The Specification Layer

The line shows up in venture-aligned conversations with increasing frequency. Sometimes attributed to Mike Novogratz, sometimes to others — the shape is consistent enough that the attribution barely matters. AI compression has eaten the engineer. Five people in Chengdu, with the right tools and a copy of any prevailing SaaS product, can ship a clone before the original raises its next round. The cost of building has collapsed. The moats are illusion. The labor itself is the asset being devoured.

It is worth taking seriously before taking apart, because the version that is correct is genuinely correct, and the founders who do not understand which version applies to them are going to be expensively wrong.

01Where the framing is right

For one specific kind of company, this is true. Business-to-consumer software, targeted at knowledge workers, where the buyer encounters the product through self-serve discovery, the value is delivered through a user interface a competent engineer can rebuild, and the relationship between vendor and customer is mediated entirely through a website.

In that segment — yes. Five engineers in Chengdu can replicate Notion, or Calendly, or Grammarly, or any number of horizontal tools whose moat was the engineering hours required to build them. The labor is being eaten there. The framing is right about that segment, and any founder building in that segment in 2026 needs to internalize what it means.

02Where the framing fails

The framing fails the moment the buyer is not a knowledge worker.

It fails for any company that touches a regulated counterparty, a procurement function, a vendor master, a governance committee, or a contract that survives across administrations. The reason it fails has nothing to do with engineering and everything to do with what kind of work is actually involved in selling and deploying software inside an enterprise.

Two layers, worth distinguishing carefully.

The engineering effort to translate a specification into working code — the database schema, the API endpoints, the proxy logic, the front-end, the tests. AI compression genuinely is eating this layer. A senior engineer in 2026 with the right tools can produce in a weekend what took a team six months in 2019. Five engineers in Chengdu can absolutely do this work, and they will do it cheaper than five engineers in San Francisco. This is true. Acknowledge it.

The work of figuring out what the software is supposed to do — which buyer it serves, against which budget line, satisfying which procurement requirements, governed by which contractual constraints, in which regulatory regime, with which audit obligations, sold through which channel, supported by which kind of professional service relationship.

This work is not engineering. It is business knowledge translated into product requirements that an engineer can then implement.

The specification layer is not being eaten. It is becoming more valuable, because the implementation layer is becoming cheaper. When implementation was expensive, the specification could be loose — the engineering investment forced discipline by economic gravity alone, and the rough specification got polished through multiple build-test-iterate cycles. When implementation is cheap, the specification has to be exactly right the first time. Cheap implementation of the wrong specification is just fast wrong-direction motion. The premium for getting the specification right has gone up, not down.

03What this looks like in B2B governance

A Chief Information Security Officer does not buy software the way a knowledge worker does. A CISO has signed off on a control framework — NIST, ISO 27001, SOC 2, whatever applies — and every product they purchase has to fit a specific control they are accountable for. They buy through a procurement organization that runs reference checks, security questionnaires, master agreement negotiations, and budget approvals that take six to nine months on the shortest version. They sit on audit committees that ask about vendor risk and need defensible answers. They evaluate against contract terms in ways that require legal review. They hold the line in board conversations about scope creep and termination clauses.

Five engineers in Chengdu cannot get to a CISO. They cannot speak to procurement. They cannot translate a master service agreement into a runtime constraint. They cannot sit on the bench when a Tier-1 financial institution decides whether to sign a pilot. They cannot answer the audit committee's question about vendor governance. They cannot price against an enterprise budget line, because they do not know which line, or how much sits on it, or who owns it.

04The four moats

The work that closes a B2B governance sale is not engineering work. It is four distinct activities, each of which compounds over years and cannot be replicated by hiring engineers offshore.

1. Sales channel
   A motion led by a person the buyer trusts, often built over years of prior relationships in adjacent companies. Five engineers in Chengdu have no relationships with American CISOs.
2. Category framework
   A defensible reason for the buyer to allocate budget. CISOs do not approve line items for which no category exists in their planning frameworks. Categories are built by people who can host conversations with multiple buyers, frame the problem the same way each time, and let the buyer recognize themselves in the description.
3. Vertical expertise
   A person who can walk into a procurement review and answer "how does this fit our existing controls" without consulting a manual. That person is a domain executive. Not an engineer.
4. Professional services
   Implementation delivered by people who have done this work before, in regulated environments, often through a series of engagements that look more like consulting than software deployment.

These four are the actual product of a B2B governance company. The software is the artifact that anchors them. The five engineers in Chengdu can build the artifact. They cannot build the moats. Without the moats, the artifact does not sell.

05Accountability as moat

There is a deeper version of this argument that goes beyond commerce.

The four moats are forms of accountability. Sales channel is accountability to the buyer. Category is accountability to the market. Vertical expertise is accountability to the regulated environment. Professional services is accountability through performance over time.

Engineering, by itself, does not generate accountability. Engineering is morally neutral as to what it builds, and AI-compressed engineering is even more so — the engineer is no longer the bottleneck on what gets made, only on what gets specified. The moats are where ethics, fiduciary duty, and regulatory compliance actually live in a software business.

AI does not eat these. AI sharpens them — because AI without governance is the threat that creates the demand for the moats in the first place.

06The inversion

This is why the framing is more than wrong for B2B governance — it is structurally inverted. The same forces that compress the implementation layer expand the specification layer. The more the world looks like the framing says it looks, the more the moats matter.

A founder who accepts the framing as universally true builds differently. They minimize headcount, minimize sales motion, minimize category investment, minimize professional services, and rely on the implementation layer alone to carry the company. That founder is correct for B2C SaaS and is wrong for governance. The mistake, in venture terms, is to project the economics of one segment onto every segment. Founders who correct the mistake at the moment of capital negotiation keep their companies. Founders who accept it hand their companies to the people running the framing.

07The follow-up question, and its answer

Sometimes there is a follow-up. Won't AI also automate the sales motion, the category framing, the vertical expertise, the professional services — eventually?

The answer is: not in the buyer's lifetime, because the moats are forms of accountability, and accountability cannot be delegated to a system the buyer does not trust. The whole reason CISOs are buying governance software in 2026 is that they have learned AI systems cannot be trusted with unsupervised access to data they care about. The same buyers who are skeptical of AI agents in their data repositories are not going to trust AI agents to negotiate their procurement contracts.

The moats are durable specifically because the buyer has internalized that they are durable. That internalization is the deepest moat of all, and it grows stronger with every poorly-governed AI deployment in the news.

08The line for the room

When this comes up — and it will, repeatedly, from people who have absorbed the framing without questioning whether their segment is the segment it applies to — the answer should be short, calm, and unmemorizable in the moment.

The framing is not a threat. The framing is a sorting mechanism. Founders who accept it for the wrong segment hand their companies away. Founders who reject it for the wrong segment build companies that fail. Founders who can name which segment they are in, and build accordingly, win.

Build the moats. Hold the category. Sell to the buyers who actually exist. Let the rest of the market sort itself out around the assumption.