Why is RPA a Good Starting Point on the Journey Towards AI?
Updated: Oct 22
RPA is a practical starting point for AI automation because it establishes repeatable workflows, structured inputs, exception handling, and clear controls. AI can then support tasks such as extracting information or reviewing documents within that workflow. My Power Automate examples show how this sequence works in procurement.
"He who would learn to fly one day must first learn to stand and walk and run and climb and dance; one cannot fly into flying." Friedrich Nietzsche
We’re often told that mastering prompt engineering is the key to thriving in the AI era.
But in my experience, the real power to optimize and streamline business processes has been at our fingertips for years.
The RPA Alert
This potential is called RPA, Robotic Process Automation.
The hype around RPA started about a decade ago, and people predicted robots would take over back-office jobs. Around the same time (November 2016), Microsoft launched Flow, which later became Power Automate.
How many of us knew about Microsoft Flow at the time? Even fewer actively used it.
Nevertheless, the "robot takes your job" alerts continued to flash until AI entered the picture. Now, we’re told to fear AI.
It's not the Revelation, just an Opinion.
In my Power Automate projects, much of the work involves triggers, conditions, approvals, data handling, and exception management. AI contributes where the process needs document interpretation, classification, extraction, or content generation. This is an observation from my projects, rather than a measured industry benchmark.
A terminology distinction matters: Power Automate supports cloud workflows and desktop RPA. Connecting SharePoint, Outlook, and approval actions through connectors is cloud workflow automation. Automating interactions with application screens is desktop RPA.
My recommendation is to understand the process before adding AI. Define the inputs, decision rules, responsibilities, and exceptions, then identify where AI offers a measurable improvement.
Check licensing against the actual design. Microsoft 365 access alone doesn't mean premium connectors, desktop automation, or AI processing are covered.
JSON news for Prompt Learners
Learning the art of prompting is essential, but not sufficient.
Learn enough JSON to recognize objects, arrays, and name–value pairs. JSON is a data format for exchanging structured information between systems. In Power Automate, distinguish the JSON that carries your data from the expressions that manipulate it.
I started with Copilot and ChatGPT by my side, which helped me structure JSON expressions, which can be particularly tricky because of their complex syntax rules.
Dynamic AI prompts
If you want AI to really help out with your workflows, your prompts need to be dynamic. This means they should change and adapt based on new inputs in your process.
Check out the screenshot below—it shows a prompt with four dynamic elements that I've used within my flow.
RFQ Evaluation Automation with Power Automate and AI Builder
This case blends JSON, multiple variables, arrays, and content controls. It’s genuine automation with a touch of AI, where I trained a model to extract data from standardized Word forms.
You’ll need to:
Tailor it to AI’s need for structured data
Manage GPTs to help assemble the flow and navigate the JSON intricacies
Run numerous test loops
Safeguard the process logic
Before you aim to fly, ensure you’ve mastered walking. Let’s not rush into AI without first grasping the fundamentals of process automation.
Budget for AI Processing
Running AI Builder models and prompts consumes AI Builder credits or Copilot Credits, separately from the applicable workflow licensing.
From 1 November 2026, new or renewed eligible Power Platform and Dynamics 365 licenses will no longer receive seeded AI Builder credits. Qualifying licenses purchased or renewed before that date retain their seeded credits through the applicable contract term.
Estimate consumption using representative documents, including retries and repeated evaluations.
P.S. Intelligent Automation.
When I wrote the above, I tried to ixplain Intelligent Automation intuitively.
It's nice that such a concept exists. Yet, it suggests that we have to develop ourselves in a balanced manner, not just getting overweight on the AI side.
P.P.S. Tender Committee Process Automation
If you think RPA is just about replacing keystrokes, look at how we run our Tender Committee workflows:
TC Entries runs 24/7, automatically classifying each submission and notifying the right people without anyone opening Outlook.
This Power Automate process monitors the SharePoint “Tender Committee (TC) Entries” list for new or changed items. Depending on the item’s state, it sends one of two transactional emails: 1) New record created → “Recorded for TC” confirmation to the SCM Officer (CCs the TC Administrator). 2) Status changed to “Scheduled for TC” (and a TC Date exists) → “Scheduled for upcoming TC” notice to the SCM Officer and the End-User. Agenda Formation pulls the next meeting’s cases, formats them beautifully, and pushes a draft to the Organizer for sign-off — no more “where’s the latest agenda?” email chains.
This Power Automate process prepares, circulates, and finalizes the agenda for the nearest upcoming Tender Committee (TC) meeting. It pulls all TC entries scheduled for the next available TC date, formats them into a styled HTML table, sends the draft to the Organizer for approval, and, upon approval, distributes the final agenda to all TC voting members while archiving it in SharePoint. Approval by Circulation runs full-cycle remote voting: requests, responses, tally, decision, recordkeeping, and summary distribution — all hands-free once triggered.
These examples are rule-based cloud workflows. AI could supplement them by summarising submissions or flagging missing information, while the workflow retains the defined routing, approvals, and recordkeeping.
That’s the sequence: automate first, then augment with intelligence. Without the RPA backbone, AI adds noise. With it, AI becomes a lever.
Comments