glossary
the wordspeople use loosely.
- forward deployed engineer
- an engineer who works inside your team and your codebase rather than delivering from a distance. the model palantir made common, and the one anthropic, openai, aws and microsoft all put money behind in 2026. the point is that whoever scopes the work is whoever ships it.
- read the full definition of forward deployed engineer
- ai consultancy
- advisory work that ends in something running, not a deck. a readiness assessment, a written plan, and usually the implementation. ask any firm two questions: what does it cost, and who exactly does the work.
- ai readiness assessment
- a short review of whether your data, systems and team can support the ai work you are considering. it should name the workflows worth changing and the ones that are not ready, and be honest enough to say 'not yet' when that is the answer.
- ai implementation
- the step between deciding to use ai and having it run in production: the integration, the access controls, the rollback plan, and the handover to whoever maintains it. most stalled ai programmes stall here, not at the model.
- ai governance
- the written rules for how ai is used in a company: what data it may see, who signs off, what is logged, and how a bad output gets caught. concretely a policy document, an audit log and a rollback plan — not a statement of principles.
- ai pilot
- a limited trial run of an ai workflow before committing to it. a pilot is useful when it has an exit criterion written down in advance. without one it becomes the thing that runs for a year and never ships.
- ai enablement
- training that changes what a team does on monday, not what they know on friday. cohort dates, a real assessment, and work produced during the programme rather than a certificate of attendance.
- large language model (llm)
- a model trained to predict text, which is why the same question can produce two different answers. useful when the task tolerates review; risky when an unchecked output goes straight to a customer or a ledger.
- retrieval-augmented generation (rag)
- giving a model your documents at question time instead of training them into it. it is how an assistant answers from your policies rather than from the internet, and it is usually cheaper and easier to correct than fine-tuning.
- ai agent
- a model given tools and permission to act — send the email, update the record, call the api. the engineering problem is not the reasoning, it is deciding what it may touch and what it must ask a human about first.
- eval
- a repeatable test that scores a model's output against known-good answers. without one you cannot tell whether a prompt change improved anything or merely changed it. this is the difference between shipping and guessing.
- fine-tuning
- further training of an existing model on your own examples so it adopts a format or a style. worth it for consistency of output; usually the wrong tool for teaching a model facts, where retrieval does the job better.