Zoral Intelligence Layer
A general-purpose AI tool does not know your credit policy, your lending authorities or who must approve what. It can give a confident answer that is wrong. And it is hard to show a regulator what it did and why.
Zoral Intelligence Layer
AI Agents Runtime
Rules & Guardrails
Zoral Data Fabric
Human Oversight
The Zoral Intelligence Layer works around the centre of Zoral fOS. Zoral Decision Engine runs its models and records every decision with its reasons. Zoral Core Orchestrator runs the workflows its agents work in. Zoral Data Fabric holds the data the models are tuned on. AI is guardrailed, with human-in-the-loop supervision.
The AI works on the same platform as your credit rules, workflows and audit trail. So it follows the same approvals and leaves the same record as everything else.
Agents gather documents, check data against your rules and draft notes for staff. A workflow can hand a task to an agent, and an agent can start a workflow.
Your bank chooses enterprise model services or open-weight models it hosts itself. They connect through a standard interface (MCP). Your data stays within the environments and agreements you control.
Your analysts, product and risk managers set up agents themselves. Each change is tested, versioned and approved before it goes live.
Your bank decides which workflows, tools and data each agent may use, with its own budget and approval steps. Agents act only through the permissions you set.
Every action, prompt, tool call and answer is recorded for your risk and security teams. These records are kept within the environments your bank controls.
Agents do not make credit decisions. Those follow rules and models your bank approves, and accountability stays with your bank. Each decision is recorded with its reasons, ready to explain.
Since 2007, Zoral has put machine-learning models into live, high-volume lending, where every decision counts. We were doing this long before the current AI wave. We tune each model on your bank’s own data. It then runs in Zoral Decision Engine in real time, at very high volumes, with every decision recorded for audit.
Machine-learning models handle default, pricing and limits. Language models read documents and draft. Your risk and compliance teams approve each model before it goes live or is retuned.
Zoral model library
Zoral Decision Engine
Customer lifecycledecisions
Portfoliooptimisation
| Category | Examples |
|---|---|
| Acquisition and conversion | Lead scoring, offer acceptance, likelihood to close |
| Fraud and identity | First- and third-party fraud, linked accounts, identity fraud |
| Credit risk and default | Credit scores, default and time to default, repayment behaviour |
| Affordability, limits and pricing | Affordability, limit setting, risk-based pricing, collateral valuation |
| Retention and lifetime value | Churn, cross-sell, early repayment, lifetime value |
| Collections and recovery | Likelihood and timing of payment, best contact method, collection strategy |
| Documents and underwriting | Document sorting and reading, cross-document checks, financial spreading |
| Portfolio and model monitoring | Portfolio scenarios, product mix, model quality |
Agents. Each agent has its own instructions, tools, knowledge base, guardrails, budget and approval steps, all versioned.
Models. One model interface routes calls to enterprise model services or self-hosted open-weight models, with budgets across all of them.
Data. Knowledge bases use PostgreSQL with vector search, in your environment.
Machine learning. Models are tuned on Zoral Data Fabric data and run in Zoral Decision Engine. Results are written back to Zoral Data Fabric.
Your own platform and code. Scores from your models on Amazon SageMaker, Databricks or Google Cloud Vertex AI feed Zoral decisions. Engineers can add Java, C# or Python libraries and connect approved systems by API, events or files. Their changes are versioned and approved like everything else.
Hosting. AWS or Azure in your chosen region, or your own data centre.
Models spot unusual patterns, and documents are checked against each other, before an offer is made.
Agents prepare the file and models score it in real time. Staff spend their time on the cases that need judgement.
Sharper risk and fraud scores help you approve more good customers.
Your next product starts from models already refined in live lending.
The Business Outcomes Layer shows what each change to models, pricing or limits did to the business. Each change becomes evidence for the next.
Models predict default and fraud for each applicant. Every applicant is assessed by the same approved rules.
The platform spots an applicant who is likely to give up, so your staff can contact them in time. Behavioural signals are used here only where your bank switches them on.
Agents help staff with servicing, know-your-customer (KYC) follow-up and customer questions. Staff approve what is sent.
Models predict which borrowers may refinance elsewhere, so your team can make a timely offer.
Zoral Decision Engine sets risk-adjusted pricing within parameters your bank’s team approves and reviews with your compliance team.
The challenge
Slow credit processing meant many applicants gave up.
What we did
The Zoral Intelligence Layer, with Zoral fOS, automates underwriting across several products and onboarding across channels.
The result
Much faster processing, more qualified loans without more risk, and better customer retention.
The challenge
High fraud exposure and slow manual checks on uploaded documents.
What we did
AI agents work with Zoral Document Fabric, our intelligent document processing (IDP) service, integrated with Zoral fOS.
The result
Less fraud, far less manual review and a faster time to close.
References are available in confidence.
Yes, through a standard interface (MCP). They read data, run checks or start workflows only within the permissions you grant. Every action is recorded. Credit decisions still follow the rules and models your bank approves.
You keep them. Their scores can feed Zoral decisions, and Zoral’s models can run alongside as challengers.
Every model version and decision is recorded. Your risk and compliance teams approve each release and each retuning. Model-risk obligations stay with your bank, and Zoral fOS keeps the evidence.
They could. Connecting the model is the smaller part of the work. The larger part is the control around it: permissions that match lending authorities, approval steps and a record of every prompt. Someone must keep that working for years. Zoral does not compete with language models. It makes them safe to use in banking: in Zoral fOS, these controls come with the platform. Because models connect through one interface, you can change models later without changing your workflows. Your engineers stay free to connect their own agents and models.
Atelier, Zoral’s senior consulting practice. Its data scientists and credit specialists tune the models on your data and set up agents to your policies. The people who built the platform are part of every programme.
Data, AI & Automation · Credit risk and decisioning · Engineering and delivery
Product brief
9 pages to read and share with your colleagues: what it changes, how it is controlled and audited, and how it works with the tools you already own.
Markets and regulation keep changing. Your teams change models, agents and rules on one platform. They do not rewrite workflows each time.
Prefer to pick a time? Schedule a meeting
Automate lending, banking and finance operations.
London
Miami
We are on Socials
+44 20 3963 2002