My grandmother made the best cornbread in the world. Am I biased? Yes; but that doesn’t mean I’m wrong. I’ve been trying to recreate it for longer than I care to recount. Because she never wrote down the recipe. In fact, there really wasn’t a recipe… a cup of this, a cup of that, a dash of salt… you get the idea. It made for some memorable cornbread but also a big challenge for me to replicate.
While losing a family recipe can feel tragic, losing a “recipe” for a critical business process can be destructive for an organization trying to transform a process or to even hand it off to another person.
One of the challenges and benefits of trying to automate a process, through AI or other technology, is that you are confronted with what you don’t know, with the lack of context around the processes that your team is running every day.
As we’ve implemented RELIability AI this year, we have uncovered nuances to the work that people are doing every day that weren’t captured on the official workflow chart or in the requirements documents. These gaps weren’t intentional. Some of that knowledge, like my grandmother’s recipe, was never written down. People simply knew what to check, when to ask a question or how to handle an exception.
Some of it is the result of processes that changed and evolved slowly (and sometimes rapidly) over time in ways that didn’t always trigger an update to the documentation. Over time, those changes created a process that didn’t quite match the original requirements.
And if we’re being fully honest with ourselves, EVERY process does this. Nothing stands still. Processes change and that’s OK. But our understanding and documentation need to keep up, too.
This is one of the aspects of agentic AI that excites me about the future. Implementing agentic AI forces you to reconcile what is actually happening with what you think is happening. The people doing the work are essential to understanding that delta. They know what happens between the documented steps and why those extra actions became necessary. Often, implementing AI forces us to ask questions that we didn’t even know we needed to ask.
For example, during our implementation, we uncovered an input process that was documented but wasn’t standardized. Someone had to interpret the information coming in, understand how to process it and then massage it into a form the system could use. The documentation captured the input, but it didn’t capture the variability or the work someone was doing to make sense of it. On paper, the step was accounted for. In practice, it depended on someone knowing what to do with whatever arrived.
Discovering that work gave us more than another requirement. It gave us questions to answer:
- Could the input be standardized?
- Could AI help interpret the variations?
- Where would a person still need to review it?
The people doing the work were essential to understanding those choices. Without their context, we would automate the documented process and still miss important parts of the work.
Once implemented with agentic AI, similar future requests may be easier to accommodate, sometimes by adding a new instruction or prompt into the workflow. For some changes, adapting the system may largely be a matter of providing the right data and asking for the desired output.
Don’t get me wrong, there is still rigor and complexity to do this. We still need to understand the request, provide appropriate access to the data, approve the change and QA the output to ensure it meets the need. But we have more flexibility to incorporate changes into the system rather than adding another manual step outside it.
The system can also help document what it does, but people still need to capture why the process changed and who approved it. Used this way, agentic AI can help us keep the process and its documentation aligned as needs evolve, reducing the need for one-off undocumented manual processes that erode the effectiveness of a system over time.
Implementing agentic AI intentionally helps us reveal the hidden steps and context in our existing processes. A well-designed platform can then help us better adapt to changing needs over time. By putting in the initial effort to find the missing ingredients, we create a recipe that can be understood, automated and improved as needed.
In the meantime, I’ll continue trying to recreate that cornbread. While that’s one thing I’d never automate, it would be nice to have the recipe.