Introducing Invoice Operator — Twin's First Agent built with Qonto and OpenAI

March 27th, 2025


Today, we're thrilled to unveil Invoice Operator, the first autonomous agent built to automate invoice collection—powered by Twin's Agent, OpenAI's state-of-the-art CUA model, and now available to Qonto customers at twin.so/qonto.


Solving the Pain of Invoice Collection

For executives and business owners, chasing down invoices from online services at month-end is notoriously tedious. On average, a Qonto customer spends over 3 hours each month manually retrieving invoices from multiple platforms. Across Qonto's entire customer base, that's 1.2 million invoices every month across more than 100,000 services! Previous attempts at automating invoice retrieval were unsuccessful because most services lack APIs for invoices, and traditional RPA approaches required creating and continuously maintaining thousands of custom scripts that would inevitably break whenever interfaces changed—until now.

Meet Invoice Operator: Automating Invoice Retrieval at Scale

To collect your invoice, simply go to twin.so/qonto, log into each of your services, and Invoice Operator handles the rest. It collects transactions directly from your bank account, identifies the missing attachments, and creates agents with the task to retrieve the matching invoice on the service provider website. It then locates and downloads the invoices, reconciles them with your transactions, and uploads them into Qonto. Invoice Operator is powered by the Twin Agent, a browser use agent capable of autonomously navigating, reasoning, and taking action within any application by controlling a browser in the cloud. Unlike traditional APIs and fragile RPA scripts, Twin Agent can adapt to any interface through real-time reasoning, making it as reliable reasoning, making it as reliable as API integrations and as universally compatible as RPA solutions.

Invoice Operator has already been used by 1,500 beta testers, and during a session has been able to retrieve 500 invoices, from 50 different service providers, in less than 10 minutes! Now, available for all Qonto customers.


Try Invoice Operator


The First AI Agent, Deployed to 500,000 SMBs

Agentic AI is revolutionizing automation, yet few agents achieve production-grade scale. Serving 500,000 European SMBs, Qonto required the highest standards of reliability, security, and scalability, which led to several key innovations:

Reliability

Collecting invoices autonomously from thousands of online services pushed our agent to its limits. We developed new techniques to (1) correct the model when it makes mistakes and teach it complex interfaces and scenarios; (2) generate domain-specific data and capabilities—such as accurately identifying URLs for each service or managing complicated reconciliation cases involving mismatched amounts, currencies, or dates; and (3) automatically evaluate the agent's trajectories and self-improve.

Security

Operating within customer applications necessitates high safety standards. On the authentication side, credentials are not stored and are immediately destroyed once the session ends. We are currently testing a feature for securely storing user credentials and automating authentication. Additionally, we implemented guardrails to prevent the agent from taking harmful actions. Finally, we enhanced our internal processes to achieve the highest standards of data security and confidentiality, achieving compliance with GDPR and SOC2 Type 2.

Scalability

The Twin Agent controls custom-built Chromium browsers in the cloud, tailored specifically for our agent. Browsers, being complex and resource-intensive, traditionally have high operational costs and latency. To overcome this, we designed a new agent-browser interface and a specialized Kubernetes infrastructure, enabling extremely low-latency interactions between agent <> browser and user <> browser even at high volumes of customers and invoices. Our infrastructure maintains a latency below 50 to 100 ms at high loads—a benchmark we believe is state-of-the-art.


Learn more about Twin


The First Production-Level Agent built on OpenAI's CUA

When we started Twin 18 months ago, we saw the potential of LLMs to automate complex enterprise workflows by autonomously controlling browsers.

As humans, we navigate digital environments seamlessly, often unaware of the intricate series of decisions and actions involved. However, if you think of it, even simple tasks require numerous small yet critical decisions and actions. The first generation of LLMs, despite strong reasoning capabilities, frequently struggled with tasks that humans find effortless.

Building on this insight, we launched our initial HTML-based agent in March 2024 using GPT-4o. By August 2024, we advanced further with a vision-based agent leveraging GPT-4o in combination with UGround. While latency, cost, and accuracy improved significantly, we experienced a substantial leap forward with our CUA-based agent.

Following OpenAI's release of Operator, Twin was selected among a small number of companies worldwide to alpha-test CUA in the API—the model behind Operator that merges GPT‑4o's visual capabilities with reinforcement learning-based reasoning. During initial testing, CUA impressed us with its human-like navigation, contextual, real-world understanding, and great history management, enabling the successful execution of long-horizon tasks exceeding 50 steps. Most importantly solved around 80% of the problematic cases, that were not solvable with our previous agent.

CUA

The speed, accuracy, and cost of our agent are internal North Stars that we are focused on. In the past 12 months, we’ve increased speed by 6x, reduced costs by 35x, and improved accuracy by 55%. We anticipate significant improvements over the next few months, enabling many more use cases across verticals — and we’re excited to partner with OpenAI’s CUA team to further improve the model capabilities from real-world scenarios.

Benchmark
Twin A1
Twin A2
Twin A3
Latency/step
35s
19s
6s
Cost/step
$1.10
$0.15
$0.03
Accuracy
54%
68%
84.0%

More info on CUA

The End-to-End Solution for Enterprise to Build and Deploy Vertical Agents

With Twin, companies can effortlessly build specialised agents able to automate complex workflows on any application - no more limited API integrations or fragile RPA scripts.

Currently, Twin is developing new agents for diverse industries, including e-commerce platforms, digital clinics, insurances, marketplaces, automation tools, call centers, and customer support platforms.

Twin provides a comprehensive, end-to-end solution, empowering teams to efficiently build, test, and deploy reliable, secure, and scalable agents at scale.

The platform includes scalable browser infrastructure, secured authentication features, domain-specific finetuning of the agent, and observability tools for continuous monitoring and performance enhancement in production.


Get in touch to discuss your agentic project


Get in touch