Efficient at the wrong things

The last 12 months for Omira’s internal processes have been an exercise in deciding which processes require AI attention, and which remain ‘human-first’. I didn’t want us to become efficient at doing the wrong things. The result is that we automated less than people assume, but the steps we did automate made us both quicker and better; our proposals take a fraction of the time they took on PowerPoint, and they land better.

What we automated

Processes within a marketing and design studio like Omira can be repetitive and time-consuming. After some time keeping an eye on the bottlenecks in our delivery, sales and admin systems, I got an idea of where I should start looking to implement AI solutions. I found opportunities in our reporting, sales, and onboarding processes and started implementing solutions from the ground up.

The AI layer in our workflows is often smaller than people imagine. We found that oftentimes workflows worth ‘automating’ are made up of a series of predictable steps, which we trust deterministic code with rather than reaching for AI. The AI may sit in gaps between some steps which previously required human action (not human approval - important distinction). At Omira we did not replace human approval with LLMs in these systems, as I decided that the responsibility for these tasks still lands on the owner of the process, as it always has.

When we stopped

When AI processes started to ‘work’ it was natural for us to be tempted to expand the reach and look for new use cases for these tools. I had to decide where to draw the line, before we started spending money and time automating things that we didn’t actually need. This ‘LLM creep’ is something I intentionally look out for before starting a project. For us, it meant that some of the new workflows are ‘AI assisted’ rather than ‘AI native’, as a member of our team remains the owner of the process, as well as the result. Our reporting system still requires a human trigger; the account owner writes in their qualitative results and thoughts, and only then does the AI take over.

What we didn’t build

Some processes remain as they were, human-first at every step. The clearest refusals are not necessarily separate systems - they are parts of the built ones that don’t exist. Features I scoped, could have built in a couple of days, and didn’t - because they automated something nobody needed automating.