Scaling AI as a Culture: A Cellular, Semantic Model of Standardizing Human to AI Interaction

I remember trying to convince one of my executives to operationalize everything into AI. Everything. This process took a trip across the Atlantic to Scotland and three months of culture investment. The payoff, as painful as it was in the moment, has led to a 60% drop in operating costs as he codified his expertise into agents that can scale and operationalize.

This situation taught me that as fast as AI models improve, such as the recent Anthropic releases into HIPAA-compliant healthcare and workstations, a primary challenge remains: how do I scale AI in an organization with transparency, consistency, and repeatability?

To solve this challenge we realized that integrating AI wasn't enough—we needed a shared language. So we built Somata: a dialectic semantic model that transforms individuals into one-person innovation labs, each owning their own discovery, design, and development pipelines. We reduced task half-life so dramatically across our domains that a designer, an engineer, and a pathologist built what would have been three separate companies into a 100% bootstrapped and unified venture: pathology and diagnostic synthesis, government contracting, and operations.

A Unified Semantic Language

A Cellular Language for AI

Semantics is the language in which we describe the world. Design, engineering, finance, and others are all ways to deconstruct complexity into manageable problems to analyze and solve. One of the key challenges I found is that everyone projects what AI is, and this informs different usage behaviors. Fine at a personal scale, difficult at enterprise and company levels where everyone needs to move toward the same vision and end states.

The language we landed on comes from systems biology and medicine—in other words, a cellular structure for data. Starting as early as middle school, we are taught the building blocks of the body—both in the abstract and the body as something we live in every day. Unlike AI, where definitions vary widely and even "using AI" means different things to different people, there is a common understanding that a "cell" is a building block towards something.

Based on this shared understanding across humanity, we built a way to structure data in this cellular interaction model.

We organized company information into Somata, or large bodies or concepts of information, such as accounting, compliance, and contracting. Cells are related themes inside that Somata, so when a prompt is written, a transformer can pick up key words and access only the data it needs to most efficiently execute the task. One of the metrics we track (more on AI metrics later) is number of prompts to task completion. Over time, as the prompt writer learns to include key Somata terms and Cells, the overall task completion rate efficiency goes up. Simultaneously, the model gets better at accessing the right data set over time.

The diagram of the process is shown below.

The goal of this efficiency then creates what we call internally a half-life of tasks. As AI gets more efficient and the human prompt writer pushes the edge of what is possible, the actual swath of tasks AI can do increases—both as a function of the human pushing their understanding of AI limitations and the better shared context across an organization.

The Modern AI Company

Mastering Dialectic Learning Loops

The result of this semantic learning has been remarkable, with operational savings approaching $1M for a small startup. We've become so efficient that all business operations are run by one person—saving approximately four headcount for an emerging startup. All of which has been reinvested into accelerating this model and its outputs.

As founders, we sat down for coffee at a local café next to the office and sketched out a simple operating north star for a modern AI company: go through as many learning loops as possible to the edges of task completion. In metrics terms, reduce the half-life of tasks in a given expertise as much as possible. As a three-person core team, we arrived at the following process that derived the best results:

Goal: Codify, scale, and optimize the Somata model's task completion and accuracy by operationalizing each founder's expertise.

Dialectic Learning Loop:

1. Map core competencies

   Identify the Venn diagram of skills across the core team—design, engineering, and pathology.

2. Structure daily work streams

   Develop clear task structures for each daily workflow.

3. Push model boundaries

   Extend the edges of what tasks models can handle until reaching the conclusion: "a human needs to do this."

4. Assess model boundaries and human necessity weekly

   Evaluate boundaries through cross-expertise syncs:

   • Was this really an edge?
   • Have we reached a point where we truly need to hire?
   • Where is model development currently, and what's the trajectory? Could this be solved in a month, or will it require more time?
5. Semantic Reports

   Daily reports aggregated for the weekly syncs measuring productivity:

   • Tasks completed
   • Number of prompts for task completion
   • Code generation—specifically important to baseline non-technical team members
   • Model speed
   • Assess prompt writing changes
6. Iterate until tasks shrink

   Continue the learning loop until the number of tasks within each domain expertise decreases.

The results were tremendous:

What these metrics show is that joint learning was taking place with varied acceleration. The engineer, better versed in AI modeling and its boundaries, reduced task half-life much faster. Over time, the designer and pathologist caught up as their shared context through the Somata the engineer architected improved. In summary, we were equally testing our assumptions of AI limits at varied paces. This dialectic learning between humans and AI within a shared context is what accelerates the productivity, innovation, and efficiency of both.

The weekly syncs to test and compare progress were crucial in validating edges (at model limits) and false edges (a boundary or limitation of the person's ability to engage with AI). The first two months were understanding the Somata structure across all major tasks. After that golden cross, the acceleration was immense. We had three people running a multi million dollar revenue company across highly complex industries such as healthcare diagnostics and federal contracting.

AI in the Market

Semantic Model to Competitive Advantage

Even with these results, the question on the top of our minds is: what does this mean for competitive advantage, and how does an AI-first company convert efficiency into real market advantage?

Across the landscape, already shown, we are dawning on an age of small person companies delivering tremendous market value. The implied value of this in traditional metrics would be lower operating costs, higher revenue per employee, and better margins. These savings and margins can be reinvested into innovation.

• Product-market fit. Building cheaper extends runway—giving us more time and iterations to find a product that truly resonates.
• Culture building. Operating this lean creates space to invest in defining what work looks like for a company that doesn't exist yet.
• Market capture. Delivering faster means we can serve more customers and expand our footprint while competitors are still scaling teams.
• AI adoption. The time we've freed up compounds when reinvested into making our AI integration even more effective.

Can AI-first companies achieve all of these at once? Yes, all of it. AI reduces the need for and raises the bar on tradeoff conversations. If you have 40% unlocked efficiency as a company and can do this with 20% less people, the need for compromises reduces and the conversations for innovation should increase. In other words, less opex and more creative productivity that generates more service innovation, adoption and marketshare.

As a designer myself, I often say that design is engineering desirability. If anyone could write a prompt and make a billion dollar product they would. It's more accepting that both the definition of what is desirable is changing while also accepting key components of it hasn't. The term "AI slop" has been widely popular lately—AI-generated content and apps that are terrible. To us at Somata, the goal of AI is to capture the problem solving of highly intelligent and creative people and help them loop through faster innovation curves towards higher quality solutions. The output of that cycle is "un-restricted one person innovation laboratories."

Measuring What Matters

Metrics as a Dialectic

Throughout this case study, we alluded to a few metrics, intentionally leaving these to the later part. Primarily because many processes for KPIs and metrics are very human-oriented. I, or we as a company, set a vision, set KPIs and metrics to that vision, and measure and adapt them. There is still a very important place for this. I often say that one of the most important jobs for executives is to be right, and being right is really hard.

At the same time, as we were working through Somata and this semantic language, we found that metrics themselves were their own semantic language. In the model of business, especially at scaled organizations where Human Capital is the most critical resource, KPIs and metrics are the orienting mechanism to direct human productivity toward singular outcomes.

In an AI-first company—and we hypothesize that AI-first companies will continually make up a larger share of the market—KPIs and metrics also become a dialectic. It becomes less about what I can get from AI and more about whether I can create models where I can consistently trust AI to understand the core KPIs and metrics for my company and industry.

Not defer, but trust. Imagine that you have a model you can trust to drive an increasing percentage of your KPIs and metrics, even just 10 or 20% lift. This means many fewer meetings, workshops, executive and organizational alignment sessions, and related activities.

This uplift translates to faster market adaptability. For example, with Somata we arrived at a singular KPI: Build Somata into a model we can trust increasingly over time for all development, design, and business operations.

As trust goes up, all measurements improve that we mentioned throughout the case study: speed, efficiency, accuracy, and task half life. These metrics have changed and evolved over time, and we found that our conversations and ability to vision also accelerated. Here are some examples:

Somata at Scale

A Unified AI Culture

To create and facilitate this shared context we built an internal application called Somata. This interface allows one person to move seamlessly between code, prompts, and instant app interface output. All of which condenses the product development cycle into one person. Each individual at Somata AI is their own designer, engineer, and product manager.

The power of this at scale is threefold:

1. Measurable engagement model. You create a measurable and understood engagement model with AI across a scaled organization, in comparison to micro-cultures or varying projections of what AI actually is.
2. Clear metrics dashboard. You get metrics that measure trust in your model over time and what that actually means across operational, strategic, and execution layers of your organization.
3. One unified context. Since everyone in your organization is working on one application, there is exponential growth in the context, leading to better outputs, and thus outcomes.

Where models are increasingly becoming the competitive advantage, Somata allows you to be an AI-first company.