Plimsoll: Maritime Classification Applied to AI in Banking

The Problem That Exists in Every Bank

Walk into any Tier-1 bank’s model risk or third-party risk team and ask how they assess the AI systems their vendors provide. The answer follows a predictable pattern:

The vendor delivers a model validation report. The bank’s internal team reviews it, finds it insufficient for their specific use case, and commissions an independent assessment. That assessment typically costs $60,000 to $100,000 depending on scope — call it $79,000 as a working average.

Now walk across the street to the next bank. They are deploying the same vendor’s AI system for the same AML alert disposition task. They commission their own independent assessment. Another $79,000. The third bank down the block does the same. Another $79,000.

Three banks. Same system. Same task. Three identical assessments. $237,000 spent to answer the same question three times.

This is not a market failure in the traditional sense. Each bank is behaving rationally — their regulators expect independent validation, and no shared infrastructure exists to provide it. The failure is structural: the absence of an independent classification body.

The Maritime Precedent

In 1764, Lloyd’s Register began classifying merchant vessels. The problem was identical: every insurer needed to know whether a specific ship was seaworthy for a specific voyage under specific conditions. Without shared classification, every insurer conducted their own survey. The duplication was expensive and the quality was inconsistent.

Lloyd’s Register solved it with three design decisions that Plimsoll replicates:

Independent survey. The classification body tests the vessel. Neither the shipowner nor the insurer controls the assessment.

Published register. The classification is available to every market participant. The evidence travels with the vessel.

Load line. The Plimsoll line — painted on every hull since 1876 — shows at a glance how far the vessel may be safely loaded under different conditions. Not a binary safe/unsafe verdict, but a graduated boundary: safe to this line, in these waters, in this season.

The Reliance Lens: L0–L3

Plimsoll translates the maritime load line into four reliance tiers for AI systems in banking:

L3 — Unrestricted action. The system has demonstrated sufficient accuracy, robustness, and explainability for this task under these conditions. It may act without human review.

L2 — Recommends and confirms. The system recommends; a qualified person must confirm before action is taken.

L1 — Assisted execution. The system provides information; the human performs the task.

L0 — Not for use. The system has not been tested for this task under these conditions, or testing revealed it cannot be relied upon.

The critical design decision: the same system may hold different reliance limits for different tasks and different operating conditions. A system classified L3 for domestic retail AML alert disposition may be L1 for cross-border correspondent banking and L0 for sanctions-adjacent transactions.

The classification is not a grade of the system. It is a boundary of use.

The Six-Step Survey Method

1. Nomination. Three or more subscribing banks nominate a system for assessment. This ensures the survey serves genuine market demand rather than vendor marketing.

2. Vendor access agreement. A standardised contract clause enables independent testing. The clause is designed to be incorporated into existing vendor agreements with minimal negotiation.

3. Testing against banking Rules. Task corpora with practitioner-verified ground truth. Stress testing under adverse conditions. Injection testing for adversarial inputs. Change-control validation for version management.

4. Load-line assignment. Reliance limits assigned per operating condition based on test results. The assignment is evidence-based, not opinion-based.

5. Registration with vendor reply. Results published in the register. The vendor receives the findings and may respond with up to 300 words. The response is published alongside the classification — not instead of it.

6. Ongoing re-testing. Quarterly reviews for performance drift. Full re-survey triggered by material model changes, significant vendor version updates, or new module releases.

The Independence Structure

The conflict of interest in AI assessment is structural. Most assessment firms are funded by vendors, retained by banks with existing relationships, or staffed by people who move between the assessed and the assessor.

Plimsoll eliminates these conflicts through legal structure, not policy:

Delaware public benefit corporation. The entity has a legal obligation to serve the public benefit — independent classification of AI systems — not shareholder returns.

Banks fund the work. Subscription tiers by total bank assets: Regional ($10–30B) at $95,000/year, National ($30–100B) at $190,000/year, Global (over $100B) at $380,000/year.

Vendors cannot pay. No vendor funding, no vendor sponsorship, no equity from assessed parties.

Member trust holds veto. Any change to the independence provisions requires approval from the member trust — not the board, not the management team, not the investors.

The Shared Economics

The economics are the moat. One bank assessing one system alone: ~$79,000. Twenty banks sharing a Plimsoll survey: ~$6,842 per bank. The 91% cost reduction is not a marketing claim — it is arithmetic.

The classification travels with the system. When the twenty-first bank joins the consortium, they inherit every survey conducted before they arrived. The economics improve with scale. The evidence compounds with time.

Why This Product Required Twenty Years of Banking

Plimsoll could not have been designed by someone who had not spent two decades inside the institutions it serves. The survey method reflects empirical knowledge of how model risk teams actually work, what vendor contracts actually permit, and where regulatory expectations are actually heading.

The choice of maritime classification as the model — rather than credit ratings, audit opinions, or certification schemes — reflects a specific insight: what banks need is not a grade (like a credit rating) or an opinion (like an audit) or a badge (like a certification). What they need is a boundary: how far may we rely on this system, for this task, under these conditions?

That is a load line. And load lines require an independent surveyor, a published register, and an economic model that makes the classification cheaper to share than to duplicate.

Twenty years of watching banks duplicate identical vendor assessments. Plimsoll is the structural answer.

More case studies