The Internal Move Needs a Shipped Thing, Not a Course

What Actually Moves an Internal Decision

What moves it is what the deciders can already see, and what it costs them to lose you. Their opinions about you were formed from watching your work, and the real question is not whether you can do the new role; it is what it costs to move you out of the one you are in. Generic career-change advice (take a course, update your profile, network in) misreads that: it is written for someone applying where they are a stranger.

That also changes what counts as evidence. A hallway conversation where a manager says they think you would be great at this is warm, and it is thin as a decision input. In a controlled study, undergraduates asked to predict a classmate's semester grades did worse after conducting an unstructured interview than they did working from the classmate's prior grade record alone ¹. That is a lab analogue rather than a hiring study, so what travels is the mechanism and not the size of it: an unscored conversation adds material that feels informative and crowds out a record that was already predictive. A shipped thing someone else used sits on the record side of that line, and it is the version of evidence a manager can describe accurately to someone above them without relying on their own memory of how a conversation felt.

Build Something Small, Then Get Someone Else to Use It

Choose a real, small problem that sits where your current work overlaps the work the destination team does, not a hypothetical one, and build a working answer to it with an AI assistant in the open. Then get it in front of someone other than you: a teammate who uses it for a week, a report someone downstream actually reads, a process someone else runs without you standing over it.

Adoption by one other person answers the question a manager actually has, whether this holds up outside your own head, in a way a polished demo only you have opened cannot. It will not, on its own, be the thing the receiving team decides on. Evaluating whether an internal candidate can really do an AI-heavy role means handing them one task built from the destination role's own material and reading back what they framed, what they refused to take from the model, and what they checked outside it, on the explicit view that a record built in the job you are leaving does not carry to the job you want. So keep the working notes as well as the thing: the frame you wrote before the first prompt, the assistant output you rejected and the reason, the check you ran outside the chat that changed the answer. That is the shape the review reads, and having it already written is what makes the artifact count for more than the demo it resembles.

Ask What the Receiving Team's Bar Actually Is

Ask before you spend more time building than you actually need to. Different teams are looking for different things at this moment, and a written list of what "AI-heavy" means on that specific team is worth more than another week spent guessing. If no such list exists yet, that is itself useful information: it means the team has not decided what it is testing for, and a direct conversation with whoever leads it may do more than a project would.

Bring more than one kind of evidence. In a 2023 article revisiting the validity of widely used selection methods, Sackett, Zhang, Berry and Lievens found that a composite of several predictors recovers most of what the older single-method estimates had promised ². That result describes measures combined deliberately, with weights, on predictors that measure genuinely different things, so it is not a licence to pile up impressions. Read the shape of it rather than the number: for an internal move, a shipped artifact plus a direct answer from someone who has seen your work plus your own honest account of what you delegated and what you checked is stronger than a project alone or a reference alone. A team that has never written its bar down may be evaluating on instinct, which cuts both ways: harder to prepare for precisely, easier to move once one person in the room has seen your work directly.

Don't Let the Backfill Question Sink the Move

Raise the backfill question yourself, early, rather than letting it surface as the unspoken reason the move keeps getting delayed. Being genuinely useful where you are is not a skills gap; it is a scheduling problem, and it is solvable in ways a manager may not have thought through until you name it: a transition period, a documented handoff, a colleague who can grow into part of your current work.

There is context to argue from, and it is worth sizing honestly. Job postings that mention AI reached 6.3% of US postings on Indeed in August 2026, past their prior peak of 3.3% in 2022, and the Bureau of Labor Statistics projects functions like HR to grow faster than the average occupation through 2034, even as the work inside them changes ³⁴. Both are measures of employer language and headcount forecasts rather than of how your company will handle one move, so carry them as background, not as an argument. What gives an internal candidate standing is the offer to solve the scheduling problem: naming a concrete transition plan, even a rough one, does more than waiting for a manager to solve it unprompted.

Should You Wait to Be Asked?

No, especially not if you are already senior and well regarded in your current role. The advice to wait for a manager to notice and offer the move assumes someone is actively looking, which is worth checking rather than assuming: a manager's attention goes to the work in front of them before it goes to career-pathing yours. Naming the specific move you want, with a shipped artifact already in hand, is what starts the conversation.

If the honest answer is that you are early in your field rather than senior in it, the more useful first question may be whether to aim for an AI job at all, and which of the three kinds. Either way, the sequence is the same: build the small real thing, get someone else to use it, ask what the bar actually is, and raise the backfill question before someone else raises it for you.

See the benchmarks >