Earlier this month I argued the future of AI is T-shaped: one general assistant everyone stands on, and firm-specific depth, written as Skills, in the few processes where judgement and regulation actually bite. The comments pushed the argument somewhere I had stopped short of, and they deserve a proper reply rather than a dozen individual ones.
I still think the shape is right. What the comments separated for me was ownership from defensibility: the Skill can be yours without being the moat. One comment put it as a correction, and I'll take it: the T-shape relocates the moat rather than removing it.
Plain language travels
A Skill is a folder of plain-language instructions that the assistant loads when the task in front of it calls for that expertise. Plain language is what makes it yours. You can read it, change it the morning a regulation changes, and take it with you if you change models. The counterpoint from the same commenter is that the property cuts both ways: the playbook is copyable. A playbook clear enough for a model to follow is clear enough for a competitor to imitate. Good patterns spread quickly across an industry, and staff move firms.
I don't think that's a problem to engineer around. Portability is judged by what it does for the owner; a moat is judged by what a competitor can cheaply reproduce. They are different tests, and the same document can pass the first while failing the second. If the words can walk out the door, the words aren't the moat.
Two firms, one Skill
Suppose two insurers install the same submission-triage Skill today, word for word. Six months later they will not be running the same system.
One wires the assistant into its underwriting platform so it can pre-fill terms; the other keeps it away from the system of record and has it draft for a person instead. Their eval banks, the sets of real past cases with known answers that each new version of the Skill is tested against, grow from different histories: a decade of near-misses at one firm, whatever the team remembers at the other. After an early scare, one tightens its authority limits, while the other decides the approvals are slowing quotes down and pulls a gate out. Each firm's exception history teaches its playbook different lessons, and the corrections compound in different directions. It's the learning loop you own doing exactly what it does, on your cases, on your side of the wall.
The Skill is where both firms started. What they compete on is everything the Skill is embedded in: the systems it can see, the cases it is tested against, where the approval gates sit, and what the exception history has taught it. Another commenter described this as getting the harness and the skills working together, so you keep the industry edge and your own company's flavour. That surrounding system is slow to build and hard to copy from the outside, which is more than you can say for a document.
Accountability was always yours
There is a quieter consequence of writing your own operational logic, and one of the comments put it better than I had: authoring the playbook pulls the expertise in-house, and the accountability with it. Owning the logic and owning the consequence, in the commenter's phrase, are now the same address.
When the logic lived inside vendor software, accountability had somewhere soft to land. The system declined the application; whose judgement was that, exactly? Boards and licensed professionals were accountable for outcomes the whole time, but the reasoning sat in a codebase nobody in the firm could read. Write the playbook yourself, in words your own people chose, and there is no vendor left to point at. That is more control and more exposure in the same move, which is why the condition in that comment matters: a named person has to own what the model does with what you wrote. The signature on the output was always going to be yours. The T-shape just made that unavoidable.
From building to proving
The past year of AI work has been about building. Anyone can now write a Skill, and the tooling for agents improves monthly.
The question that arrives later is harder. A recommendation gets challenged, and someone asks why the assistant reached it. Which version of the Skill was loaded? What information did it rely on? Who approved the result, and could another reviewer reproduce it? Every organisation will eventually want answers. For consequential regulated work, they are questions a firm should expect to have to answer. The sharpest version came from the comments: in regulated work, defensibility lives in verification, not the prompt. The commenter was talking about evals, and I think the same logic extends one step further, to the record of what actually happened in production.
This is where the moved moat gets its depth. A competitor can copy your playbook. They cannot copy your eval bank, because it grows out of your own cases and corrections. And they cannot copy the retained record of work your firm has already done and stood behind, because that record only exists where the work actually ran. The first can be built. The second can only accumulate.
The general assistant is becoming the surface everyone works on. The Skill is becoming a document anyone can write. What remains is the part that was always hardest to copy: how your firm actually works, and whether you can show it.