Meet Frank: How Agentic AI Shifted My Service Model from Linear to Parallel

I’ve been running an experiment.

Frank is an AI agent running via OpenClaw in a container, on a dedicated Fedora laptop sitting on my desk. While this started as an experiment, the use case was real: Could I enable Frank to be a true digital assistant for my photography business, Trail Nutz Media?

When we talk about AI use cases, we often forget a core truth: AI is successful only when we have people who have actually done the job, who recognize when the job is done well, and who understand when it has gone sideways.

I tasked Frank with three specific tasks to see if I could shift my business from a manual, linear grind to a parallel, automated service model.

  1. The Digital Concierge First, Frank had to be able to interact with people, understand our brand voice, and know where our assets were. I configured Frank with read-only access to our style guides and processes stored on Google Drive.

Then, we had to give Frank a way to communicate. We went with texting, powered by Twilio. Frank, utilizing the Google Gemini Pro model, was instructed to respond politely, ask what the potential question is, and use the artifacts in Drive and our public website to answer based strictly on the facts he has access to. If he couldn’t answer, he was to send an email to me with the contact information and advise the customer we would be in touch within one business day.

It performed this task exceptionally well. I was even able to configure API capabilities to our CRM running in a Notion Kanban board, so if a question was about booking a session, a new entry was automatically made.

  1. The First-Pass Review Next was culling a photo gallery. I take action photos; I routinely shoot 1,500 photos and need to cull that down to a usable 150-200.

In order to cull, you need to review a photo and decide if the intended moment was captured. Is the primary subject in focus? Did you inadvertently chop off a foot or a hockey stick? If the photo wasn’t quite level, could it be fixed by cropping and changing the angle slightly? Some of these are highly subjective, and an AI agent can’t technically “see” a photo the way a human does.

To tackle this, I uploaded the full gallery to a locked-down backup space. Frank was given access to this directory, instructed to make a copy of the master, and use the Google Vision API to inspect each image and make the decision to cull or not. I told Frank to be conservative—if it was not a clear cull selection, leave it. It cleared the noise, allowing me to focus only on the creative selections.

  1. The Proofing Engineer Lastly, the final piece: preparing a gallery of proofs for customers.

When we receive orders for photos, we edit each photo to ensure proper calibration of colors, highlighting key elements like eyes and smiles, and just uplifting the photo overall. We do a lot of hockey photography, and the lighting in rinks is notoriously bad. If you’ve ever seen a yellow-tinted shot of a hockey game, that’s the bad lighting in action.

I wanted Frank to do three things: review each image to correct the white balance (white ice, not yellow), lift the exposure (not too dark, not too light), and change the file name to “Player Number - Player Name.jpg” using the Vision API.

The white balance correction and exposure uplifting? Chef’s kiss, perfect.

But the reality of AI is that it’s not perfect. The problem? The kids. It’s always the kids. Between folded jerseys and kids tucking the back of their jerseys into their pants, a 4 can look like a 9. A 15 could look like a 1 or a 5. While this still saved significant time, it requires a manual human review—and sometimes even the human can’t figure it out.

The Business Impact From a business perspective, this has massive value. It lets me focus on what’s core to the business—capturing high-quality action photos—and less time performing tedious tasks.

More importantly, this is a shift in Unit Economics and Service Velocity. The turnaround time and the labor cost going into delivering a job are way down. I can now deliver multiple jobs almost perfectly parallel in post-processing, where before it was strictly linear depending on the operator performing the tasks.

The Cautionary Tale This all sounds great, but it is a cautionary tale. It is truly incredible the power AI is bringing to our businesses, but the reality is that it’s not magic. I was able to build and apply this AI to the photography business because I intimately knew the process and flow. I know what “good” looks like.

Stay tuned on the Frank journey. We still have to dig into more of what “Enterprise Grade” governance looks like for these workflows.

This is Part 2 of my “Frank” Series. Check out Part 1, AI Velocity, Governance and Breaking Points, Oh MY!

Photo Credits: Banner photo is from Trail Nutz Media. This is a photo of my son eating it, after making a poor on-ice decision. He is fine, only his pride was hurt.