New Joiner Development · Reference

AI Foundations — reference sheet

A practical reference for briefing AI, reviewing outputs and using approved tools responsibly.

Structure a clear brief — R · C · T · F

Role

Define the relevant perspective and expertise: “Act as a senior consultant drafting professional client communication.”

Context

Provide the material facts, audience, purpose, source information and constraints. Do not include data that is not authorised for the tool.

Task

Use a clear action and quality criteria: “Draft a polite but clear follow-up that confirms Friday's deadline and offers support.”

Format

Specify the structure, length and language: “Subject line plus three short paragraphs, under 150 words, in British English.” Add an example when useful.

Then review and refine. Give measurable feedback, correct assumptions and ask the model to identify gaps: “reduce this to 100 words” · “ask up to five clarifying questions first” · “identify weak assumptions and the evidence needed to test them”.

Responsible use in every task

1
Protect data and use approved tools. Confirm that the tool, information and purpose are authorised under EY policy, client commitments and applicable requirements.
2
Verify evidence and exercise judgement. Check material facts, figures, citations and conclusions against authoritative sources. You remain accountable for the work.
3
Be transparent and ask when unsure. Permitted use, review and disclosure requirements can vary. Confirm them with engagement leadership before proceeding.

Choose the proportionate intervention

Start by clarifying the brief. If the model needs current or organisation-specific information, consider authorised retrieval with source citations. Consider fine-tuning when repeatable, evaluated performance gains justify the build and governance effort. In every case, define success, test representative scenarios and design the human review and escalation process.

Ten terms you will hear on engagements

Model / parameters: A mathematical function with adjustable values. Training changes those values to reduce a defined measure of error.
Token: A unit of text that a language model processes. A token may be a whole word, part of a word, punctuation or another character sequence.
Context window: The limited set of tokens available for a response, including the prompt, supplied conversation history and retrieved content.
Temperature: A setting that influences variation in generated output. Lower is generally more consistent; higher allows more variety. It does not guarantee accuracy.
Hallucination: A plausible-sounding but unsupported or incorrect output. Material claims must be checked against authoritative evidence.
Embedding: A numerical representation of content that captures patterns of meaning and use, enabling similarity-based retrieval.
Attention: A mechanism that lets a model weigh relevant parts of the context when building each token representation.
RAG: Retrieval-augmented generation: retrieve relevant material, add it to the context and generate a grounded response that can cite its sources.
Fine-tuning: Further training on examples to adapt model behaviour, style or task performance. It requires clear evidence, evaluation and governance.
Agentic system: A system that combines a model, tools and instructions to plan or execute multi-step work within defined controls.