The rest of this blog has examined what looks like a narrow problem: why AI struggles with the rule-bound back-office work inside banks. The guessing that turns into fines (see why 95% of enterprise AI pilots fail inside regulated institutions). The data that cannot leave the building (see the question that kills AI projects in a bank). The restraint that turns out to be the real feature (see what AI has to be before a compliance officer signs off). The rules that rot the moment they are encoded (see why a static AI model quietly stops being correct).
Zooming out from those four pieces reveals something more general than a banking problem: a shape, and once identified, it appears almost everywhere in regulated and semi-regulated work.
The shape
Strip any of these workflows down to its structure and the same four elements recur:
A case arrives. There is a rulebook it has to be checked against. There is a deadline. And there is a decision that someone has to be able to defend later.
That is the entire pattern. Something arrives, published rules govern the outcome, a clock is running, and whoever makes the call may have to justify it to someone with authority over them.
Holding that template up against a single institution shows how many distinct functions it fits:
- Onboarding a business customer. A file arrives. Verification rules define what must be checked. There is a turnaround expectation. Approving someone incorrectly has to be explainable after the fact.
- A compliance alert. A flag is raised. Rules and thresholds determine whether it is reportable. There is a filing window. The decision to file, or not, has to be defensible.
- A reconciliation break. A mismatch surfaces. Rules govern how it should tie out. It cannot remain open indefinitely. Someone has to be able to justify how it was resolved.
- A tender response. A document arrives with embedded requirements. Pass-or-fail rules are buried inside it. There is a submission deadline. The decision to bid, or not, has to hold up to scrutiny.
- A payment dispute. A case opens. The scheme rulebook defines what evidence counts. There is a hard deadline. Whether the case is fought or conceded has to be justifiable.
Five different departments. Five different rulebooks. People with entirely different job titles who would not describe their work as similar at all. Underneath, it is the same shape wearing five different uniforms.
These are not five separate problems that happen to involve AI. They are one problem, repeated, and the reason each instance is hard is identical: published rules, real deadlines, real liability, and a general-purpose model that guesses precisely where guessing is unacceptable.
It extends well beyond banking
Outside finance, the same structure recurs with only the rulebook changed. An insurance claim: a case arrives, policy rules define what is covered, there is a settlement clock, and the decision has to be defensible. A healthcare prior authorisation: a request arrives, clinical and coverage rules determine the outcome, there is a turnaround requirement, and someone has to stand behind the call. A legal or regulatory filing. A customs declaration at a border. Each one is a case, a rulebook, a deadline, and a decision someone must defend.
The shape is identical. Only the rulebook changes.
Which reframes the underlying question. Banking is not where the problem originates, it is simply where the shape is densest and the stakes are most visible. The broader pattern is that a substantial share of the serious, high-stakes, rule-bound work running the modern economy shares this exact structure, and all of it fails generic AI for the same underlying reasons.
The implication for how this gets solved
If it is genuinely one shape, the interesting problem was never "can AI do disputes" or "can AI do onboarding." Those are individual instances. The real question is whether a system can be built for the shape itself: a case, a rulebook, a deadline, a decision to defend, engineered so it does not guess, does not require data to leave the institution, understands what it is not permitted to decide, and stays current as the rules move.
Solve that once, and the result is not a single workflow automated. It is a pattern that recurs across every department in an institution, and then across industries that otherwise have nothing in common.
This is the premise CaseClear is built on: one governed engine, with the case-rulebook-deadline-decision structure encoded once, applied across modules for onboarding, financial crime, reconciliation, compliance, and payments operations, each running against its own rulebook inside the same deployment.
Continue reading. See why 95% of enterprise AI pilots fail inside regulated institutions, or see how the pattern maps onto CaseClear's module architecture on the main site.