On the Topic of Workflow Autonomy

Yesterday I read a really interesting article (apologies for the portuguese) that prompted me to write this piece. The author made a distinction that's been rattling around in my head: the difference between infusing a process step with intelligence versus deploying an autonomous workflow intelligent to the point of being capable of deciding its own steps.

This is a big distinction.

Throughout my interactions with Intelligent Automation leaders, I've been noticing a theme. When asked if they've started infusing AI into their programme, the most recurrent answer is some variation of: "We've added LLMs on specific steps on our automations" or "We're using AI for document extraction in our IDP workflows."

This pattern is consistent across organizations, regardless of their automation programme maturity level. Everyone's starting with the same playbook: take your existing automations, identify the step that handles unstructured data, and drop in an LLM. AI checkbox: ticked.

This low-hanging fruit makes sense to me. It's contained, measurable, doesn't require massive methodology redirection, and most importantly builds on the skills you have within your automation team.

We've previously made our AI thesis public that AI will play itself out in Intelligent Automation programmes in three ways: a) new capabilities (via new features and use case possibilities), b) as a new delivery tool (requiring tailored discovery approaches, skillset and methodologies) and c) as a dramatic redesign force of operating models.

Either because LLMs were made available in their organization or because their preferred tools now allow them to leverage AI as new functionalities, the point made above is that AI is being used as capability augmentation. Not yet as a standalone delivery tool or as an operating model redesign.

This, of course, brings us to the point of autonomy. "Agentic AI," which seems to be the buzz of the moment, has led all major Intelligent Automation players to rebrand. One year ago, Druid and Cognigy were conversational AI platforms—now they're agentic AI platforms. UiPath, same same, and so on.

If you define agentic AI (in the context of automation) as goal-oriented workflows with a high degree of autonomy, I'd argue that most Intelligent Automation programmes aren't yet there. I'd also argue that perhaps the number of use cases suitable for true autonomy is smaller than the marketing hype suggests.

One of the final points of the article I mentioned, which I agree with, is that embedded AI capabilities might actually be a force for a new wave of Intelligent Automation implementations with old horses of the toolstack—RPA and BPM leading the pack in our eyes.

This is because it's way easier to build guardrails and orchestrate AI with predictable Intelligent Automation tools than to do a pure play AI project, which is still missing many of the functionalities needed to scale Intelligent Automation. Starting with trust and monitoring capabilities.