The 90% is moving
Tim Kellogg’s You Can’t Sell an Agent gets the structural fact right. Enterprise agents fail from organizational mismatch, not capability. The 10%/90% decomposition — the agent itself versus everything around it that has to absorb the agent — is the keeper idea. Reporting paths, escalation routes, exception handling, accountability lines, the 3am watch. None of that ships with the model.
Where the essay ages badly is in treating the 90% as one thing. It isn’t. It’s two layers.
The first layer is technical integration. The wiring. Connecting the agent to the systems where work actually happens — CRM, docs, email, ticketing, data warehouses. Building the harness. Writing glue code. Standing up monitoring. The second layer is organizational tissue proper. Domain judgment about what to auto-approve and what to escalate. Policy design for when the agent disagrees with a specialist. Accountability architecture — who signs, who audits, who gets fired. Drift monitoring that knows what quiet failure looks like in this specific context.
Platforms are eating the first layer fast. MCP standardizes integrations that used to need bespoke connectors. Skills let non-engineers shape agent behavior without touching harness code. Computer use replaces per-tool custom wiring. Project and memory primitives shrink bespoke state engineering. Vertical fine-tunes pull domain capability into the model layer. Each of these shipped in the last twelve months. In another eighteen, “how do I connect an agent to my systems” will be as boring as “how do I connect an app to a database.”
The second layer doesn’t compress uniformly. Some of it is domain-generic and can be templated — what a legal citation check should do, what a clinical scribe note must contain, what a tier-1 support escalation looks like. This part is productizable, and it’s what vertical specialists are selling. Harvey, Sierra, Abridge don’t ship raw agents — they ship the domain-generic policy envelope with the agent inside. The per-customer bespoke part — this firm’s escalation chain, this hospital’s EHR quirks, this underwriter’s override rules — shrinks but doesn’t disappear. Delivery still carries service load. The envelope is pre-formed but has to be fitted.
Cloud in 2012 needed an integrator to lift and shift workloads. By 2020 that work was self-serve plus vertical SaaS — and Accenture and Deloitte were still billing billions, but for cloud strategy, migration governance, and change management. The glue code got eaten. The policy work didn’t. Salesforce in 2005 meant a twelve-month consulting engagement to stand up the system. By 2015 AppExchange and Lightning had absorbed most of that, and the partner ecosystem survived by selling customization of process, reporting, and org design. AI deployment is tracing the same arc, faster, and the operator role is moving up-stack with it.
Kellogg describes the operator as someone who wires the agent into the organization. That’s the current version. The 2027 version is someone who can tell when the agent’s summary is subtly wrong and why — someone shaping policy around a working primitive, not someone building the primitive. Implementation consultant becomes embedded product manager. Judgment scales poorly but ages well. Mechanics scale well but age fast.
The empirical winners already show this hybrid. Vertical AI specialists own both layers: compressed wiring absorbed from the platform, domain-generic policy baked into the product, per-customer fitting delivered as service.
The bets
On October 18, 2027, eighteen months from today, here’s what the compression thesis says we should see. If I’m wrong on most, Kellogg was right. If I’m right on most, “you can’t sell an agent” needed an asterisk.
- Integration is boring. MCP or an MCP-equivalent is shipping as a first-class endpoint on every major enterprise SaaS — Salesforce, ServiceNow, Microsoft 365, Workday, Snowflake, Google Workspace. Enterprise IT treats it as a standard connector, not a project.
- Vertical specialists keep winning. Harvey, Sierra, and Abridge are all larger by customer count, revenue, and valuation than today. Not acquired at a discount. Not flat.
- No horizontal agent SKU dominates enterprise. Platform companies’ biggest enterprise revenue lines remain infrastructure-shaped — API, compute, general-purpose seats — not product-shaped horizontal agents pitched as the-agent-that-does-your-job.
- The operator role shifts visibly. Job postings for bespoke AI-integration engineering flatten or decline. Inside enterprises, “AI product manager” and “AI policy lead” roles grow faster than “AI integration engineer.”
The 90% isn’t a moat. It’s two moats, and the water is draining from one into the other. If it isn’t, this post should be embarrassing to read in October 2027. Fair enough.