Twin is putting AI agents into a practical business workflow: finding missing invoices. The Paris-based company has released Invoice Operator in partnership with Qonto, the fintech startup that offers business bank accounts to more than 500,000 customers across Europe.
The launch marks a shift from AI agents as a broad idea to AI agents as a targeted automation product. Instead of asking users to build scripts, connect APIs, or prompt a general-purpose assistant, Twin is focusing on one repeated task that Qonto customers already have to perform.
What Invoice Operator does
Qonto handles millions of invoices per month, and customers spend several hours per month gathering invoices and uploading them to Qonto. Invoice Operator is designed to reduce that manual work by retrieving missing invoice documents and attaching them to the right transactions.
When a user launches Invoice Operator, Twin first fetches the list of transactions with missing invoices. It then displays the services it needs to access, alongside a browser window where the user can see the agent’s actions.
If a service requires a login, the browser pauses and asks the user to enter credentials manually. After the user completes that step, they can click a button and let the agent continue.
From there, Invoice Operator searches for the relevant past transactions, downloads invoices, and attaches the PDFs to transactions in the user’s Qonto account. The product is built around a narrow but common finance workflow: invoice retrieval, document download, and transaction matching.
Why Twin is using AI instead of older automation tools
Businesses already have several ways to automate repetitive tasks. Some use API-based no-code or low-code automation products like Zapier. Others rely on RPA software such as UiPath.
Twin argues that invoice retrieval is a difficult fit for those older approaches because the task spans a large and changing set of services. Qonto customers may need invoices from thousands, tens of thousands and soon hundreds of thousands of different services that everyone is using, according to Twin co-founder and CEO Hugo Mercier.
That creates a long-tail problem. An RPA approach would require a custom script for each website, and those scripts would need to be updated whenever a website changes. API-based automation has its own limits: Mercier said it took Zapier 10 years to support 8,000 applications on its platform.
Twin says Invoice Operator already supports thousands of applications just a couple of months after work on the product began. The claim is central to its pitch: an AI agent that can operate a browser may cover more services faster than a system that depends on a separate integration or script for every site.
The technology behind the browser agent
Behind the scenes, Twin runs a Chromium-based web browser on a server. The startup uses OpenAI’s CUA (computer-using agent) model, which is designed for agents that can use computers through a browser-like environment.
Twin was one of the 15 companies that got to try CUA in beta. The same model also powers OpenAI’s Operator, a prosumer product that lets a user enter a prompt so an agent can perform an action.
Invoice Operator applies that kind of computer-using agent to a business banking workflow. But Twin is not presenting it as a prompt-heavy tool. Mercier said the company worked on making the experience simple for end users, including people who may not be tech-friendly.
That design choice matters. The user does not have to configure a workflow or tell the agent what to do in natural language. They log into accounts when needed, launch the task, and the agent navigates services to find invoices.
What this says about B2B AI agents
Invoice Operator is Twin’s first product designed for Qonto, but the company sees a broader opening for B2B agentic applications. The common thread is not general conversation. It is the ability to perform repeated work across web services where manual navigation still takes time.
Twin has pointed to several possible use cases beyond invoice retrieval:
- Automatically managing orders for an e-commerce company.
- Classifying the catalog of a marketplace.
- Retrieving information for call center agents.
Those examples share a similar pattern. A user or company needs information from software systems, websites, or accounts; the work is repetitive; and the task may not be fully covered by a single API or a stable script.
The larger question is whether Twin can turn the platform behind Invoice Operator into something developers can use in their own applications. The company is pitching a future where AI agents become cheaper, faster, and more accurate across many kinds of business work.
For now, the Qonto partnership gives Twin a concrete test case. Invoice retrieval is not an abstract demo. It is a recurring finance task with real friction, especially when missing documents are spread across many services. If Invoice Operator can reliably handle that long tail, it will show one way AI agents can move from theory into daily operations.