Glossary
Plain definitions for the vocabulary this market uses carelessly. Where a term is commonly used to mean two different things, we say which one we mean.
The short version: the vocabulary of AI agents, defined plainly and without hype — what each term actually means in a deployment, not what a vendor deck wants it to mean.
Last updated:
Terms
- Agent
- A system that takes a goal, decides what steps to take, uses tools to take them, and adapts when the result isn’t what it expected. The distinction from automation is the decision: automation follows a path you defined, an agent chooses one.
- AgentOps
- The ongoing work of keeping deployed agents working: monitoring, evaluation runs, model upgrades, integration repair, drift detection. An unmaintained agent degrades quietly rather than failing loudly.
- Agent Opportunity Map
- The deliverable of our audit. Every candidate workflow in a business, ranked by recoverable hours, with integration feasibility tested against live systems and a build estimate for each.
- Confidence score
- A number the agent attaches to an extraction or a decision, indicating how certain it is. Its purpose is routing: low-confidence outputs go to a human rather than being silently accepted.
- Drift
- The gradual degradation of an agent’s output quality after deployment, caused by changing data, changing models, or changing business rules. Detectable only with an evaluation suite; invisible without one.
- Escalation
- The point at which an agent stops and hands to a human. Designing escalation well is a product decision, not an engineering one, and it is where trust is won or lost.
- Eval suite
- A regression test set built from real cases, run against every prompt or model change before deployment. An agent without one is an agent nobody is measuring.
- Fixed price
- A quoted amount that does not change unless you change the scope. Distinct from an estimate, which is a number that grows.
- Hallucination
- A model producing a confident, fluent, wrong answer. In agent systems the mitigation is architectural — retrieval from systems of record, validation rules, and escalation on low confidence — not prompt wording.
- Happy path
- The version of a workflow where nothing goes wrong. Most software handles it. Exceptions are where operations actually spend money.
- Human in the loop
- A design where an agent prepares work and a person approves it. Mandatory for anything financial, clinical, or contractual — and a deliberate design choice rather than a limitation being engineered away.
- Latency
- Round-trip response time. For voice agents the target is under 1.2 seconds; past roughly 1.5 seconds callers start talking over the agent and the conversation degrades.
- LLM
- Large language model. The reasoning component inside an agent, not the agent itself. Confusing the two is the most common category error in this market.
- MCP
- Model Context Protocol. A standard for exposing tools and data sources to models, so an agent can reach a system without a bespoke integration per model.
- Multi-agent system
- Several agents coordinating on a task. Powerful and hard to evaluate. Always a second engagement — never attempt one first.
- Observability
- Tracing what an agent did and why: the tool calls, the reasoning, the confidence, the decision. Without it you cannot debug a failure or defend a decision.
- Orchestration framework
- The layer that manages an agent’s control flow, state and tool calls. LangGraph, the OpenAI Agents SDK and CrewAI are examples. Anyone who uses the same one for every problem is selling their comfort, not your outcome.
- Pilot purgatory
- The state of an organisation with many agent proofs of concept and nothing in production. Usually caused by scoping for demonstration rather than for deployment.
- RAG
- Retrieval-augmented generation. Giving a model access to your documents or data at query time. Effective retrieval is hybrid — semantic plus keyword — because pure vector search fails on exactly the specifics a business cares about: part numbers, permit IDs, unit references.
- Rollback path
- The documented, tested route back to the previous way of working. Every deployment stage has one before it goes live, not after something breaks.
- Shadow mode
- Running an agent alongside your team on live data, logging every action it would have taken without executing any of them. You compare its decisions against your team’s, see the disagreements, and tune thresholds before it goes live.
- System of record
- The authoritative source for a piece of data — your ERP, CMMS, PMS, or practice management system. Agents that only read documentation cannot complete transactions; agents that write to systems of record can.
- The 70% trap
- Most agent projects reach 70% capability quickly and then stall. The remaining 30% — edge cases, error handling, escalation quality — is most of the work and all of the trust.
- Tool call
- An agent invoking an external function: querying a database, sending a message, creating a work order. The mechanism by which an agent acts rather than merely answers.
- Vertical product
- Software built for one industry’s specific workflow. Often the right answer, and if one fits you we will say so instead of quoting you a build.
Frequently asked questions
Why publish a glossary at all?
Because the vendor vocabulary in this category is doing a lot of hiding. Plain definitions let you compare claims across vendors — including ours — on equal terms.
Are these definitions standard across the industry?
Mostly, with honest variation flagged where it exists. Where a term is used loosely in the market (agent, autonomous, human-in-the-loop), the definition here states what we mean when we use it in a proposal.
Can I cite these definitions?
Yes, with a link. They’re written to be quoted.
A note on the words we avoid
We don’t use “AI-powered”, “transformative”, “seamless” or “revolutionise”, because none of them carry information. We don’t call an agent “autonomous” unless it acts without human approval, and we say what it’s allowed to act on.
If a vendor’s vocabulary is doing the persuading, ask what each word means operationally. Vague answers are the signal.