point of-view · Draft for discussion · September 2026
The compensating structure
Why most institutional AI work is tooling, and what the alternative actually looks like.
The cases
Click a case to expand it.
Market risk — the second machine
Unreadable
Each regime's risk representation, to every other regimeCompensation
Parallel risk estates, plus permanent reconciliation between themStop doing
Building the second machine — and reconciling itTransaction banking — the fungible dollar
Unreadable
The provenance, purpose and forward implication of each individual flowCompensation
Treating flows as fungible; buffers sized for ignoranceStop doing
Pricing a relationship advantage as though it were only a relationshipPrivate credit — the calendar
Unreadable
Contractual predicate state, at portfolio scale and between report datesCompensation
The quarterly cycle; smoothed, appraisal-based valuationStop doing
Treating a quarterly mark as though it were a priceAsk a room of senior bankers whether AI will change their business and every hand goes up. Ask what it is being used for today and you will hear: summarising documents, drafting client correspondence, accelerating code review, answering policy questions faster. Both answers are honest. They describe entirely different things, and the distance between them is where the next decade of institutional advantage sits.
The second list is tooling. Tooling makes an existing process cheaper, faster or less unpleasant. It is worth doing, but it leaves the shape of the business exactly as it found it. There is a test that separates the two, and it takes one question: what does this let us stop doing?
Every large process is compensating for something unreadable
Somewhere in the history of any substantial institutional process there is a piece of information that could not be read — not missing, but illegible at the speed or scale the business needed. Around that illegibility the institution built a structure. Call it the compensating structure. The genuine contribution of AI — the part that is not tooling — is that it dissolves the constraint, by making previously illegible material computable at all.
What follows for how you build
- Separate interpretation from calculation. AI reads unstructured evidence and produces structured state; the financial calculation that consumes it stays deterministic, transparent and governed.
- Provenance is the precondition, not a feature. Every derived value carries the clause, line, document and date it came from.
- Uncertainty must survive to the output. A system that cannot return unknown is manufacturing precision it does not have.
- Decision support long before decision authority. Authority is sought one class at a time, lowest consequence first.
- Every stage independently worth doing. Each stage delivers something the business would want even if the next stage is refused.
What this does not claim
That the compensating structures were errors; that any of this reduces underlying risk (it changes when and how existing risk becomes observable); that it moves quickly; or that the reading is reliable enough yet in every domain — that is an empirical question per corpus, and it should be measured before anything is built on top of it.
The constraint was never capability. It was legibility.