Every enterprise board now wants an agent strategy. Few enterprises have the governance to match. In a Thoughtworks survey of 3,200 CIOs published this month, nearly 3,000 said AI adoption was outpacing their ability to govern it.
The gap is widest further down the market.
McKinsey found that only 22% of organisations with under $1 billion in annual revenue were scaling AI agents, against 40% of larger firms. Mid-sized companies often lack the engineers to build agents in-house. But relying on a vendor’s out-of-the-box agents means giving up control of data, business logic and oversight.
AI product teams are increasingly weighing up building agents they have complete control over, with the overhead of building an agent workbench and keeping it up-to-date with the latest and greatest models and engineering practices.
Rasa is betting firms don’t have to choose.
The company has developed an open-code platform for building repeatable AI agents that can be deployed across the enterprise with roots in Rasa’s earlier open-source framework for building chat and voice AI assistants.
Its platform is used by Swisscom, BNP Paribas and N26, and Forrester named it a Strong Performer in its 2026 Wave for customer-service conversational AI.
Rasa offers the customers the ability to read and modify the platform’s source code, which it markets as “open-code.” Several rivals, including Cognigy and Kore.ai, offer on-premises deployment, but not the code itself.
“Making a smoothie out of prompts”
A core problem enterprises can run into scaling AI agents, said Rasa's CTO and cofounder Alan Nichol, is that most teams try to steer agents to desired outcomes by rewriting prompts. That works in a demo but it can break at scale.
“Knowing that when you fix something it stays fixed is important for maintaining velocity in a complex project,” he says. If you tweak a prompt to fix one issue, nothing stops a feature shipped next month from bringing it back, “because everything’s just based on prompts and [you’re] making a smoothie out of all the prompts.”
Multiply that across hundreds of agents and you are “building on quicksand.”
Rasa’s solution was creating an agent orchestration framework where customers can see into the code and customise each agent with deterministic, code-enforced controls.
Nichol called the result a “spectrum of supervision.” He explained that an agent on the Rasa platform can be created using natural language prompts but when a more high-stakes action needs to be taken, for example, processing a refund, then a deterministic turn can be added in.
Determinism where it counts
Rasa agents are ultimately software artifacts, not collections of prompts. That means they go through the same CI/CD pipeline as any other code, with automated tests, version comparison, progressive delivery and automatic rollback.
Before release, Nichol explained, Rasa runs each agent through a hundred simulated voice and text conversations and flags what failed. Because of this centralised governance, an enterprise can scale agents faster, trusting teams that understand the line of business to create agents without the bottleneck of a centre of excellence having to build out every use case.
Nichol said that unlocking that “true team-of-teams setup” is near impossible if an engineer is building agents within a SaaS tenant. “There's no way to reuse that [work] across projects or bring any of that development lifecycle in house. You're beholden to their fit, and like whatever buttons they have for you to push, those are the buttons you can push.”
Low-code on top, real code underneath
That code-first design is what makes it safe to let non-engineers build agents.
In Rasa’s Studio workspace, product and operations staff build, review and test agents visually, and their changes are written as code into a shared repository.
Developers work on the same codebase from their IDE, and coding agents can connect through Rasa’s MCP tools. The enterprise owns the result and can run it on its own infrastructure with, critically, its own free choice of model.
Switzerland’s largest telco Swisscom runs Rasa agents across its customer service hotline and in-app chat. They cover thousands of request types across landline, broadband and mobile, for households and businesses, in four languages.
“There’s no possibility for a single team to be responsible for all of these agents and to know how to build all of these skills,” Nichol said. Using Rasa, AI product teams can work in a hub-and-spoke model: hand development to the teams that know each part of the business, “but centralise the governance, the provisioning, the building and the shipping of agents.”
A skill built once, such as identity verification, can then be reused across channels. According to Rasa, Swisscom was able to cut operating costs of its digital customer service by 50% after partnering with Rasa to build AI assistants.
Nichol said companies can also save money in who they staff on their product teams.
“The teams that enterprises used to staff to build chatbots often had one or two natural language processing PhDs in there. Really heavily technical people who would deeply know about Rasa and train their own models and customize things. And definitely now it's more of a sort of citizen developer or someone who is semi-technical. So there's definitely been an evolution of how enterprises expect to staff these teams and who they expect to be able to build agents and put them in production.”
From deflection to revenue
Measuring return is where many agent projects stall.
In customer support, the metrics are familiar: satisfaction, retention and resolution time. Nichol argues the prize isn’t keeping customers away from human staff. It’s “actively trying to add value to their users”, with in-app agents that answer questions and point customers to services they didn’t initially know about.
Per Nichol, about a third of Rasa customers now build internal agents. Because Rasa is open-code and can be deployed and run entirely within a customers’ environment, customers can use the AI agents on data that they wouldn’t want to ship out to a “vanilla OpenAI API.”
Nichol said customers can run agents on LLMs accessed via enterprise agreements (Claude in AWS Sagemaker or OpenAI via Azure) – or even on their own models in airgapped environments. Built on principles rooted in open source, Rasa’s model is appealing to customers who are increasingly viewing their enterprise data as a "valuable asset to build upon.”
The early wins from internal agents come from making revenue-generating staff more productive, Nichol said.
One banking customer gives each trader an agent that works as a round-the-clock data-science intern, supplying research, numbers and analysis on demand. “Because these people make mega money for the bank, it’s like a virtual employee on their team,” Nichol said.
“That’s truly revenue generating.”
Delivered in partnership with Rasa.