“Conversational fluency is not the hard part. Control is.”
From scripted chatbots to governed conversations
For years, a banking chatbot was a separate tool. It matched a request to a short list of topics and returned a scripted reply. Language models have changed what customers expect. They can describe their situation in their own words and get a natural answer.
For a regulated bank, a fluent answer is not enough. A model may guess its way through onboarding. It may not explain its choices. It may reach into systems it should not touch. Each of these is a compliance risk.
Zoral takes a different approach. In the Zoral Finance Operating System (fOS), chat is not a separate application added to the bank. It is one more way into the same onboarding journey:
- the chat screens are built in Zoral Portal Studio;
- the process is run by Zoral Core Orchestrator;
- the decisions are made by Zoral Decision Engine;
- the language models are connected through Zoral Intelligence Layer, using a standard interface (MCP).
The same workflows, credit rules, permissions and audit trail apply whether the customer fills in a web form, uses the mobile app or talks to the chat window.
Four components, one place to control them
Zoral Portal Studio: the chat window
Zoral Portal Studio builds web, mobile and chat screens in one studio. The chat window is set up like any other screen, and released through the same approvals. The bank’s security, workflows, credit rules and audit trail apply to chat exactly as they do to every other screen.
Zoral Core Orchestrator: the onboarding process
The workflow sets the structure: the steps, the data required, the checks and the approvals. AI agents handle the parts that need interpretation, such as reading a free-text answer. The workflow can ask an agent for help. The agent returns its answer in a fixed format, and the workflow checks it. The workflow keeps the state of the application and records every step in Zoral Data Fabric.
Zoral Decision Engine: the decisions
Onboarding decisions follow the identity, fraud and credit policies the bank has configured. Zoral Decision Engine brings together document data, internal and external data, and open-banking data. It explains each decision in plain English, so risk managers and auditors can see the logic at any time.
Zoral Intelligence Layer: the governed connection to language models
Zoral Intelligence Layer connects agents to the language models, prompts and knowledge sources the bank approves. Agents read only approved data and use only approved tools, through separate, controlled connections. They have no direct access to your core systems. Your bank chooses its language models, through enterprise model services or open-weight models it hosts itself.
Architecture
Conversational onboarding on Zoral fOS
How it works
A message from the chat widget travels through the onboarding workflow, an AI agent connected to tuned LLMs through MCP, and the Decision Engine — all inside the Zoral fOS security and governance. Hover or tap a block to see what it does.
The onboarding journey, step by step
The conversation starts. A prospective customer opens the chat window in a Zoral Portal Studio portal. The portal starts the onboarding workflow in Zoral Core Orchestrator, which holds the application from that moment.
The agent gathers the information. The agent asks for what the workflow needs and interprets the customer’s answers. What it extracts goes into the workflow’s structured fields. The conversation is flexible; the data model is not.
Documents go to Zoral Document Fabric. When the journey needs identity or financial documents, the workflow requests them. Zoral Document Fabric, our intelligent document processing (IDP) service, integrated with Zoral fOS, sorts them, reads the data and checks it across documents and with external sources.
Zoral Decision Engine decides. With the data in place, the workflow calls Zoral Decision Engine. The identity, fraud and credit policies run as configured. The result is a decision with a full audit record, not a model’s opinion.
The agent explains; people handle the exceptions. The agent tells the customer the outcome and the next steps in plain language. Cases outside the configured policies go to the bank’s staff for review, through the approvals set in Zoral Core Orchestrator.
The journey continues in any channel. The chat window uses the same application as the web, mobile and branch screens. The customer can continue on the website or in the app. A relationship manager can open the same application in the branch, without asking the customer to start again.
Onboarding journey
Six steps from first message to decision
How it works
The highlighted step follows the application through the journey. Hover or tap a step to read it in full.
Governance and human review, built in
Conversational onboarding works for a regulated bank only if every automated step can be limited, traced and explained. In Zoral fOS, these controls are part of the design.
Governance
Every agent action stays inside bank-defined controls
How it works
Each agent action passes through every control layer before it reaches the bank's systems. Hover or tap a layer to see what it enforces.
- Agents act only within the bank’s permissions The bank decides which tools, workflows, payment systems and data fields an agent may use. Prompts and guardrails block instructions the bank has not approved.
- Full traceability Zoral Core Orchestrator keeps a permanent record of every workflow step, prompt, guardrail setting and agent action. Zoral Decision Engine records every change to rules, tables and policies.
- Decisions explained Every credit score, fraud check and pricing decision comes with a plain-English explanation, ready for audit.
- Personal data protected Personal data can be encrypted, masked or removed field by field, inside the workflow.
- People approve what matters Zoral Intelligence Layer adds human review points, and Zoral Core Orchestrator combines automated checks with staff approvals.
- Changes are tested first Risk managers change agent instructions in Zoral Core Orchestrator’s visual studio. They test new rules in a simulator and run regression tests before customers see them. New policies in Zoral Decision Engine can run in shadow mode on live volumes without changing real approvals.
- Credit decisions stay with the bank AI agents do not make credit decisions. Those follow rules and models the bank approves, and accountability stays with the bank.
For your architects
The chat window keeps a live connection to the workflow (WebSockets and GraphQL), so the screen updates without reloading. Agents reach approved tools and data through MCP servers that Zoral Intelligence Layer configures. Zoral Decision Engine also offers an MCP interface, so the bank’s own agents can use approved decision services.
Conclusion
Language models make conversation a natural way to open a banking relationship. Zoral fOS makes it a governed one. The conversation is handled by the language models your bank chooses. The process is run by Zoral Core Orchestrator. The decisions are made by Zoral Decision Engine. And every step stays within the same permissions, audit trail and human review as the bank’s other channels.
+44 20 3963 2002