From chatbot to autonomous agent
Large Language Models (LLMs) have come a long way since they first became mainstream back in 2022, when OpenAI released ChatGPT to the general public and gained its first million users in just five days, then reaching 100 million users early in 2023. If you used those first models, you certainly encountered hallucinations, images with six fingers per hand, and conversations that went in circles with the chatbot contradicting itself and apologizing repeatedly. Nowadays models can analyze vast amounts of data at a PhD research level. You can create fully consistent videos with a single text prompt, and with the use of special tools, give hands to your AI, allowing it to handle your taxes directly (not financial advice).
But how do you transition from using AI as a simple chat partner to delegating complex, automated tasks? There are a few tools you can use to start experimenting with agentic automation, and it certainly depends on your environment. If you are looking to deploy workflows within your company, you might already have Power Automate from Microsoft. If you are not tied to those services, tools like n8n or upcoming orchestration tools from Anthropic and OpenAI are worth exploring.
Case Study
How AI can support the development of a Class III medical device
Tools like n8n or Power Automate rely on the definition and use of agents that can take on separate tasks, work in parallel, and deliver information or context to each other as a regular team of people would do. In our webinar, we go through a workflow generated to simulate the development of a Class III medical device, an AED. The workflow takes clients’ requirements as an input and distributes tasks to different agents much like departments within a company. There is a requirements department (in our case a requirements agent) that takes the client’s requirements and makes sure there are no inconsistencies before writing down the technical requirements. Those requirements are then passed to the development team (a set of software, hardware, and mechanical agents) that implements those requirements and communicate with each other to avoid conflicts in implementation. As a last step, and to simulate the process of MDR certification, we have two agents that go through the documentation generated from the complete workflow and compare it to the MDR requirements. This step functions as a small internal audit your team can use to try to reduce certification rejections by making sure all required documentation is available once you reach out to your notified body.
The workflow can be run in a matter of minutes and provides the development team with key notes before they even start the development process. This gives the team a head start and transforms a significant share of the development effort into review tasks experienced developers and engineers can go through.
This is just one example of what can be achieved with agentic automation, and teams are already building workflows to handle repetitive or time-consuming tasks.
More use cases
Imagine that as a medical device developer, you have a workflow that scans through the incoming emails of your company and identifies reports from the market. Since you are required to react to these reports within a time window, the workflow can automatically summarize the email, classify the severity of the reported issue, and send a direct urgent message to your PRRC (Person Responsible for Regulatory Compliance) to review the case.
Or maybe you want to have regular internal audits to make sure all your project documentation is up to date in case of any client or notified body visit, but your team is already overbooked. You could have a workflow that runs once every three months, checking all documentation generated on each project for inconsistencies or conflicts. This way, your team always knows exactly where to focus.
Examples like these show that AI is increasingly taking on tasks that previously required experienced individuals or, in the case of developing an AED, entire teams of highly trained professionals, reducing hours or weeks of work to a fraction of the time. Our human responsibility is shifting from development to review and approval. Some of the teams at Corscience are already exploring the use of AI agents and automatic workflows to give them a head start on new projects.
Our teams are exploring the use of AI in topics like cybersecurity, requirements engineering, test automation, and more. If you are curious about these topics, make sure to check out our recently published podcast episode on cybersecurity, as well as our upcoming AI Act episode on the regulatory implications these tools bring.
The question that remains: what will you or your team do with the extra time gained?
Lamborghini Sotelo | AI Consultant
