Most silos are unintentional, and they exist because your company is partially made out of meat1. In this post, I’ll explain how information fragmentation arises inevitably when humans are performing critical organizational functions, the negative consequences of that fragmentation, and propose an alternative organizational structure: the intelligence layer.
Meanwhile, the en vogue approach is to build “assistants” or “copilots” by connecting them to existing data silos with the intention of helping an individual human perform their narrowly scoped role. It’s easy to build toys like these, and easy to adopt them within existing company processes. Ultimately, this actually results in three kinds of silos: data, expertise, and intelligence.
To see why that assistant model will fail, and why the intelligence layer is inevitable, let us take a lesson from history.
History Rhymes
Prior to the Industrial Revolution, humans performed manual work in factories and farms, operating tools designed for humans (e.g., sewing machines and hammers). We devised machines that performed those same tasks better, faster, cheaper, and with fewer errors.
Notably, these machines not only replaced the humans but also the tools and interfaces those humans required. We didn’t build humanoid robots that operated sewing machines; we didn’t build an assistant that helps a human operate the sewing machine; instead, we built machines that directly performed the task.
Today, echoing the pre-Industrial story, humans perform manual knowledge work in offices, operating computer interfaces (UIs) designed for humans. They are gathering and consuming information; synthesizing and analyzing that information; and generating new information that they put back into the computer—all of these functions are being performed with increasing competence by AI.
Some humans who were replaced by machines during the Industrial Revolution took on the work of designing, building, operating, maintaining, and improving those machines. As AI begins the decades-long process of replacing humans that perform tasks better done by machines—indeed eventually transitioning humans out of all operational functions2—we can begin to shift the focus of human labor to the task of creating a machine that is the company.
Welcome to the Machine
A company is a machine that makes money by producing and selling goods and services. Today, many of the critical functions of that machine are performed by humans.
Humans have intrinsic limitations that prevent us from being experts at every topic or holding unbounded amounts of information in our heads. As a result, we specialize to serve narrow functions within the larger corporate machine. Each human becomes an expertise silo, with critical information and skills that only they can access.
The larger the machine, the more narrow the specialization. At a startup, you might have a full-stack engineer who touches every part of the product or an account executive who speaks with every prospect. As the company grows, an engineer might specialize first on backend and then on, say, containerization infrastructure for a specific cloud provider. An account executive might only target large insurance companies in the mid-Atlantic region.
Companies necessarily introduce tools and processes to coordinate human activity, limiting access to the context and information they need. Some humans are responsible entirely for performing this coordination for groups of humans with related functions; we call them “managers.”
Most of the tools companies introduce are human-usable interfaces (UIs) sitting on top of databases (e.g., a CRM or ATS). Typically, those databases are exclusive to the tool; each tool is a data silo. These UIs provide humans a way to request information, view it in a form they can digest, and input new information. A machine needs none of these interfaces, or indeed any UI, and can interact with the information layer directly. In a sense, most of the software that companies use is human overhead.
Even worse, because humans can be unpredictable and unreliable, companies introduce mechanisms to actively prevent certain information from being shared. Broad swaths of compliance and regulatory apparatus are concerned with controlling the flow of information between humans. Much of this, too, is human overhead.
Let's see how this situation is exacerbated by introducing AI.
Meat Makes Silos
Conway’s Law3 states that “any organization that designs a system (defined broadly) will produce a design whose structure is a copy of the organization’s communication structure.” When it comes to AI, a company that’s partly made of meat and its accompanying communication silos will build intelligences reflecting that siloed nature. That’s how you end up with a proliferation of “assistants” that are just as myopic as the humans they are built to help.
Consider a product manager who wants to prioritize feature work based on what would drive the most new revenue. They don’t have access to the CRM, however, because they aren’t working in the field. In fact, that data silo might explicitly charge based on how many humans have access to it, euphemistically called “seats.” So the product manager has to find someone who does have access to the data (say, an account executive) and ask them to help. That account executive may have access to the data but lack the data science expertise to write the query, so they pull in someone from the data analytics team. Now there are three humans involved: one with a question, one with access to data, and one with access to expertise.
The product manager’s eventual decision about prioritization came slowly, because it required the involvement of so many independent (meat) components. It was likely not globally optimal, because it was limited to the information available to the humans involved. It was prone to errors, because no one had a holistic view of how the decision was made. A mistake by one human would propagate outward. And, of course, it all relied on a human asking the right questions.
Humans are expertise silos. They use tools that create data silos. The “assistant” model of artificial intelligence introduces a third kind of silo: intelligence silos. This new silo explicitly codifies the interactions above as standard operating procedure, thus reinforcing the other two silos and inheriting the limitations of their meat predecessors.
Without human components, however, these silos can and should be avoided. The only silos should be intentional ones. We accomplish this with a novel approach to AI.
Introducing: The Intelligence Layer
An alternative company architecture is a shared intelligence layer. An intelligence layer consists of data (found in data stores, tools, and documents) and expertise (found in peoples’ heads), all on an AI-driven computational substrate that does the heavy lifting to find, understand, enrich, and apply this intelligence to drive business outcomes. The layer may be partitioned to isolate sensitive information, but the default is to share and accumulate intelligence in one place.
Now, rather than having to consult multiple datastores and humans, and doing most of the analysis and decision-making herself, our product manager above can simply ask the intelligence layer. It will pull relevant information from every source that matters, perform the necessary analytics, and generate the requested response. The account executive (or any other role) can do the same: ask the intelligence layer about other parts of the business that enable them to perform better.
As companies begin to offload more responsibilities to the intelligence layer, it transitions from being the invaluable tool for humans it is today to being the literal brains of the operation.
Preparing for the Inevitable
To be clear, I’m not anti-human. Some of my best friends are humans. But we are seeing the first examples of software-based alternatives to meat-cogs that can perform their specialized functions better, faster, and cheaper. Capitalism (and history) tell us how this will play out; it behooves us to begin preparing for that transition now and reaping major benefits along the way.
Critically, you can build and use an intelligence layer even while part of your company is made out of meat. Individuals can continue employing the tools they’re used to, while teams, departments, or the organization as a whole can begin to graft an intelligence layer on top of them. Indeed, doing so yields immediate rewards in the form of operational efficiency, knowledge continuity, and improved business outcomes.
At Graft, we built the world’s first intelligence layer. Connect it to your company’s information layer, enrich it with expertise, and apply it to run a better business and build better products. Under the hood, our intelligence layer is a comprehensive AI platform that’s agnostic to technology choices like which foundation model to use. That makes it possible to do much more than simply find a document or answer a question about a narrow slice of the business. An intelligence layer explicitly breaks down data and expertise silos while avoiding the trap of creating intelligence silos.
Unlike a specific technology stack or short list of use cases, which are vulnerable to uncertainty risks and short-term thinking, the intelligence layer is an AI strategy. Companies that adopt it think differently about how they approach these investments because they have a true north star.
We envision company-building to one day be a matter of literally building a machine, the same way that today we build software. Humans may be involved in designing, building, and operating that machine, but they are no longer employed as machinery. As the company hums along, generating value all on its own—with an intelligence layer at its core—I think we’ll marvel at the idea that, once, such a machine was partially made out of meat.
Footnotes
[1] Terry Bisson, 1991 They're Made out of Meat https://www.mit.edu/people/dpolicar/writing/prose/text/thinkingMeat.html (with apologies)
[2] Humans Need Not Apply Made nearly a decade ago and more relevant than ever.
.png){kind=icon}
