The model is a commodity that gets better every quarter without you. The reviewable action pipeline is the thing that is actually hard, actually yours, and actually worth money.
Almost every AI purchase decision is framed around the wrong object. Which model. Which vendor. Which benchmark. Meanwhile the thing that determines whether the project survives contact with your operations team is barely discussed: what happens between the model deciding something and your systems changing.
That gap is the product. Everything else is a dependency you will swap out twice in three years anyway.
Models improve on somebody else's roadmap and budget. The one you deploy this quarter will be outclassed by something cheaper within a year, and if your system is built properly, swapping it will be a config change and a re-run of your evaluation set.
What does not commoditise:
Every one of those is specific to your business, none of them come in a box, and together they are what makes AI deployable rather than demonstrable.
One rule, and everything else falls out of it: the model never calls your API.
It emits a proposal — a typed, validated, reversible object describing exactly what would change, with the evidence attached and a declared way to undo it. A separate execution layer holds the credentials, checks the proposal against policy, and decides whether to run it now, queue it, or refuse it.
model → proposal → [ policy + trust level ] → execute | queue | refuse
↓
append to ledger
Three things you get for free from that shape, all of which are otherwise very hard:
Prompt injection stops being existential. A malicious instruction in a customer email can, at worst, produce a proposal that fails validation or gets rejected by a human. It cannot reach your database, because the thing that talks to your database does not read customer email.
Everything is reversible by construction. The reverse operation is declared alongside the action. You do not need to invent a rollback story during an incident at 2am.
Autonomy becomes a dial, not a decision. Because the execution layer already checks a trust level before running anything, raising autonomy for one action type is a configuration change backed by measured data — not a rewrite, and not a leap of faith.
Most approval workflows die because reviewing is slower than doing the task yourself. If that is true, your team will — correctly — route around it.
So the queue gets treated as a first-class product surface. Keyboard-first. One decision per screen. The diff against current state and the cited evidence visible without scrolling. Bulk approve for runs of identical low-risk items. And rejections captured with a reason from a fixed taxonomy, in one keystroke.
That last point is where most of the long-term value hides. A queue that records only yes and no is discarding the highest-quality training signal in the entire company. "Wrong policy" and "right action, wrong amount" are different failures demanding different fixes, and you only learn that if somebody built the button.
Not "which model do you use". Ask:
If the answers are vague, you are buying a demo. If they are concrete, you are buying a system — and the model inside it is a detail you will change twice and barely notice.
Support automation, AI phone agents, n8n back-office work, and the engineering loop itself — always behind a gate you control.