Handbook / Catalogues / Financial Services
Download PDFFor institutions where "usually right" is not a release criterion.
Cohorte's programs for banks, insurers, asset managers, and fintechs. Built on the founder's published research in conformal prediction and self-consistency. Designed to survive the regulators you actually answer to.
Why FS needs different training
In a bank or insurer, an AI output is not "a nice generation." It is a credit decision, a fraud flag, an underwriting note, a suitability assessment, a transaction-monitoring alert. Each is governed. Each has a regulator. Each has a model-risk function whose job is to ask: what does this system do when it is wrong, and how would you know?
The supervisors named the gap.
Both regimes require documented evaluation, verification, and ongoing monitoring of any model influencing a business decision. Most "AI training" programs teach the model. They do not teach independent validation, ongoing monitoring, or governance.
Two regimes, one operational obligation.
DORA (in force since January 2025) covers ICT third-party risk. The EU AI Act puts most credit-scoring and insurance pricing into the High-Risk Article 6 category, triggering Articles 12 (logging), 14 (human oversight), 15 (accuracy/robustness).
The market is moving.
JPMorgan deploys LLM Suite to ~200,000 staff under strict access controls. Goldman has GS AI Platform with internal evaluation gates. The question is no longer "should we train on AI." It is "how do we train on AI without breaking model risk."
Why Cohorte
The methodology your model-risk team is going to ask for, taught by the practitioner who published it.
"Provably right within stated bounds."
The mathematical primitive that converts a model output into a statistically rigorous confidence interval. The closest existing analogue to the VaR / ES discipline already familiar to a market-risk team.
Source: The conformal-calibration method, from our reliability-certification paper. See /research.
Detecting confabulation.
If the same prompt run five times produces five different answers, the model is not reasoning. The verification gate that catches this before output reaches a customer-facing channel or a regulator-facing log.
Source: The self-consistency method, from the same paper. See /research.
The red team your CISO asked for.
Prompt injection. Indirect injection through retrieval corpora. Tool-call hijacking. The systematic taxonomy your security team uses to scope penetration testing of an LLM application.
Source: Our 10,000-trial exploitation-surface taxonomy. See /research.
Open-source. Inspectable. Yours.
TrustGate (verification gates), Guardrails (policy enforcement), Agent-Auth (authorization), Agent-Monitor (observability). Six public repositories. Your engineers can audit the layer before deploying it.
Where: github.com/Cohorte-ai.
The use cases we train against
The AI work financial institutions are actually shipping in 2026.
| Use case | Where it sits | Risk category | Verification primitive |
|---|---|---|---|
| Customer service copilot | Retail bank, insurer | Conduct & mis-selling | Self-consistency, response guardrails, escalation gates |
| Underwriting assistant | Life / P&C insurer | AI Act high-risk · fairness | Conformal bounds, adversarial fairness tests |
| Credit memo drafting | Commercial / corporate bank | Credit risk · disclosure | Source attribution, citation-grounded generation, audit logs |
| KYC / AML alert triage | Compliance, financial crime | AML & sanctions | Decision logs, human-in-the-loop gates, explanation generation |
| Suitability assessment | Wealth, retail brokerage | MiFID II suitability | Rule-based gates, output classification, regulator-facing logs |
| Fraud / transaction monitoring | Payments, retail bank | SR 11-7 / SS1/23 | Conformal anomaly bounds, false-positive tracking |
| Equity research / sell-side notes | Asset manager, sell-side | Disclosure · market abuse | Source attribution, content classifier, publication gate |
| Code generation for trading systems | Quant, market risk | Model risk | Static analysis, regression evaluation |
| Internal knowledge agent | Bank-wide | Conduct · IP leakage | Access control, DLP gates, query auditing |
Sample curriculum
What a typical 12-week FS Team Bootcamp covers, module by module. The actual curriculum is tuned to your stack and use cases in the pre-bootcamp scoping call with Charafeddine.
| Module | Topic | What your team produces |
|---|---|---|
| 01 | Scoping & operating discipline. Process First. The Thinking Stack (Clarify, Structure, Generate, Verify). LUMEN scoping brief drafted live for one real FS workflow. | Mission charter and LUMEN scoping brief for one production AI workflow, signed off by the sponsor. |
| 02 | AI Engineering Foundations. The Three V's (Vibes, Variance, Vendor). The Accountable Development Lifecycle. Prompt architecture, multi-model patterns, MCP tool use. | Working system v0.1 running against your stack and your data. |
| 03 | Trust & Verification. The Confidence Problem. Self-consistency sampling and canonicalization. Conformal prediction. TrustGate integration end-to-end. | Verification gates wired into the system. Statistical guarantees on output, with documented bounds. |
| 04 | Accountable Agents and the 4-Layer Architecture. Platform Protocol. Agent-Auth (least-privilege access for AI). Guardrails as architectural constraints. Agent-Monitor. | Agent-Auth and Guardrails layered into the system. Least-privilege access patterns documented for model-risk review. |
| 05 | Model Risk Governance. SR 11-7, SS1/23, DORA mapping. AI Act Articles 9, 10, 12, 14, 15. ISO 42001 clauses 6 to 10. Immutable audit-trail design. | Governance log, regulatory mapping cited by section number, audit-trail design defensible to model-risk function. |
| 06 | Capstone and Sponsor Brief. Red-team workshop. Founder critique. Sponsor demo to leadership. Operating brief written, reviewed, signed. | Production-grade FS system. Sponsor brief that survives a model-risk review without amendments. |
The FS portfolio
Same operating layer underneath every program. Vertical-specific tuning bends the exercises and the regulatory annex. Pricing is fixed; scope is negotiable.
FS Pilot · €8K–€12K · 4 weeks
Verification scopingFS Team Bootcamp · €4,200/seat · 12 weeks
Private cohortFS Curriculum License · €12,000 / year
Up to 25 seatsFS AI Readiness Program · from €35,000 · 3-6 months
Assessment + advisory + playbookBeyond the four programs: ongoing, founder-led advisory is available on a retainer, for firms that want Charafeddine on call between engagements. Operating counsel on your systems, not strategy decks. Email [email protected].
The FS objections we always hear
Asked, answered, on the table.
"You don't have a named retail-bank reference."
Correct. PwC is the closest analogue (60+ systems, 4,000 Copilot users, governance built in). For FS buyers needing a named-FS reference, the Pilot is the right entry point.
"Our model-risk team has SR 11-7 expectations you may not meet."
Curriculum is explicitly mapped to SR 11-7 §III and SS1/23 principles 1 through 4. We walk your model-risk team through the mapping on the discovery call.
"We have an internal AI / ML team."
Good. We are not a replacement. Internal teams teach the stack; Cohorte teaches the operating discipline (taste, scope, verification, governance) that internal teams typically do not own.
"We need on-site delivery inside the firewall."
Supported. AI Readiness assessment and two of four Team Bootcamp sessions on-site in Paris, London, Frankfurt, Amsterdam, Brussels, Casablanca. No Cohorte infrastructure dependencies.
"Procurement / vendor risk has a 60-day onboarding."
The Pilot is the way around this: fixed-price, fixed-scope, often clears procurement under a faster threshold. We have run Pilots while full vendor onboarding is in progress.
"AI training is not in this year's budget."
Cohorte is not Qualiopi-certified today (in progress, 2026 roadmap), so direct CPF and direct OPCO funding aren't available right now. The standard path is direct L&D budget; OPCO-routed engagements go via a Qualiopi-certified partner. The Pilot at €8K-€12K typically fits inside discretionary spend mid-year.
Bring your worst FS use case.
Suitability copilot. Credit memo. KYC triage. Fraud assistant. €8,000 fixed. Four weeks. The brief that survives the model-risk meeting.