#8: How to Hire for AI (and Get Hired): The Four Roles of Intelligence Transformation

TL;DR: New AI roles are forming around the same practical problem: someone has to secure the decisions a deployment depends on, specify how the workflow really operates, encode what the company's data and rules mean, and build a system whose behavior can be verified. Digital transformation delivered the storage layer and was measured on whether the data arrived; intelligence transformation has to deliver the meaning layer, and that work has no established owner yet. This final installment of our series explains the four roles that carry it, what titles like Forward-Deployed Engineer and AI Operations Lead map onto, and what the shift means whether you are hiring or looking for work.

Swyx followed this with a deliberately exaggerated comparison: one year managing ten agents may now provide more relevant experience than ten years managing teams of ten to one hundred people. The comparison is at least half-right, because agents teach decomposition, context design, evaluation, and exception handling, but they cannot teach you how an institution responds when automation reaches work tied to people's roles and career paths. What is true in the hiring signal is that seniority alone has stopped serving as a reliable proxy for operating ability.

To understand why, start with a question we now ask on every enterprise project.

Not interested in this topic? Watch this video about how the inference race is shifting from installing more chips to extracting more useful work from them.

Follow us on

Ask a large company whether its data is ready for AI and you will get a confident yes, with a warehouse to point at. Ask a narrower question and the picture changes. On one project we asked how many data feeds bypass the warehouse entirely and land directly in the systems that consume them. Nobody knew, and not because anyone was careless, but because nobody had ever been asked to count. The program that built the warehouse was scoped and measured on whether the data got in, and by that measure it worked.

Once you start looking, the pattern is everywhere. Several hundred published views turn out to be dynamically generated JSON blobs rather than typed tables, which means they can be queried but cannot be built on. Business-logic validation does not happen at ingestion, so a bad number from a partner's file can reach an executive dashboard before any human examines it. The authoritative list of channels exists in three versions at once: a value hardcoded inside a pipeline, an Airtable maintained by one person with known gaps, and a view that refreshes every morning and is currently correct, with nothing recording which of the three should win.

None of this is failure. All of it works, because people are holding it together by hand.

The industry named the situation accurately and then forgot what the name meant. A warehouse is where you store things; it does not classify them, catalog them, or explain what they are, because that is a different building with different staff. Digital transformation delivered the warehouse well: paper became records, records became databases, desktop spreadsheets were pulled into systems of record, and every company we walk into has a data team with a serious budget and a director who can put a query in front of you within ninety seconds.

Nobody built the library, and nobody noticed, because humans were performing the library's functions by hand the whole time. The analyst who has been there four years is the card catalog. The reconciliation spreadsheet is the index. The person who knows which feed lies, and why, is the reference desk. Those workarounds were cheaper and faster than building the real thing, and for dozens of years nothing required otherwise.

Then the goal moved. Something is now standing at the end of the pipe that has never worked at the company before, with no institutional memory and no instinct for which numbers carry weight, and it cannot use the card catalog, because the card catalog is a person. The gap between "the data is in there" and "the data can be reasoned over" is what intelligence transformation has to close.

There is a reason this became urgent over the last eighteen months rather than gradually, and the reason explains why the work is organizational rather than technical. Any piece of knowledge work moves through four phases: you align on what you are doing, you specify it, you execute, and you verify. The phases have not changed. Their relative sizes have, drastically.

Building used to be the expensive part, and part of the expense was invisible work nobody accounted for: across months of implementation, shared understanding accumulated through code review comments and hallway arguments, so alignment arrived as a byproduct of slow execution and was never budgeted separately. Now that implementation has collapsed, the saved time relocates to the end of the process, where disagreement surfaces under a finished pull request, the most expensive place to discover you wanted something else. Maggie Appleton compressed the dynamic into six words: when implementation gets cheap, alignment is what's left.

Here is the version a CFO can act on: headcount does not need to shrink, it needs to move. Most organizations are staffed for the left-hand column of that table, with the bulk of their people in a phase that has narrowed by an order of magnitude, while the two widest phases in 2026 are the two that no existing department owns.

One line has to hold through all of this, and it drifts unless someone watches it. An agent can discover how your pricing currently works by reading the tables and tracing the logic. It cannot decide whether enterprise discounting is in scope this quarter. When a planning agent begins answering its own business questions and nobody notices, you have not accelerated the work; you have transferred authority without recording the transfer.

LinkedIn's 2026 Jobs on the Rise list places AI Engineer first and AI Consultant or Strategist second, with a reported median of 8.2 years of prior experience for the latter – this part of the market already skews toward people who have seen how organizations work. With the only caveat that AI in 2018 was very different. The range of new titles is wide, and the mandates behind them overlap heavily:

Forward-deployed engineering is the most visible version: OpenAI's current FDE role spans discovery, scoping, system design, building, rollout, adoption, and measurable workflow impact, and inside companies, AI Operations Leads are receiving a remarkably similar assignment. The hype suggests a new class of AI superhero has arrived. The plainer explanation is that companies had neatly contained strategy, operations, data, and engineering within separate departments, then AI mixed them together, and projects began failing at the boundaries. Underneath the titles, the work divides along the four phases:

Those are the four roles that encode a business.

And we will walk through them using examples from our current enterprise projects.

Learn from those who work directly with companies navigating these transitions.

Below, we are going to discuss:

Join Premium members from top companies like Microsoft, Nvidia, Google, Hugging Face, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.

⬅️ #7: AI Saved Time. Where Did the Value Go?

The layer above digital transformation. Digital transformation moved data into systems and was measured on whether the data arrived. Intelligence transformation writes down what the data means, so that a system with no institutional memory can act on it.

No. Data quality asks whether values are correct, complete, and fresh. Intelligence transformation asks what they mean: agreed vocabulary, entity definitions, invariants, and refusal rules. A warehouse can pass every data quality check and still be unusable by an agent.

A cross-departmental champion who owns alignment, an AI-enabled analyst who owns specification, a semantic modeler who owns the shared vocabulary, and an AI engineer who builds the tools and the verification machinery. Below roughly a hundred people, one strong generalist can carry all four across a narrow scope.

A vendor-side engineer who works directly with customers to scope, build, deploy, and improve production AI systems. The FDE can leave after a deployment; the internal owner cannot.

It is the market's current name for the four roles collapsed into one posting: identifying valuable workflows, defining how they should change, building or coordinating systems, driving adoption, and measuring the operating result. It typically reports into business operations or the COO rather than engineering.

Because it is the only phase without a natural owner. It widened more than any other phase while remaining everybody's second priority, which is why closed-loop systems stay open in practice.