Why the role appeared now

For twenty years the binding constraint on enterprise software was build capacity. If a customer needed a system, the question was whether anyone could build it in time. That constraint has moved, and it has moved far enough to change what a company should hire.

The arithmetic

The handbook puts the shift in one line. A traditional project is roughly 20 units of understanding plus 80 units of execution: 100 in total. With current tooling the same project is 20 plus 8: 28. Understanding's share of the work rises from about 20% to about 71%.

Two things follow, and only one of them is comfortable. The comfortable one: the same scope now costs a little over a quarter of what it did. The uncomfortable one: the part that got cheap was never the part that decided whether the project was worth doing. When 71% of the remaining effort is understanding, hiring more builders buys very little. This is an in-book estimate, not a measured study — the ratio is the argument, not the decimal.

The four-layer information gap

The second reason is structural. An AI system fails to connect to a business for one of four reasons, and only the first is a data problem.

1. The data is not in the system. It lives in someone's spreadsheet, a photographed form, a group chat.
2. Field meanings have been lost. The column still exists; nobody remembers what value 7 means, or why the field is nullable.
3. The real process was never written down. The documented procedure is one thing; the version people actually run — the shadow system — is another.
4. Nobody has the authority to say no.

The fourth layer is the one engineers most often miss, because it does not look technical. A system that contradicts an undocumented process needs someone with the standing to retire that process. No amount of model quality supplies that. Leave any one layer unbridged and the deployment stops at the demo, which is exactly where most of them stop.

Usage is not return

The survey record is consistent. McKinsey's 2025 State of AI finds 88% of organisations using AI regularly, with only 39% reporting enterprise-level EBIT impact. BCG's 2025 work puts 5% of companies at scale value and 60% at almost none. Gartner expects more than 80% of enterprises to have used GenAI APIs or deployed GenAI applications by 2026, and reported in 2025 that 63% of organisations do not have, or are unsure they have, the data management practices AI requires. Microsoft's 2024 Work Trend Index found 75% of knowledge workers using AI and 78% of AI users bringing their own tools to work. Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027.

Read together, these say the constraint is not usage. The 78% shadow-AI figure is the tell: demand already exists and official channels are not serving it. The distance between 88% and 39% is not a capability gap in the model. It is the four layers above, priced.

The obvious objection is that better models will close this on their own. They will close part of layer one and some of layer two. They will not create an authority to retire a process, and they will not recover a meaning that left with the person who knew it. The constraint migrates; it does not disappear.

Why this is a job and not a project

When execution is 8 and understanding is 20, the person who can do the 20 is worth more than the person who can do the 8, and the 8 is turning into a commodity input. That is the whole argument for a role that owns understanding end to end. The work is not new. The work is now the majority of the cost, which is a different sentence.

There is a second-order effect worth stating plainly. Cheaper execution makes it possible to start more projects, including projects whose problem was never defined. The failure moves upstream and gets quieter: a project that dies in the third week of discovery looks like nothing happened, while a project that dies at go-live looks like an incident. Organisations that only measure the second kind will keep buying execution, and will keep being surprised.

The test

Take one AI initiative that stalled. Name the layer — one, two, three or four — that was never bridged. If you cannot name it, you have diagnosed the model, not the deployment. Then look at the dashboard: if the only number on it is users, you measured adoption, and adoption was never the scarce thing.

For the organisational shape this role took, see what a forward deployed engineer is responsible for. For how reconnaissance surfaces layers one to three, see the five-stage delivery method.