YP / Breaking
Rasa closes $30M Series C - co-led by PayPal Ventures & StepStone 50,000,000+ open-source downloads and counting CALM ships: LLM nuance, rule-bound certainty, zero hallucination Customers include American Express, Deutsche Telekom & two of the top three banks on earth ARR roughly doubled in 2023 under CEO Melissa Gordon Founded 2016 in a Berlin kitchen - now SF + Berlin, ~150 people
Rasa logo
Yespress // Profile No. 0422

Rasa

The conversational AI platform that refuses to let its chatbots make things up - and the reason a couple of the world's biggest banks let machines talk to their customers in the first place.
Photograph: Rasa speech-bubble mark, found in the wild on a developer's laptop sticker, San Francisco, May 2026.
HQ / San Francisco & Berlin
Founded / 2016
Funding / $70M+
Team / ~150
Stack / Python, LLMs, Kubernetes
License / Apache 2.0 + Commercial

01 / Who they are nowThe bot in your bank app probably has Rasa in it

It is 2026 and a customer in Munich is asking her banking app, in German, to dispute a charge and split it across two cards. The app answers. The answer is right. Behind that quiet little exchange sits a stack the customer will never see - and a company most consumers have never heard of.

That company is Rasa. If you have used a customer service assistant from a global telecom, dialed into a healthcare hotline that didn't waste your time, or argued with a chatbot that actually knew its job, the odds are uncomfortably high that Rasa's software was on the other side of the wire. The company doesn't really announce itself. Its customers don't either. That's sort of the point.

Rasa is the conversational AI platform you've used without knowing - which is exactly how its customers want it.Yespress editorial

02 / The problem they sawLLMs are amazing. Until they aren't.

Generative AI changed the chatbot conversation overnight. Suddenly, anyone could spin up a charming little assistant that sounded human and handled small talk like a sommelier. The trouble began the moment those assistants had to do something real - quote a price, file a claim, move money, look a regulator in the eye.

Large language models hallucinate. They invent. They cheerfully recommend products that don't exist, refunds that aren't owed, and policies the company has never written. For a Saturday-night side project, charming. For a Tier-1 bank, catastrophic.

A chatbot is funny right up until it issues a refund nobody approved.An anonymous compliance officer, paraphrased

The bet Rasa has made since the GPT-era opened up is straightforward, even a bit boring on purpose: you can have the magic of language models without the liability. You just can't let them speak unsupervised. Rasa's whole architecture is built around that one inconvenient idea.

03 / The founders' betTwo physicists, an entrepreneur, and a kitchen in Berlin

Rasa was founded in 2016 by Alan Nichol, Alexander Weidauer, and Tom Bocklisch. Nichol had drifted out of physics - he had spent his PhD applying machine learning to small molecules - and into natural language processing, which by 2014 was the more interesting room to be in. Weidauer brought commercial instincts. Bocklisch brought the engineering rigor.

The story they tell is that the idea was hatched in Weidauer's kitchen, trying to make Dialogflow do something it wasn't really built for. The founders did the unfashionable thing: instead of selling a slick demo, they shipped an open-source Python library and let developers wreck it for them.

We didn't want to build a chatbot. We wanted to build the tool other people would build chatbots with.Alan Nichol, CTO & Co-founder

It worked. The framework has been downloaded over 50 million times. Hundreds of thousands of developers have used it. Basis Set Ventures wrote the first check in 2017, Accel led the Series A in 2019, Andreessen Horowitz came in for the Series B in 2021, and PayPal Ventures and StepStone co-led the $30M Series C in February 2024. Total raised: north of $70 million.

04 / The productCALM, Rasa Pro, and a low-code panic button

Today the product splits into three doors. Rasa Open Source is still there, Apache 2.0, the framework the community knows. Rasa Pro is the commercial runtime - hardened, observable, supported, the thing you put in production when the auditors are watching. Rasa Studio is the low-code UI that lets conversation designers and product managers sit in the same room as developers without anyone having to learn each other's vocabulary.

The interesting bit is CALM - Conversational AI with Language Models. CALM is Rasa's answer to the hallucination problem: an LLM handles understanding and routing, but the assistant can only emit responses that a human has pre-approved. The model gets to be nuanced. The output gets to be controlled. The auditors get to sleep.

Framework

Rasa Open Source

Apache 2.0 Python library for NLU, dialogue management and integrations. 50M+ downloads.

Runtime

Rasa Pro

Enterprise-grade, observable, supported. The version regulated industries deploy.

UI

Rasa Studio

Low-code design surface for non-engineers. Bridges product, design and dev teams.

Architecture

CALM

LLM reasoning paired with validated, pre-approved responses. Zero hallucination by design.

CALM is the part where Rasa stops being a chatbot library and starts being an insurance policy.Yespress editorial

05 / The road hereNine years, four rounds, a Series C

A short Rasa history, with dates

2016
Founded in Berlin by Alan Nichol, Alex Weidauer, Tom Bocklisch.
2017
Seed round; open-source framework starts spreading through developer communities.
2019
Accel leads Series A. First wave of enterprise customers.
2021
$26M Series B led by a16z. Rasa X and product expansion.
2023
Melissa Gordon (ex-Oracle) joins as CEO. ARR roughly doubles.
2024
$30M Series C. CALM and Rasa Studio launch into the generative-AI era.
2026
Two SF + Berlin offices. ~150 people. Banks, telecoms, healthcare onboard.

Timeline assembled from public filings, founder interviews, and the occasional press release we trust.

06 / The proofBanks. Telcos. Numbers that don't really need adjectives.

Rasa's enterprise list reads like a who's-who of "things you do not want a hallucinating chatbot near": American Express. Deutsche Telekom. Two of the world's three largest banks (which Rasa, politely, declines to name). Two of the largest US banks. Healthcare. Travel. Telecom. The pattern is consistent - the more regulated the room, the more likely Rasa is in it.

Rasa, by the numbers

A bar chart only an open-source company could love.
$13M
Series A
2019
$26M
Series B
2021
$30M
Series C
2024
50M+
OSS downloads
~$18M
ARR (est.)
The fastest way to tell whether a conversational AI company is serious is to ask who their customers are - and how loudly they're allowed to say so.Yespress editorial

The other number that matters is 50 million. That's roughly the number of times Rasa Open Source has been pulled off PyPI and onto somebody's laptop. It is, depending on your taste, either a vanity metric or the most honest distribution channel in enterprise software. Either way, it predates every glossy LLM startup of the current cycle by several years.

07 / The missionTrustworthy AI, said without smirking

Every AI company says it builds trustworthy AI. The phrase has roughly the same nutritional value as "delicious" on a menu. What is unusual about Rasa is that the architecture forces it to be true. The product literally cannot say a sentence a human hasn't pre-approved. The mission has a contract under it.

Trust isn't a slogan when your runtime refuses to ship sentences your operator hasn't signed off on.Yespress editorial

CEO Melissa Gordon, who joined in 2023 after a long run at Oracle, has framed the company's pitch in less ethereal terms: enterprises do not want a wizard, they want a worker. They want predictability, observability, deployment options - on-prem, cloud, hybrid - and they want to know exactly what the assistant said to which customer at 11:14 last Tuesday. Rasa was built for that audience before that audience knew it had a name.

08 / Why it matters tomorrowAgents are coming. So are the regulators.

The next wave of AI is agentic - assistants that don't just talk, they act. They move tickets. They pay invoices. They cancel flights. The blast radius of a bad answer doubles each time you give an agent a new permission.

Which is why the boring questions matter again. Who watches the agent? Who can replay its decisions? Who can prove, in writing, that it didn't go off-piste in a way that the bank examiner would find amusing? Those are the questions Rasa has been answering since before they were fashionable.

The next decade of AI is going to be won by whoever makes the boring parts work.Yespress editorial

Back in Munich. The customer closes the app. The dispute is filed. Two cards, split correctly, in German, on a Saturday night. She doesn't post a review. She doesn't tell anyone. The whole transaction was, in the best possible sense, uneventful. That's the product. That's the bet. That's Rasa.

09 / Tell someoneShare, then go read the source