Glossary
Explainable AI, in plain language
Every term you will hear from us, defined in a sentence or two, without the jargon.
By product part
The Rulebook decides. Agents do the work. The Ledger proves it.
Most terms belong to one of Palladium’s three parts. Start here for the big picture.
Decides in code, against your policy.
Seals every step for audit.
A to Z
Every term, one link each
Each definition has its own anchor, so you can link straight to it.
- Agentic AI
- AI that takes actions toward a goal, such as updating a record or issuing a credit, instead of only answering. In Palladium, an agent acts only after the Rulebook approves. Technical overview
- AI audit trail
- The record of what an AI system was asked, what made the decision, and what it said. In Palladium this is the Ledger: sealed, replayable and exportable. What the Ledger records
- Audit trail
- A timestamped, sealed record of every step in a conversation: what was asked, which rule decided, and what was said. See a decision replay
- Concierge
- The Palladium agent that talks to your customers and staff on voice, chat, SMS, WhatsApp and email. It explains decisions. It never makes them. Agents
- Conversational AI
- AI that understands and replies in natural language across chat, voice, SMS and email. Palladium uses it to explain decisions, not to make them. Chatbot vs. agent
- Data stays put
- Sensitive details are masked before anything reaches an AI model, and your data is never used to train models, by default and by contract. Trust and security
- Decision trace
- The step-by-step replay of a single request, from the customer’s question to the sealed record. Replay one
- Deterministic AI
- A design where written rules make the decision, so the same input always gets the same outcome. A language model only reads the request and explains the result. Explainable AI
- Deterministic core
- An earlier name for rules in code: the part of Palladium that decides, using your written rules, so the same input always gets the same outcome. Today it is called the Rulebook. The Rulebook
- Deterministic guardrails
- Hard limits, such as dollar thresholds, enforced in code so no AI model can go past them. AI compliance checklist
- Explainable by design
- Every answer can be traced back to the rule and the record that produced it. Explainable AI
- Human-in-the-loop
- A person takes over wherever a decision is ambiguous or high-stakes, with the full context handed over. Human-in-the-loop AI
- Intent whitelist
- The fixed list of topics the Concierge may handle. Anything else goes to a person. Example: legal case status
- Ledger
- Where every sealed record lives. Records are replayable and exportable, and nothing written to the Ledger ever changes. Trust and security
- LLM guardrails
- Controls that limit what a language model can see, say and do: input screening, output screening, and hard limits kept outside the model. See the controls
- Model risk management
- How banks develop, validate, monitor and govern the models they rely on. In the US, current expectations are set out in SR 26-2. SR 26-2 explained
- Orchestrator
- Routes each request to the right specialist agent, such as disputes, payments or identity. All of them run on the same Rulebook and Ledger. Specialist agents
- Private deployment
- Palladium running fully inside your own cloud (AWS, Azure or Google Cloud), with private model options. Deployment options
- Proactive trigger
- An event, such as a fee posting or a deadline approaching, that starts the right conversation before the customer asks. Platform
- Provisional credit
- A temporary credit a bank gives a customer while it investigates an electronic transfer error under Reg E. Reg E disputes
- Reg E
- Regulation E, the US rule under the Electronic Fund Transfer Act that sets how banks resolve consumer disputes over electronic transfers, including deadlines. Reg E disputes
- Rulebook
- Your policies, thresholds and eligibility rules, written as code that Palladium enforces before any model sees the request. The Rulebook
- Sealed record
- A record that cannot be edited after the fact, so an examiner can trust what it shows. See one
- SR 26-2
- Model risk management guidance issued by the Federal Reserve, OCC and FDIC in April 2026. It revises the SR 11-7 framework. SR 26-2 explained
- UDAAP
- Unfair, deceptive, or abusive acts or practices: the consumer-protection standard the CFPB and bank regulators enforce. Treating customers consistently is the first defense. Palladium for banking
- Warm handoff
- Passing a conversation to a person with identity, progress and context already verified. For service teams
See every term in action
A live walkthrough, not a slide deck. We resolve a real request end to end and replay the full audit trail on screen.