The Operator Stack

MIT's Project NANDA studied 300 public AI deployments, interviewed 150 executives and surveyed 350 employees, and found that 95% of enterprise generative AI pilots produced no measurable business return despite $30 to $40 billion in investment. The cause was not model quality. It was integration: brittle workflows, tools that did not learn the work, and deployments misaligned with what people actually did each day. Individuals repeat the same failure at smaller scale. They collect tools and call it a system.

The principle: AI is a subordinate

Not a co-pilot. Not a replacement. A subordinate. It executes, you decide. That distinction is the whole difference between a stack that compounds and a subscription list that drains a card each month.

A subordinate gets a defined job. Anything in your setup that cannot answer "what is this for" is not part of a system — it is an expense you have not audited yet.

The four layers

Layer 1 — Capture. Everything that comes in: reading, notes, transcripts, saved material. The failure here is collecting without retrieving. If you cannot find something six weeks later, it was never captured, only stored.

Layer 2 — Processing. Thinking and analysis. This is where a model earns its place: arguing against you, finding the hole in your reasoning, drafting the version you then cut. Note what it is not doing — deciding.

Layer 3 — Automation. Grunt work elimination. The rule is narrow: automate what is repetitive and verifiable. Repetitive but unverifiable work automated is just error at scale, produced faster than you can catch it.

Layer 4 — Leverage. Assets and systems that keep working when you stop. Writing, products, processes others follow. Layers 1 to 3 make you faster. Only Layer 4 makes you bigger than your own hours.

Most stacks are all Layer 3 and call it a strategy. Automation without leverage means doing the same job faster, forever.

The layers are ordered deliberately. Automating a process you have not captured or thought about produces a faster version of a bad process, and the speed makes the badness harder to see.

Rules of engagement

Three rules keep a stack from turning back into a subscription list.

Every tool has one job, written down. If a tool's job overlaps another's, one of them goes. Overlap is where cost hides.

The model drafts, you decide. The moment a model's output goes out unread, the subordinate relationship has inverted and you are the one executing — checking, cleaning, apologising for it.

Verifiability gates automation. Automate what you can check. Anything else stays manual until you can define what wrong looks like.

Why the MIT number matters to one person

The 95% failure had a specific character: generic tools work for individuals because they are flexible, and stall in organisations because they never learn the workflow. The individual lesson is the mirror image. Your flexibility is exactly what lets a tool sit unused for months while still being paid for, because nothing in your week depends on it.

A system is defined by what breaks when you remove a piece. If nothing breaks, you did not have a system. You had a collection.

Your Next Move

Write down your four layers on one page. Every row gets a tool and a job-to-be-done. Any row where the job is vague is the row to cut.

Set up Layers 1 and 2 first. Capture and processing are where compounding starts. Automation before them speeds up a process you have not examined.

Audit at ninety days. Anything untouched in thirty days leaves the stack. The test is not whether a tool is good. It is whether your work depends on it.