Agent Harness
The software around an AI agent that manages its task loop, tools, progress and controls.
The term overlaps with AI harness; check the actual responsibilities supplied by the application or platform.
AI, explained clearly.
Definitions for AI engineering and managed business automation, with practical examples and links to the relevant work.
58 terms
Architecture
Agent Orchestration
Coordination of the steps, tools and state used by one or more agents in a workflow.
It can combine fixed software paths and model-selected steps. It does not inherently require many agents.
Agentic AI
A broad label for AI systems able to help plan or carry out sequences of work, often through tools.
Describe the actual autonomy and available actions rather than relying on the label.
Agentic Workflow
A workflow in which a model helps choose steps or tool requests within a defined task boundary. Implementations vary in how much planning is model-directed.
The application still needs authoritative state, permissions, validation and stopping conditions.
AI Employee
A business description for software carrying out a bounded recurring role with defined access and supervision.
It does not imply a human worker, independent professional authority or an unlimited ability to act.
Example A defined role might read incoming requests and prepare drafts, with staff reviewing the actions that need approval.
AI Harness
The surrounding runtime and controls that connect a model to a defined task, including context, tools, state, permissions, verification and operating evidence as needed.
It is broader than the model and is not a universal proprietary product or mandatory component list.
Example An agent prepares a record change. The harness checks access, holds it for approval and records the result.
Deterministic Workflow
A workflow whose transition rules are specified in software, while external inputs and dependencies can still vary.
Predictable reminders or routing may suit fixed rules without model-directed planning.
Example A confirmed appointment triggers a reminder at an agreed time.
Model Context Protocol (MCP)
An open protocol for connecting AI applications with external capabilities and data through supported clients and servers.
Using a protocol does not automatically make a tool trusted or authorised. Identity, permissions and validation still need design.
Multi-Agent System
A system coordinating multiple agents or model-driven roles, potentially with different tasks or tools.
Multiple agents add communication and failure paths; they are not required for every useful AI application.
Tool Calling (Function Calling)
A model interface for proposing a structured request to an external function or tool.
Application code validates, authorises and executes the request. A proposed call is not evidence of a completed action.
Example A model requests a customer lookup. Application code checks permission before querying the customer system.
Governance
AI Bias Audit
An assessment of defined performance or impact differences across relevant groups, using an explicit scope, data and evaluation method.
Finding no disparity in one test does not prove that a system is fair in every setting.
AI Governance Framework
An organised set of responsibilities, policies and controls for deciding how AI is developed, used, reviewed and operated.
Its value depends on implementation and evidence, rather than the existence of a policy document alone.
Audit Trail
A chronological record of actions, checks and decisions that helps a reviewer establish what happened.
Its usefulness depends on what is recorded, how records are protected and who can access them. It does not reveal a model’s complete internal reasoning.
Example A record shows the proposed change, the person who approved it and the result returned by the connected system.
Confidence Scoring
A method of assigning a score to an estimate, prediction or output. Its meaning depends on how it is produced and validated.
A model’s self-reported percentage is not automatically a calibrated probability or authority to execute an action.
Explainability
The ability to provide information that helps a person understand aspects of a system, output or decision.
A generated explanation is not necessarily a faithful causal account of the model’s internal computation.
Human-in-the-Loop (HITL)
A process in which a person participates at a defined stage of AI-assisted work, such as review or approval.
The person needs suitable evidence and authority, and software must enforce approval where required.
Example A reviewer sees the proposed action and its supporting information before deciding whether it can proceed.
Least Privilege
Giving a person or software process only the access needed to carry out its assigned task.
Apply the boundary in the connected systems and application code. A prompt asking a model to stay within its role is not an access control.
Example An agent can read appointment availability without being able to delete calendar entries.
Prompt Injection
Instructions in user input or external material that try to redirect a model away from its intended task or rules.
Retrieved documents and tool results can carry these instructions. Permissions, validation and approval controls help limit the actions a manipulated model can cause.
Example A retrieved document tells an assistant to ignore its task and send customer records elsewhere.
Risk Scoring
A defined method for summarising selected risk factors to support prioritisation or review.
A score does not replace legal classification, task-specific evidence or a responsible person’s judgement.
Shadow AI
AI tools or uses operating outside an organisation’s approved inventory, access arrangements or oversight.
Discovery and usable approved alternatives can help bring actual data flows and responsibilities into review.
Models and retrieval
Chunking
Dividing source material into units for indexing or model processing.
Poor boundaries can separate a value from its heading or remove a necessary qualification.
Context Window
The capacity available for a model request under its interface’s token-accounting rules.
It is not durable application memory; important task state should remain in an authoritative store.
Embedding
A numerical representation used to compare or process items such as text passages.
Similarity is useful for retrieval but is not proof that a passage supports an answer.
Fine-Tuning
Additional training that updates model parameters using examples for a particular task or behaviour.
It can complement retrieval, but does not provide a dependable inspectable store of changing business facts.
Hallucination
Generated content that is unsupported, inaccurate or fabricated relative to the task or evidence.
A fluent response and a genuine citation can still contain an unsupported claim.
Hybrid Search
Combining retrieval methods, often lexical and vector search, to select relevant information.
The combination needs testing on actual queries; it does not guarantee a fixed relevance improvement.
LLM (Large Language Model)
A language model trained at scale to model and generate sequences, sometimes with additional modalities.
An application adds retrieval, tools and controls; these capabilities should not be confused with inherent model authority.
LoRA (Low-Rank Adaptation)
A parameter-efficient adaptation method that trains low-rank updates while keeping base-model weights fixed.
Quality, memory and cost effects depend on the model, task and configuration; there is no universal performance percentage.
Prompt Engineering
Designing and evaluating the instructions, examples and information supplied to a model.
A prompt can guide behaviour but is not an enforceable permission boundary.
RAG (Retrieval-Augmented Generation)
Supplying externally retrieved information to a model when it generates a response.
Source quality, access, freshness, retrieval relevance and answer support all need evaluation.
Example A team member asks about an internal policy. The system retrieves relevant passages and includes links with its answer.
Re-Ranking
Scoring or reordering retrieved candidates with another relevance method before selecting evidence.
It may improve selection but adds latency, cost and another component to evaluate.
Token
A unit used by a model’s encoding of information, such as part of a word or another representation.
Counts depend on the tokenizer and modality. Word counts do not translate into one universal token ratio.
Operations
Circuit Breaker (AI)
A mechanism that stops or limits a defined class of work when specified failure or risk conditions occur.
Stopping new requests does not automatically reverse actions already committed in external systems.
Evaluation
Testing an AI system against defined tasks and criteria to understand its behaviour and limitations.
Use representative work and failure cases. Check the application’s permissions and tool outcomes as well as the model’s answers.
Example A knowledge assistant is tested on supported questions, missing information and documents the user may not access.
Idempotency
A property where repeating an operation has the same intended effect as performing it once, within the relevant system contract.
It helps handle retries, but requires appropriate keys or target-system guarantees; not every external action supports it.
Example A retried request uses the same operation identifier so a supported payment interface can recognise an existing charge.
MLOps
Practices and tooling for developing, releasing, monitoring and maintaining machine-learning systems.
Ownership, versioning, evaluation and recovery remain necessary when models are used through hosted interfaces.
Model Drift
A change in model behaviour or performance over time. Related application degradation can also result from changing input data, retrieval or tools.
Diagnose the failing layer rather than assuming every decline requires retraining.
Shadow Mode
Running a candidate process in parallel for comparison without letting its outputs drive the main operational action.
It may still process sensitive data and needs appropriate access, privacy and operating controls.
System of Record
The authoritative source for a particular class of business information, such as an accepted payment or approved job status.
A model summary or local copy should not silently replace the authoritative record.
Example A payment workflow checks the accounting system before marking an invoice as paid.
Privacy and regulation
Data Sovereignty
Control and legal considerations concerning data, including the jurisdictions and authorities governing its processing and access.
Data residency is one aspect. Contracts, remote access and transfers may remain relevant even when storage stays local.
EU AI Act
The EU regulation establishing obligations for specified AI practices, systems and models according to its scope and actor roles.
Classification and application dates differ by provision. Use current consolidated legislation and guidance.
GDPR (General Data Protection Regulation)
The EU data-protection regulation governing covered processing of personal data and related rights and responsibilities.
It provides several lawful bases, not only consent. Applicability, transfers and automated-decision rules require scoped assessment.
ISO 42001
Common shorthand for ISO/IEC 42001, a standard specifying requirements for an AI management system.
Certification, where pursued, concerns the defined management-system scope; it does not guarantee every model output is correct.
NIST AI RMF (AI Risk Management Framework)
A voluntary risk-management framework with Govern, Map, Measure and Manage functions. It is intended to help organisations address AI risks through the lifecycle.
It is a useful reference, not a regulatory approval or a certification issued to every system using it.
PII (Personally Identifiable Information)
Information that can identify or relate to an identifiable person, with terminology and legal scope varying by jurisdiction.
Removing direct identifiers does not necessarily anonymise a record. Indirect details and linkage may still identify someone.
Infrastructure
ETL (Extract, Transform, Load)
A data-integration pattern that extracts records, transforms them and loads them into a target store.
Schema checks, record ownership and error handling remain important when AI participates in the pipeline.
LLM Gateway
An intermediary service between applications and model endpoints, used for capabilities such as routing, authentication, usage records and policy checks.
A gateway can centralise defined controls, but must not be assumed to enforce every application permission or data rule.
Sovereign AI
An approach to AI emphasising control over data, processing, model access and dependencies within chosen organisational or jurisdictional boundaries. Definitions vary by context.
Hosting location alone does not resolve every access, legal or operational requirement.
Sovereign Infrastructure
Infrastructure arranged to meet defined requirements for jurisdiction, access and operational control. It may involve on-premise or appropriately controlled cloud resources.
Specify the actual processing, contracts and access boundaries rather than treating the label as a guarantee.
Vector Database
A data store supporting search over numerical vector representations.
It can support retrieval, but does not by itself provide source authority, freshness or complete permission handling.
Zero-Retention Architecture
A design intended to limit specified parties from retaining specified data after processing. The exact promise depends on the services, logs, settings and agreements involved.
Non-retention is different from non-processing, no logging anywhere or a guarantee that no information crosses a boundary.
Strategy
Friction Mapping
A practical review of delays, handoffs, repeated work and exceptions in a process.
It can identify opportunities for integration, ordinary automation or AI; it is not an independently validated universal method.
Use Case Prioritization
Comparing proposed uses by their expected value, feasibility, risks and evidence needs.
A useful decision includes the current baseline, accountable owner and constraints, not just an attractive demonstration.
Business automation
Lead Leakage
Gaps in the enquiry-to-sales process where a potential customer is missed, left without follow-up or lost between hand-offs.
Review each stage and its next action before choosing an automation. A recorded enquiry alone does not mean someone followed it up.
Example A missed call creates no record or follow-up task, so the enquiry never reaches the sales team.
Managed Automation
Business workflows that a provider sets up, monitors and maintains for an agreed scope of work.
For Ryzolv Automate, the scope defines the workflows, connections and ongoing management your business receives.
Example Ryzolv configures quote follow-ups, then reviews the workflow with the team when the sales process changes.
Missed-Call Follow-Up
A configured process for acknowledging a missed call, recording the enquiry and routing the next action through supported channels.
Message channels, timing and staff hand-offs depend on the agreed phone and enquiry setup.
Operations Center
A shared system connecting customer records, sales, jobs, schedules, payments and the next actions around them.
Ryzolv Automate connects and manages the agreed workflows. The business still needs people to deliver the work and keep records current.
Example An accepted quote moves into job preparation, while scheduling and customer updates stay connected to the same record.
Workflow Trigger
An event or condition that starts a defined sequence of actions.
A trigger needs a reliable source event and rules for timing, repeat events and exceptions.
Example Marking a job as complete starts an agreed review-request workflow.