Glossary
Glossary
Short definitions of terms that come up in our work and on this site. Half the confusion in an AI project starts with two people meaning different things by "agent" or "prototype".
A
- AI agent
-
Software that uses an AI model to plan and carry out multi-step tasks, calling tools and systems along the way, within limits set by its designers.
- AI strategy
-
A decision about which problems an organization should solve with AI, in what order, and what solving each will take.
B
- Build or buy
-
The decision between building a custom solution and buying an existing product, made per problem by comparing fit, lifetime cost and strategic value.
C
- Custom AI solution
-
Software written for one organization's problem that uses AI, such as a language model or an agent, somewhere inside it.
E
- Evaluation
-
Testing an AI solution against a set of real cases and agreed measures of a good result, before launch and every time the solution changes.
F
- Fine-tuning
-
Further training an existing AI model on examples from a specific task or domain, to change how it behaves.
- Forward-deployed architect
-
An architect embedded with a customer who designs how a solution fits the customer's systems, data, security and operations.
- Forward-deployed engineer
-
An engineer who works directly inside a customer's organization, learns the problem from the people who have it, and builds the software where it will be used.
G
- Grounding
-
Giving an AI model trusted source material, such as an organization's own documents or records, so its answers are based on that material rather than on its general training.
- Guardrails
-
Controls that limit what an AI solution can do and catch it when it goes wrong, such as restricted actions, input and output checks, and required human approval.
H
- Hallucination
-
An AI output that is fluent and confident but not supported by fact or by the source material it was given.
- Human in the loop
-
A design in which a person reviews or approves an AI system's work at defined points, typically before costly, irreversible or high-risk steps.
L
- Large language model (LLM)
-
An AI model trained on large amounts of text that can read, write, summarize and reason over language. Most AI agents and assistants are built on one.
P
- Pain-point map
-
A record of where time, money and quality are lost in a process: each pain point, who feels it, how often, what it costs and what causes it.
- Production-ready
-
Integrated, secured, evaluated against agreed measures, guarded against failure, monitored and owned: fit for daily use rather than demonstration.
- Prototype
-
An early working version of the software, run on the organization's own data and tried by the people who will use it before the main build starts.
R
- Retrieval-augmented generation (RAG)
-
A technique that retrieves relevant passages from a body of documents and gives them to a language model, so its answer is grounded in that material.
S
- Solution blueprint
-
The design of a solution before it is built: architecture, data, AI components, integrations, security and the plan to build it.
U
- Use case
-
A specific problem, with specific users and a measurable outcome, that a solution is designed to address.
W
- Workflow automation
-
Software that carries a process from start to finish across the tools and data involved, with people handling only the steps that need judgement.
