The Evolution of Workflow Automation: AI, Signal Processing, and Human-in-the-Loop Systems

Automation has always played a crucial role in optimizing workflows, from publishing to industrial processes. But recent advancements in signal processing toolchains are pushing workflow management into new territory, where AI and human input coexist in more dynamic ways.

Signal Processing and Early Workflow Automation

Signal processing refers to the manipulation, transformation, and analysis of data streams—historically applied to audio, video, and communications. Software like MaxMSP and PureData (as it happens I co-wrote an early-ish book on PureData) emerged as powerful tools for artists and engineers, enabling real-time signal manipulation in music and multimedia projects. These tools belong to a broader category of visual programming environments for signal processing, which allow users to construct workflows using modular, node-based structures.

When I first went to PLOS, I wanted to design a system like this for publishing and later wrote some blogs about it, but at the time, it was too hard to build a signal processing workflow because it required too much abstraction and the technology wasn’t quite there. Instead, we settled for the modernizing paradigm of the time—the Kanban workflow, a card-based system that structured task management more fluidly. While this was a significant improvement, the signal processing model has always remained interesting to me as a more dynamic and adaptable approach to workflow automation.

In 2017 or so Jure Triglav and I collaborated with Simon Porter and Jared Watts from Digital Science to experiment again with early Coko tech and workflow automation. Simon and Jared did most of the work (and it was their idea) and produced some interesting results but ultimately it still felt a little early for this approach.

However tools for workflow automation have become more sophisticated, and a lot of the problems that made signal processing workflows difficult have been solved. As a result, the potential of these systems is now starting to get really interesting.

The Rise of Modern Workflow Automation Tools

Modern workflow automation tools provide a structured way to process information using linear process chains (workflows). At their core, these tools allow users to build out logical flow diagrams that represent step-by-step processing chains. Each workflow consists of:

• Input triggers – Events that start the workflow, such as an incoming email, a form submission, a file being uploaded, or a scheduled time-based event (etc).
• Processing nodes – Intermediate steps that manipulate or evaluate the data. This might include transformations, logical conditionals (e.g., “if this, then that”), enrichment from external data sources, and routing logic.
• Output actions – Once processing is complete, the workflow sends the final data to its destination, which could be a content management system (CMS), a spreadsheet, an email notification, a database entry, or even an API call to another service.

For example, consider a basic linear workflow handling incoming emails:

1. An email arrives in an inbox.
2. The system extracts the attachments and stores them in a database.
3. The metadata (such as sender, subject, and timestamp) is logged.
4. A notification email is sent to a relevant user to inform them of the new stored attachments.

This system is entirely deterministic and rule-based, following predefined steps. It automates a tedious process, reducing manual effort and increasing efficiency, but it doesn’t adapt or interpret the content dynamically.

Examples of Linear Workflow Automation Tools

• Make (formerly Integromat) – automate and control over how data flows between services.
• N8N – An open-core workflow automation platform that allows for more customization and self-hosting options.
• Flowise - similar to N8N but entirely open source

These tools have transformed how businesses and organizations manage workflows, reducing manual effort and improving efficiency by automating routine processes. However, until recently, these systems were entirely rule-based and followed strictly linear, pre-defined workflows.

The Integration of AI into Workflow Automation

In the last year, AI has started to interweave into these deterministic workflows, introducing probabilistic, adaptive decision-making into previously static processes. Workflow automation tools like N8N and Flowise have introduced AI nodes, which expand the capabilities of linear workflows by enabling:

• Dynamic data analysis – AI-powered nodes can classify, summarize, and extract meaning from raw input data before passing it further down the workflow chain.
• Decision-making with machine learning – Instead of simple “if-then” statements, AI nodes can weigh different factors, make probabilistic choices, and even adjust workflows based on contextual understanding.
• Generating outputs dynamically – AI nodes can create reports, draft responses, or even generate structured data in response to input triggers.

Using the same email workflow example, but now incorporating AI:

1. An email arrives in an inbox.
2. The AI interprets the content of the email, classifying it into categories like invoices, job applications, or general inquiries.
3. The AI determines whether the attachments contain key data (e.g., invoices with missing fields or contracts needing approval).
4. If necessary, the AI routes the email to specific recipients based on content (e.g., finance for invoices, HR for job applications).
5. AI drafts a summary email with recommended actions, which a human can review before sending.

This integration of AI shifts workflow automation from being static and rule-based to adaptive and context-aware. The system can now handle ambiguous inputs, make smart decisions, and even generate responses, vastly improving efficiency while reducing the need for constant manual oversight. Agentic approaches can further enhance this by introducing AI agents that actively iterate towards specific goals rather than merely responding to inputs (for a good definition of AI vs AI agents there is a very good post Anthropic, the article classifies an agentic AI system as one that not only processes and generates text but also refines its outputs through iterative reasoning, adjusting based on context to achieve a desired outcome). This transforms AI-driven automation into a more dynamic, problem-solving system capable of self-correcting and optimizing workflows.

The Human-in-the-Loop Challenge and the Future of Publishing Workflows

While automation significantly improves efficiency, many workflows still require human oversight and decision-making at key points. This is where Human-in-the-Loop (HITL) processes come in. Rather than executing tasks entirely autonomously, AI-driven workflows pause at critical moments for human validation, modification, or approval before proceeding.

A major challenge in this space is that while signal processing workflows are highly valuable, the way humans are currently integrated into the process remains clunky. Most implementations rely on existing technologies like email to surface content for review and trigger the next step in a workflow. This often means AI-processed outputs are sent via email for validation, and responses to those emails determine how the workflow proceeds. While functional, this approach is clearly not a sustainable long-term solution.

Relying on 'pre-AI' tools like email for human oversight introduces friction and inefficiencies. As workflow automation evolves, the need for seamless, purpose-built HITL integration is becoming increasingly clear. Instead of forcing human input through disjointed communication channels, the next generation of workflow systems must enable direct, intuitive engagement within the process itself. Developing smarter, more integrated validation tools will be key to making AI-enhanced workflows truly effective.

Publishing today still depends heavily on manual processes, with workflows fragmented across CMS platforms, spreadsheets, wikis, and email chains. Many of these steps require human judgment, from submission triage to peer review coordination and editorial decision-making, making them ideal candidates for intelligent automation.

This creates a prime opportunity for workflow automation tools. If I were to predict the evolution of these AI enhanced generation of tools in publishing I'd say there are three stages:

1. Initially, many manual publishing processes will be automated by specialized AI-enhanced HITL automation tools like N8N. These are fast and cost effective to assemble and importantly, easy to modify and extend in the fast changing technical world of AI.
2. As these approaches demonstrate their efficiency, legacy platforms will likely attempt to incorporate them—though many may struggle due to their rigid architectures.
3. Ultimately, the large-scale manuscript management systems developed during the transition from paper to digital will give way to a new generation of HITL automation solutions. These advanced tools will redefine publishing workflows, blending AI-driven efficiency with essential human oversight to create a more streamlined, intelligent, and adaptable publishing ecosystem.

The Future of Workflow Automation in Publishing

AI-driven workflow automation is evolving beyond rigid, rule-based systems into adaptive, context-aware processes that integrate human oversight seamlessly. By leveraging signal processing paradigms and agentic AI approaches, these systems can dynamically interpret inputs, optimize decision-making, and iterate toward specific goals. As automation advances across sectors, from publishing to scientific research and beyond, these AI-enhanced workflows will play a critical role in reducing manual effort while ensuring precision and adaptability. The future of automation isn’t just about efficiency—it’s about creating intelligent, human-aligned systems that enhance productivity and decision-making at every level.