The Mediocrity Tax: Why the Economic Logic Behind Enterprise Software Has Finally Collapsed
The replacement is not a better suite. It is a different logic entirely.
This is the first post in a series called The Headless Firm. It draws on our paper: "The Headless Firm: How AI Reshapes Enterprise Boundaries."
If you have ever sat in a procurement meeting and signed off on an enterprise software suite you knew was not the best tool for the job, you were not making a bad decision. You were making a perfectly rational one. The system your organisation needed was not the most capable one on the market. It was the one that could be made to work with everything else already running in your environment, and the one your team would not spend the next eighteen months integrating. You paid for coherence, not excellence. Almost every large organisation in the world has been doing exactly this for decades, and almost nobody talks about what that has actually cost.
Why Enterprises Settled for Less
In 1937, Ronald Coase asked a deceptively simple question: if markets are so efficient at allocating resources, why do firms exist at all? Why not just contract out everything? His answer was that coordination is not free. Searching for a supplier, negotiating terms, verifying quality, enforcing a contract: these activities consume time and money. The firm exists as a mechanism for avoiding those costs by bringing activities inside a single organisational boundary. You internalise a function when doing so is cheaper than transacting for it in the market. This insight, which earned Coase a Nobel Prize in 1991, maps almost exactly onto the logic that built the enterprise software industry.
For decades, large enterprises did not buy best-of-breed software because best-of-breed software was impractical to operate. The cost of integrating fifteen specialised tools from fifteen different vendors was prohibitive. Every connection between systems was a custom engineering project. Every time two applications needed to share data, someone had to build and maintain the bridge. Every vendor update risked breaking something downstream, and the total coordination overhead was high enough that paying a premium for a single integrated suite was the rational choice, even when individual modules were mediocre.
This is what we call the mediocrity tax: the price enterprises paid to avoid the integration headache. The suite was, in Coasean terms, a digital firm — a walled garden that absorbed coordination costs internally and shielded users from the friction of the market. Vendors understood this and acted accordingly, expanding their perimeters through acquisition, bundling adjacent functionality, and engineering switching costs deep into their platforms. The value proposition was not excellence. It was containment. That logic held for roughly forty years.
What Large Language Models Actually Change
The common framing around large language models focuses on task automation: writing, coding, summarisation. That framing undersells what is actually happening at a structural level. What large language models introduce, for the first time at scale, is the ability to act as intelligent translators between systems that were never designed to work together. A model can take data out of one business application and make sense of it in the context of another, without anyone writing custom integration code. It can operate software the same way a person would, which means it can work with systems that have no programmatic interface at all. It can read a new regulatory requirement and reason about what needs to change in an existing process, rather than waiting for the next software release to incorporate it.
None of these capabilities are individually transformative. Together, they reduce the cost of coordinating between specialised tools — and that changes everything. This is the mechanism behind what we describe in the paper as the Coasean Singularity: the point at which transaction costs of coordinating with external, specialised agents drop below the cost of maintaining a monolith. When that happens, the economic case for the integrated suite inverts. The mediocrity tax stops being worth paying. The current repricing of enterprise software reflects this inversion. It is not a sentiment shift. It is the market catching up to a structural reality.
The Architecture That Follows
When Coasean logic inverts, the organisational form changes. The pyramid of the integrated monolith gives way to what we call the hourglass.
At the top sits the intent layer. Rather than learning a fixed application interface, users express a goal and the system works out the process needed to address it. “Optimise inventory allocation given the updated supplier lead times.” “Flag every contract that references the discontinued component line.” The application as a static destination with a fixed feature set is not a natural endpoint of software development. It is an artifact of a world where dynamic composition was too expensive to build.
At the bottom sits a competitive market of vertical micro-agents, each doing a narrow thing well and evaluated on outcomes rather than features. Procurement shifts from multi-year licensing to something closer to hiring for a specific capability. Agents get replaced when something better comes along, and competition plays out on results rather than on sales relationships or switching costs.
In the middle sits the layer that keeps this from becoming ungovernable: a thin orchestration layer that manages identity, permissions, budget, and audit trails. It does not execute work. It ensures work is executed within boundaries the enterprise can inspect and trust. Even in a world of near-zero coordination cost, trust still has to be established somewhere. The headless firm is not a firm without governance. It is a firm where governance is decoupled from execution in a way the monolithic suite never permitted.
What This Does and Does Not Mean
The hourglass is a useful frame, but it is worth being clear about what it does not resolve. The intent layer only works if agents can actually observe and execute real workflows — not just the clean ones exposed through tidy interfaces, but the full complexity of how enterprise software is used in production. The vertical agent market only functions if those agents can handle exceptions, permissions, and audit requirements without breaking. The orchestration layer requires a trust framework that enterprise security teams will accept in a regulated environment, not just in a proof of concept.
None of these are solved problems. The gap between the model and working production software is real and significant. What has shifted is not that the headless firm is fully buildable today. What has shifted is that the direction is no longer ambiguous. The economics have changed, and the software industry is repricing accordingly. The harder question now is which parts of this architecture prove durable, and which teams are actually equipped to build them. The next post looks at the most immediate obstacle: why agents being deployed today cannot see most of what they are supposed to automate.
Next in this series: AI Agents Never Graduate from Onboarding — on why the majority of enterprise workflows are invisible to the agents being deployed to automate them, and what that means for every organisation currently investing in agentic AI.
Tassilo Klein and Sebastian Wieczorek are the founders of Mantix.
 | A guest post by
|