Why We Stopped Prompting
and Started Building —
The Story Behind Ultra Studio
The previous two articles made the case for moving from subscriptions to API, and addressed the objection that most people cannot build what we built. This one goes inside the actual problem we were trying to solve — and what building the solution has looked like from the inside.
The real problem we were solving
The cost problem came first. Five AI subscriptions across the team, flat monthly billing, usage that never justified the rate. I wrote about that.
But underneath the cost problem was a deeper one that the billing dashboard did not show. And it was this: every piece of content we produced required a human to manually carry work from one tool to the next.
A staff member would open Claude to research and draft a script. Then open ChatGPT to refine the tone. Then open Gemini to generate image prompts from the script. Then open Seedance or another video tool to generate the visuals. Then come back and stitch it together. Four tools. Four separate browser tabs. Four separate sessions with no memory of each other. The human in the middle was the integration layer, doing nothing but copy, paste, and context-rebuild for a significant portion of their working time.
This is not a small inefficiency. If a team member spends 30% of their AI-assisted workflow time doing manual handoffs between tools, that is 30% of their productive capacity going into work that produces nothing — no output, no quality, no thinking. Just movement of information between platforms that should have been talking to each other already.
We were also hitting a quality ceiling. When context does not carry across tools, output quality is anchored to the context you can manually re-establish in each session. A staff member who becomes skilled at prompting Claude will produce better work from Claude. But the moment they switch to a different tool for a different part of the workflow, they are starting from zero again. The compound value of a tool that knows your brand, your standards, your previous work — that compound value resets to nothing every time you open a new tab.
The decision: buy vs build
The obvious question at this point was: can we buy something that solves this?
We looked. There are products that wrap AI models behind a unified interface. There are workflow automation tools that chain prompts together. There are content platforms that plug into APIs and give you a dashboard. None of them were wrong. Some of them were genuinely useful for specific tasks.
But none of them solved the specific shape of our problem. The workflow we needed was not generic — it was built around the exact content types IMA AI produces, the exact quality standards we hold, and the specific sequence of steps that takes a brief from an idea to a finished piece of content ready for distribution. A general-purpose tool would solve 60% of it and leave the rest back in manual territory.
We also came back to the ownership argument. If we built our workflow inside a third-party platform, we were dependent on that platform's product decisions, pricing changes, and feature roadmap. We had already learned that lesson with subscription lock-in. Building our own meant our workflow was ours — it could be changed, extended, and adapted as our needs changed, without waiting for a vendor to prioritise our feature request.
So we decided to build. Not because buying was impossible, but because the thing we were building was specific enough to our operation that buying it would have meant accepting significant compromise on the parts that mattered most.
What Ultra Studio is
Ultra Studio is an internal automation platform built on API. It is not a product, and it is not available publicly — it is the infrastructure that powers IMA AI's content and generation workflows.
The core idea is a pipeline: a brief goes in one end, and the system carries it through to finished output — script, image, video — without a human acting as the integration layer between steps. Each stage of the pipeline is driven by a model chosen for that specific task. The output of one stage becomes the structured input of the next. No copy-paste. No context rebuild. No switching tabs.
The pipeline is built on API access to multiple models — the best available model for each specific task in the workflow, not a single model trying to do everything. Language tasks go to the model that handles language best at that moment. Image generation goes to whichever model produces the right output for the visual style we need. Video generation goes to whichever model can handle the assembly.
Because we own the API layer, switching models when a better one arrives takes minutes, not months. We change one configuration. The workflow continues. No platform migration, no retraining the team on a new interface, no waiting for a subscription service to support the new model.
What works today
The unified access layer is fully operational. The team no longer maintains separate subscriptions or logs into separate platforms for different tasks. All model access flows through one internal system, billed at API consumption rates. The monthly cost is lower than the subscription stack it replaced, and it scales accurately with actual usage rather than headcount.
Context consistency has improved meaningfully. Brand voice, quality standards, and workflow history are maintained inside the system rather than in individual staff members' prompt habits. A new team member does not need weeks to learn which prompts produce the right output — the system carries the institutional knowledge.
Script generation is running well. Given a brief, the system produces a structured script suitable for the target format — short video, long-form content, social copy — with the right pacing, tone, and structure already in place. The gap between brief and usable draft has shortened significantly.
What we are still building
The full pipeline — brief to finished video, no human handoffs — is not complete. Script generation works. The visual generation stage is functional but not yet integrated into a seamless pipeline with the script stage. The video assembly step is the furthest out. Each model in the video generation space has specific strengths and specific limitations, and finding the right way to structure inputs so that the output is consistently usable at production quality is still work in progress.
This is the part I want to be honest about. Ultra Studio is not a finished product. It is a working system with some stages running well and other stages still being refined. The direction is clear. The destination is a pipeline where a brief produces finished content without a human acting as the glue between stages. We are building toward that, and we are not there yet.
An open door
These three articles — on cost, on accessibility, and now on what we are actually building — are not a sales sequence. There is no product to buy at the end of this. Ultra Studio is an internal tool and will stay that way for the foreseeable future.
What these articles are is an honest account of how one company thought through a problem that a lot of teams are facing right now: how do you use AI at scale, at reasonable cost, with consistent quality, without a human doing manual integration work all day?
We do not have all the answers. We have a direction and some working components and a lot of problems still being solved. If you are working through similar questions — whether that is the cost structure, the infrastructure decisions, or the pipeline architecture — feel free to reach out. Not for a pitch. Just to talk. Some of the best thinking on this comes out of honest conversations between people who are building through the same fog.
The contact details are below. No form. No funnel. Just a conversation if you want one.
Published by IMA AI — June 2026. Part three of a series on AI operations and infrastructure decisions.