Glossary

The unit of perception in AI — what the system takes in before it acts. Perception before action is how we work, and how reliable agents are built.

Software that uses a model to decide steps and call tools — APIs, databases, workflows — to complete a task. Not the same as a chatbot that only generates text.

An assistant embedded in a workflow that drafts, suggests or retrieves — with a human approving outcomes. Narrower autonomy than an agent.

Automated tests that run against a model or agent before release: representative queries, edge cases, compliance scenarios and regression checks. Part of CI, not a one-off benchmark.

The operational discipline for large language models in production: deployment, versioning, monitoring, evaluation, incident response and cost control — parallel to MLOps and DevOps.

When production behaviour degrades over time — because data changed, usage shifted, models updated or retrieval corpora went stale. Detected by monitoring and evaluation, not user complaints alone.

Fetching relevant documents or records from a corpus to ground a model's response. Retrieval quality depends on data curation, chunking, metadata and refresh — it is a data product.

When a model calls structured functions — read account, create ticket, post journal — instead of guessing. Requires reliable APIs, clear schemas and error handling.

Rules and classifiers that block or redirect unsafe inputs and outputs: policy violations, PII leakage, off-topic requests, unauthorized actions.

Structured attempts to break the system — prompt injection, jailbreaks, data exfiltration — so fixes land before users or attackers find them.

Step-by-step instructions for operators: deploy, rollback, respond to incidents, refresh retrieval, escalate to engineering.

Architecture combining data lake flexibility with warehouse reliability — often the foundation for analytics and AI workloads on open table formats.

A managed repository for ML features with consistent definitions between training and serving. Reduces silent skew when models go to production.

Database optimised for similarity search on embeddings. Used for retrieval when the corpus is unstructured text or documents.

Recorded path from source data through transformations to model input or agent action. Required for audit and debugging.

Where data is stored and processed geographically. Configured in infrastructure and contract, not assumed.

Discipline for cloud and inference spend: visibility, budgets, allocation, optimisation. For AI, model routing and caching are FinOps levers.

Fixed-scope, fixed-price, fixed-date delivery slice. The first increment usually ends in a production release by week 12.

Deliverable at the end of two-week discovery: ranked use cases, architecture, data readiness, governance baseline, cost model and build proposal.

The risk that a model behaves incorrectly or outside intended use — with financial, regulatory or safety consequences. Managed through evaluation, oversight, logging and limits on autonomy.

Human and automated oversight of agent actions: sampling, dashboards, escalation queues, approval gates.

The work of getting people to use the system correctly: training, comms, feedback loops, metrics on override rates and trust.

Authoritative applications where business data lives — ERP, CRM, core banking, case management. Agents integrate here, not in spreadsheets.

Operational technology — plant-floor systems, SCADA, industrial control. Different security and availability rules than IT. Relevant for manufacturing and energy agents.

Network isolation with no internet path. Some OT and high-security sites require deployment models that account for air-gapped constraints.

Instructions and context sent to a model. Prompts are versioned, tested and reviewed like code — not edited ad hoc in production.

Coordination of multiple steps, models or agents in a workflow — with state, retries and human handoffs.

A person approves, edits or completes selected steps. Required for high-risk actions and early rollout phases.