For most of my career, the breadth of my background was a problem to be managed in interviews. Architecture, data governance, user research, benefits management, development, testing. Recruiters preferred a clean line. Hiring managers wanted a specialist. The polite word for the alternative was generalist.
Right now, I'm about to ship the first version of a production housing product, built through five iterations in under a fortnight. There was no product manager, no business analyst, no user researcher, no UX designer, no security reviewer, and no test team. The work each of those disciplines would have done still happened. It was done by me, working with AI, through a set of skills I wrote to encode the questions each discipline asks.
What a skill is doing
A skill, in this context, is a short instruction file telling the model how a particular discipline thinks. Not how to produce code in its style. How to think in its manner.
I have skills covering strategic discovery, user psychology, conceptual architecture, security review, business continuity and disaster recovery, end-to-end testing, UX audit, and marketing positioning through behavioural science. Each one runs the work a senior practitioner in the field would do, structured around the artefacts a board would expect to see: PRDs, ADRs, threat models, recovery plans, UAT scripts, persona maps.
Each is governed by the rules I would apply to a contractor producing the same output. Version control. Named documents. Evidence trails.
The skills are not doing magic. They are doing what a calibrated specialist would do when someone tells them exactly what good looks like.
Why this is not the story most boards are being told
The prevailing narrative on AI is replacement. The headcount cost of a function gets reframed as a model subscription. The saving flows to the bottom line. The agenda question becomes which roles to thin out first.
It is the wrong question, because it gets the dependency backwards.
The skills work not because the model is good enough to do user research. The model is not good enough to do user research. The model is good enough to do user research when someone with twenty-five years of context tells it which questions matter, which answers are dangerous, which evidence is load-bearing, and which polished-looking output is wrong.
Without the calibration, the same model produces work which looks competent and is structurally flawed. The personas have no behavioural depth. The architecture passes a glance test and falls apart at integration. The security review misses the threat model the regulated sector would care about. The test suite covers the happy path and ignores the failure modes a regulator exists to catch.
The senior generalist is the calibration layer. It is the role AI does not remove. It is the role AI makes economically scalable for the first time.
The inversion
For most of the last two decades, the cost structure of professional work penalised breadth. Specialists scaled. Generalists did not, because their value depended on context built over years and applied to one engagement at a time. Boards optimised for specialist throughput and treated breadth as a tax.
The arithmetic has inverted. The work an old delivery model would have given to five specialists over six weeks now sits with one experienced practitioner and a calibrated model for under a fortnight. The bottleneck is no longer specialist supply. It is the judgement to direct the model and the judgement to catch it when it is wrong.
The judgement is the thing the previous twenty years of supposedly unfocused career history quietly built. Knowing what a good PRD looks like in the third week of a discovery. Knowing which user research findings are signal and which are the participant trying to be helpful. Knowing the difference between an architecture which is correct and one which is operable. Knowing when a security finding is a checkbox and when it is a board-level risk.
None of it lives in the model. All of it lives in the operator.
What this means for boards
Two implications follow, and both run against the prevailing investor narrative.
The first: the productivity gains from AI accrue to organisations which retain experienced generalists rather than thin them out. The cuts being celebrated in the trade press are, in many cases, removing the calibration layer which determined whether the model produced anything usable. The cost saving is real. The capability loss arrives later.
The second: the executive profile most undervalued by current hiring practice is the one most multiplied by the new arithmetic. Boards optimising for specialist depth are buying the input AI substitutes for. Boards optimising for breadth and judgement are buying the input AI compounds. Those are different bets, and they are not equally priced.
The career history I spent twenty years apologising for in interviews is the one which scales now. I doubt I am alone in this.