept coming back to customer support.

It was a good choice for a new AI company. Support is expensive, measured closely and full of repetitive questions. A company already knows how many tickets it gets, what each one costs and how satisfied customers are. That makes it easy to prove whether AI is working.

Zhang told Forbes he named the company after the 10-sided shape because "we think AI agents can be 10x employees."

a16z wrote the first check at inception and led the seed round in 2023.

2. Decagon sold to fast-growing tech companies first

Decagon's first customers were not airlines or banks. They were software and consumer-tech companies that move quickly and already had heavy support volume.

When Decagon came out of stealth in June 2024, it named Eventbrite, Bilt, Webflow, Substack and Rippling as customers. It announced a $5 million seed led by a16z and a $30 million Series A led by Accel, for $35 million in total.

The angel list was also a customer list in waiting. It included Box CEO Aaron Levie, Airtable CEO Howie Liu and Rippling COO Matt MacInnis.

Zhang's pitch was about trust, not demos. "In the AI customer support chatbot space, we see more flashy demos than companies that can truly win the trust of an enterprise company," he told VentureBeat at launch.

Bilt, the rewards and rent-payments company, became an early proof point. In Decagon's Series B announcement, Bilt's VP of customer support said Bilt got 60,000 tickets a month and Decagon's agents handled 70% of them.

"Working with Decagon was like hiring 65 agents overnight," he said.

These customers did two jobs. They paid, and they gave Decagon specific, quotable numbers it could take to the next, larger buyer.

3. Decagon won customers in bake-offs

AI customer support became crowded almost immediately after ChatGPT. Buyers responded by testing vendors against each other.

Forbes reported in October 2024 that almost all of Decagon's clients ran a "bake-off" between Decagon and competing tools, including Salesforce's. Each vendor's agent answered the same customer questions, and the buyer compared which resolved the most accurately. Zhang told Forbes that Decagon came out ahead each time.

That is Zhang's claim, not an independent tally. Intercom's CEO later told Forbes that Intercom wins every bake-off it runs against Decagon.

The best documented example is ClassPass.

According to Decagon's ClassPass case study, ClassPass ran a formal RFP and evaluated Decagon against 12 other AI products. Decagon says it won on response accuracy, no-code tools for the support team, responsiveness and cost savings. Forbes later described the same contest as Decagon against 11 other vendors.

The results became marketing. In the same case study, ClassPass reported a 95% cost reduction and a 32% increase in deflection, and its support lead said deflection at launch was 10 times higher than expected. ClassPass also expanded its chat support from 16 hours a day, five days a week, to 24/7.

The bake-off suited Decagon. A new vendor with no brand could still win if its agent simply answered more questions correctly on the buyer's own data.

4. Decagon put agent logic in plain English

Winning the first test was only half the problem. Support policies change every week, and an AI agent that only the vendor can edit goes stale quickly.

Decagon's answer was Agent Operating Procedures, or AOPs.

In its Series C announcement, Decagon described AOPs as natural-language instructions that compile into code. A support manager can write how the agent should handle a lost card or a late delivery, the way they would write a training document for a new hire. Engineers keep control of the underlying code, integrations and guardrails.

Decagon framed this as a product-driven approach rather than "the same professional services approach as the rest of the industry."

That matters for growth in a specific way. If every workflow change needs a vendor engineer, a vendor can only take on as many customers as its engineering team can serve. If the customer's own support team can make changes, each new account costs Decagon less to run.

Decagon kept adding tools around the same idea: simulations to test agents before release, versioning to A/B test changes, and Watchtower to monitor live conversations. Its March 2026 Duet launch post argued that "the initial build is only 10% of the work; the other 90% is what comes after."

The Series C results gave the approach credibility. Decagon said average deflection was nearing 70%, Duolingo was well above 80%, and Oura reported a 3x increase in customer satisfaction scores.

5. Decagon priced per conversation, not per seat

Traditional support software charges per human agent seat. That model breaks when the software is doing the work of those agents.

In a December 2024 pricing post, Decagon's director of product said the company supports two models. Per-conversation pricing charges a fixed rate for every incoming conversation. Per-resolution pricing charges a higher rate only when the AI fully resolves the conversation, with no charge for escalations.

Decagon said "the vast majority" of its customers choose per-conversation pricing.

Its reasoning was practical. Outcome pricing creates arguments over what counts as a resolution. If a frustrated customer gives up and leaves, is that resolved? Per-conversation pricing gives the buyer a predictable budget and removes the incentive for the vendor to count weak answers as wins.

The pricing also ties Decagon's revenue to usage. When a customer routes more channels, regions or products through Decagon, Decagon earns more without a new contract for every seat.

Decagon does not publish rates.

6. Decagon added voice and moved up-market

Chat and email were the starting channels. But many of the most expensive support interactions still happen on the phone.

In February 2025, Decagon launched voice agents built in partnership with ElevenLabs, the AI audio company. The agents ran on the same AOPs as chat and handled inbound calls about account access, returns and charge disputes.

In September 2025, at its first Decagon Dialogues conference, Decagon launched Voice 2.0. It said latency fell 65% and the release added cross-channel memory and outbound calling.

Voice opened the door to a different kind of customer. Airlines, car rental companies, banks and telecom operators run large call centers, and their support budgets are much larger than a software startup's.

The customer list changed accordingly. In its January 2026 Series D announcement, Decagon said it signed more than 100 new enterprise customers in 2025. It named Avis Budget Group, Hertz, Block, Affirm, Chime, Oura Health, Mercado Libre and Deutsche Telekom, among dozens of others.

Deutsche Telekom came in through a commercial pilot announced in November 2025. At the same time, T.Capital, Deutsche Telekom's corporate venture arm, made a strategic investment of undisclosed size.

By October 2026, Decagon's company description named Delta Air Lines, American Airlines, Deutsche Telekom, Ticketmaster, Hertz, Affirm and Snap as customers.

The Series D release also showed where those customers came from. Decagon said 33% had no prior AI automation, 53% replaced legacy systems such as phone menus or CRM-based agents, and 14% chose Decagon over building in-house.

7. Decagon trained its own models

Most AI application companies start on top of OpenAI or Anthropic models. Decagon did too.

Over time, it moved most of its traffic onto models it trains itself. In March 2026, Decagon said in its Decagon Labs announcement that more than 80% of all model traffic runs on its own models.

Its argument was that general models are not built for support work. A support agent has to detect when a caller has finished speaking, follow a workflow exactly, and catch its own hallucinations, all with low latency. Decagon built a network of specialized models, each post-trained for one of those jobs.

This helps growth in three ways.

First, cost. Smaller specialized models are cheaper to run than frontier models on every customer conversation, which matters when pricing is per conversation.

Second, speed. Voice agents feel broken if they pause too long.

Third, defensibility. If every vendor uses the same frontier model, the product becomes easy to copy. Models trained on the patterns of hundreds of millions of support conversations are harder to replicate.

Decagon has not walked away from frontier labs. In October 2026, it described OpenAI's models as "an important part of Decagon's constellation of models," alongside its own.

The first speech model from Decagon Labs, called Chord, shipped with Voice 3 in October 2026.

8. Decagon bet against forward-deployed engineers

The biggest strategic difference between Decagon and its main rival, Sierra, is how much hands-on engineering each sale requires.

Forward-deployed engineers are vendor engineers who work inside a customer to customize the software. The role has spread quickly across enterprise AI.

Zhang thinks long-term reliance on them is a warning sign. In an August 2026 interview with Newcomer, he argued that embedding engineers long term means the software doesn't work. "We are in a use case where speed matters a lot," he said. "The way we're going to prove it to you is that you're able to go live way faster and you're able to make changes and hill climb the quality of the agent a lot faster."

Decagon still uses deployment engineers to launch customers. PYMNTS reported that Zhang has said they get deployments live in roughly six weeks. The difference, in Decagon's telling, is that the customer's own team runs the agent afterwards.

Zhang also described one customer that spent a year with Sierra's embedded engineers and built three workflows, as PYMNTS summarized from the Newcomer interview. After switching to Decagon, he said, the same customer built seven new workflows in about a month. That is a single anecdote from a competitor, so treat it as Decagon's pitch rather than proof.

Duet, launched in March 2026, pushes the same idea further. It is an AI assistant inside Decagon that reads real customer transcripts, drafts AOPs, generates test scenarios and suggests fixes when performance dips.

The growth logic is simple. If customers can build and fix agents themselves, Decagon's revenue can grow faster than its implementation headcount.

9. Decagon turned support agents into a "front door"

Support was the entry point. The longer-term plan is to handle more of the customer relationship.

Voice 2.0's outbound calling let agents make appointment reminders, onboarding calls and renewal prompts. Decagon's June 2026 product roundup said Hunter Douglas generated $1 million in AI-attributed revenue through Decagon, with an 85% higher average order value.

In October 2026, Decagon said its agents now qualify leads, onboard customers, collect payments and grow relationships, not only resolve tickets. It launched Agent Modules to extend agents into those journeys for financial services, travel, healthcare, retail and telecom.

"The promise of AI within customer experience is not just a support channel that happens to use AI, but a singular 'front door' for customers to interact with a business," Zhang said in the October 2026 launch release.

Decagon is also preparing for customers who send their own AI agents. In its Personal Agent Gateway post, Zhang noted that Meta launched Muse, OpenAI introduced dots, and Instinct started placing phone calls for its users in the month before. Decagon's gateway detects those agents, routes them to a dedicated channel and checks what the customer has authorized. It also published a protocol called PACT for agents to identify themselves.

Each new journey adds conversations to the same platform. With per-conversation pricing, that is how a support contract can grow into a much larger one.

10. Decagon expanded through partners and new countries

Through 2026, Decagon added distribution channels that let it reach buyers without a direct sale each time.

In June 2026, it became an accredited integration on Five9's contact center marketplace, with joint customers including Wealthsimple, Chime and Oura.
In September 2026, it partnered with Plaid so customers can connect bank accounts inside a support conversation. Decagon called financial services one of its fastest-growing verticals.
Also in September 2026, it partnered with Databricks and became a launch partner on Databricks Marketplace.
In October 2026, it became a launch partner on OpenAI's B2B marketplace, where eligible OpenAI enterprise customers can apply part of their existing OpenAI commitment toward Decagon.

The OpenAI marketplace deal targets procurement, which is often the slowest step in an enterprise sale. A buyer that has already committed spend to OpenAI can use some of it on Decagon instead of opening a new budget line.

Geography expanded too. Decagon's January 2026 release listed offices in San Francisco, New York and London. In September 2026, it opened a São Paulo office, citing Mercado Libre and NG.Cash as existing customers in the region. A week later it announced a Benelux expansion and said Rituals had rolled its Decagon agent out across 19 countries and 15 languages in two months.

By October 2026, Decagon said it had six offices worldwide.

How fast did Decagon grow?

Decagon's revenue numbers come from a mix of company statements and reporting, so it helps to lay them out by source.

End of 2024: Decagon said in its Series C announcement that it grew from zero to eight-figure ARR during 2024. The Information reported the figure as about $10 million.
November 2025: The Information reported Decagon was generating "significantly more" than $30 million in ARR.
End of 2025: Research firm Sacra estimates about $44 million in annualized revenue. This is an outside estimate.
August 2026: Decagon told Newcomer it had crossed $100 million in annualized revenue. Sacra separately estimates the company reached $100 million in July 2026.

That is roughly 10 times growth in about 20 months, from about $10 million to more than $100 million.

The valuation moved even faster:

October 2024: $650 million after the Series B, as Forbes reported.
June 2025: $1.5 billion after the Series C, as Reuters reported.
January 2026: $4.5 billion after the Series D, including the new money, as Bloomberg reported.
March 2026: $4.5 billion again in the employee tender.

That is about a 7x increase in 15 months. Decagon's own Series C release said that round drew 5 times more investor demand than it had room for.

At $100 million in annualized revenue, the $4.5 billion valuation is about 45 times revenue. The multiple was much higher when the price was set in January 2026, because revenue was lower then.

A few caveats apply. "Annualized revenue" is a run rate: a recent period's revenue multiplied out to a year. It is not trailing 12-month revenue, and Decagon has not published audited figures. It has also not disclosed net revenue retention or churn.

Decagon's growth loop

Decagon's growth comes down to one loop:

A company tests Decagon against other vendors on its own support data.
Decagon wins on accuracy and cost, and goes live in weeks.
The customer's own support team edits the agent through AOPs and Duet, so the agent keeps improving without a vendor ticket.
Better results let the customer route more conversations, channels and regions through Decagon.
Per-conversation pricing turns that extra volume into more revenue.
Customer metrics become case studies and conference talks.
Those proof points help win the next bake-off, now with larger companies.

Two things strengthen the loop over time.

The first is data. More conversations give Decagon Labs more material to train specialized models, which lowers cost and latency.

The second is scope. As agents move from support into onboarding, sales and collections, each customer has more reasons to send conversations through Decagon.

What could slow Decagon down?

The first risk is competition. By Newcomer's account, Decagon is "far from first place" in AI customer service. Newcomer reported in August 2026 that Sierra has three times Decagon's funding and has "well-surpassed" $200 million in annualized revenue, according to a source close to Sierra.

Salesforce is also pushing hard. The same Newcomer report said Salesforce's Agentforce passed $1 billion in ARR and that Salesforce agreed in June 2026 to buy Fin, formerly Intercom, for $3.6 billion. Incumbents already own the CRM and help desk systems that Decagon plugs into.

The second risk is the bet against embedded engineers. Some large enterprises want white-glove help and are willing to pay for it. If they keep choosing vendors that provide it, Decagon's lighter-touch model could cost it some of the biggest accounts.

The third risk is that frontier models keep getting cheaper and better. Decagon's own models are an advantage today. If general models close the gap on latency and accuracy, that edge narrows, and foundation labs could sell more directly to enterprises.

The fourth is the valuation itself. A $4.5 billion price on roughly $100 million of run-rate revenue leaves little room for a slowdown. Decagon has not published retention data, so outsiders cannot yet tell how much of its growth comes from existing customers expanding versus new logos.

Finally, there is reputation risk. Decagon's agents speak directly to millions of consumers for brands like airlines and banks. A widely reported failure could slow adoption across the whole category.

Borrowing Decagon's customer proof

Decagon kept turning customer results, such as ClassPass's cost reduction and Rituals' zero backlog, into case studies and posts that helped it win the next deal. Okara's Writer Agent drafts strategy-led articles from results like these and can publish them. The LinkedIn Agent drafts the posts that put those numbers in front of the buyers you want to reach.

Frequently asked questions

What is Decagon's valuation?

Decagon is valued at $4.5 billion as of October 2026. That price was set by its $250 million Series D in January 2026, led by Coatue and Index Ventures, and reused in a March 2026 employee tender offer. It was up from $1.5 billion in June 2025 and $650 million in October 2024. Decagon has not announced a newer round.

What is Decagon's revenue?

Decagon told Newcomer in August 2026 that it had crossed $100 million in annualized revenue. That is a run-rate figure disclosed by the company, not audited revenue. Earlier, The Information reported about $10 million in ARR at the end of 2024 and "significantly more" than $30 million in November 2025.

How much funding has Decagon raised?

Decagon has disclosed about $481 million across five priced rounds: a $5 million seed, a $30 million Series A, a $65 million Series B, a $131 million Series C and a $250 million Series D. It also took a strategic investment of undisclosed size from T.Capital, Deutsche Telekom's venture arm, in November 2025. Forbes has reported total funding of more than $500 million.

Who are Decagon's investors?

Andreessen Horowitz led the seed round and wrote the first check. Accel led the Series A, Bain Capital Ventures led the Series B, Accel and a16z co-led the Series C, and Coatue and Index Ventures co-led the Series D. Other investors include A*, BOND, Elad Gil, Forerunner, Ribbit Capital, Avra, Definition Capital, ChemistryVC, Starwood Capital and T.Capital.

Who founded Decagon?

Jesse Zhang, the CEO, and Ashwin Sreenivas, now co-founder and president, started Decagon in August 2023. They met at a founders' retreat hosted by a16z. Zhang previously founded Lowkey, which Niantic acquired in 2021. Sreenivas previously founded Helia, which Scale AI acquired in 2020.

Who uses Decagon?

Decagon has named customers including Delta Air Lines, American Airlines, Deutsche Telekom, Ticketmaster, Hertz, Avis Budget Group, Affirm, Chime, Block, Snap, Mercado Libre, Oura, Duolingo, Notion, Rippling, ClassPass and Bilt. It said it signed more than 100 new enterprise customers in 2025.

How does Decagon make money?

Decagon charges for usage rather than seats. Customers pay either a fixed rate per conversation or a higher rate per fully resolved conversation, with no charge for escalations. Decagon says most customers choose per-conversation pricing. It does not publish its rates.

How is Decagon different from Sierra?

Both sell AI customer service agents to large companies. Sierra is larger: Newcomer reported in August 2026 that it has three times Decagon's funding and well over $200 million in annualized revenue. Decagon's pitch is speed and self-service. It argues customers should be able to change their own agents through natural-language AOPs, rather than relying long term on the vendor's embedded engineers.