Plimsoll: A Clear Line for AI Reliance

Every bank that deploys a third-party AI system faces the same problem. The vendor says the system works. The bank’s model risk team must verify independently whether that claim holds for their specific tasks, their specific data, and their specific operating conditions. The assessment costs roughly $79,000 per system. And the bank across the street is paying the same $79,000 to assess the same system for the same tasks.

There is no shared answer. No independent register. No classification that travels with the system.

In maritime shipping, this problem was solved in 1764. Lloyd’s Register classifies vessels independently: the shipowner pays for the survey, the classification is published, and every insurer in the market can rely on it. The Plimsoll line — the load mark painted on every hull — shows at a glance how far the vessel may be safely loaded under different conditions.

Plimsoll applies the same model to AI in banking.

What Plimsoll does

Plimsoll is an independent classification body. It tests AI systems against banking-specific task corpora with practitioner-verified ground truth, assigns reliance limits per operating condition, and publishes results in a register with vendor right-of-reply.

The reliance lens: L0–L3

Every AI system receives a reliance limit for each banking task it is assessed against, under each operating condition:

• L3 — AI may act independently. The system has demonstrated sufficient accuracy, robustness, and explainability for the task under these conditions.
• L2 — AI recommends, person confirms. The system assists the decision but a qualified person must review and approve.
• L1 — AI assists, person decides. The system provides information but the decision rests entirely with the human.
• L0 — AI out of scope. The system has not been tested for this task under these conditions, or testing revealed it cannot be relied upon.

The same system may be L3 for domestic retail AML alert disposition and L0 for sanctions-adjacent correspondent banking. The reliance limit 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.
2. Vendor access agreement — standardised contract language enabling independent testing.
3. Testing against banking Rules — task corpora with practitioner-verified ground truth, stress and injection testing, change-control validation.
4. Load-line assignment — reliance limits assigned per operating condition.
5. Registration with vendor reply — results published in the register; the vendor may respond (up to 300 words).
6. Ongoing re-testing — quarterly reviews for drift; full re-survey if the underlying model changes.

The register

Nine specimen entries demonstrate the register format: system name, surveyed use, class designation, status (in class, suspended, awaiting survey), test scope and findings, vendor conditions and replies, and reliance limits by task.

All specimen data is fictional. The register format is real.

Why shared economics change the calculation

One bank assessing one AI system alone: ~$79,000.

Twenty banks sharing a Plimsoll survey of the same system: ~$6,842 per bank.

91% cost reduction. The classification travels with the system across every subscribing bank. The bank that joins the consortium in year three inherits every survey conducted in years one and two. The economics compound with scale.

Subscription tiers (by total bank assets)

• Regional ($10–30B): $95,000/year
• National ($30–100B): $190,000/year
• Global (over $100B): $380,000/year
• Overlay service: $22,000 per system reviewed

The independence structure

The conflict of interest in AI assessment is structural: the vendor wants a favourable classification, and most assessment firms are paid by the vendor or funded by the same investors.

Plimsoll eliminates this:

• Banks fund the work. One bank, one vote.
• Vendors cannot pay or sponsor. No vendor funding, no equity from assessed parties.
• Delaware public benefit corporation. A member trust holds veto over any change to independence provisions.

The independence is not a policy. It is a legal structure.

The regulatory context

The timing is not accidental. OCC SR 26-2 and Federal Reserve guidance are increasing expectations for independent validation of AI systems in banking. OSFI E-23 in Canada follows the same trajectory. Banks that can demonstrate independent, evidence-grade classification of their AI systems are better positioned for examination than banks relying on vendor self-attestation and internal model validation alone.

Plimsoll does not replace model risk management. It provides the independent, shared evidence layer that model risk teams can incorporate into their own frameworks.

The Founder’s File

The platform includes a twelve-section business plan: the thesis, competitive landscape, shared economics, 60-month financial model, addressable market ($64.2M US, ~$146M extended), go-to-market strategy, survey clause contract language, risks and stop rules, first 90 days, team and capital requirements, visual identity, and sources.

Proposed launch: January 2027. Seed funding: $4.5M. EBITDA-positive: March 2030.

Why this product belongs in this portfolio

Plimsoll is the product that connects twenty years of banking regulatory experience to the AI governance thesis that runs through the entire portfolio.

RegTwin AI monitors regulatory change. AgentPMO governs AI agents. BLACKBOX records what agents do. Bulwark enforces guardrails at runtime. Sea Trial validates the guardrail stack. The Fracture measures governance decay.

Plimsoll classifies the AI system itself — independently, with shared economics, and with a legal structure that guarantees the classification cannot be bought.

It is the load line for AI in banking. A clear mark showing how far you may rely — and where you must stop.

Plimsoll is live as a business concept and demonstration platform.

Open Plimsoll →

Open the Portfolio Briefing →

Explore the full portfolio →

More articles