AI Fundamentals and Fluency for the Scrum Master
What "AI" actually is here
Generative AI, notably large language models, produces output by predicting likely next tokens from patterns in training data. It has no model of your team, your product, or your Definition of Done. Predictive machine learning (classification, forecasting, recommendation) is a different family of tools with different failure modes - do not reason about one using the other's rules.
Core vocabulary you should be able to use precisely:
- Tokens and context window - what the model can "see" in one interaction; content outside the window is invisible to it.
- Training vs. inference - the model is built once, then used; it does not learn from your session.
- Temperature and sampling - variation between runs is a designed feature, not a bug.
- RAG (retrieval-augmented generation) - facts supplied at query time from your documents.
- Fine-tuning - additional training that shapes behaviour and style rather than acting as a reliable fact store.
Traps that appear in questions
- Believing that an identical prompt must produce an identical answer. Temperature 0 reduces variation; it does not guarantee determinism.
- Confusing RAG with fine-tuning, or assuming a model "knows" your internal policies.
- Anthropomorphising: treating confident tone as evidence of correctness. Hallucination is fluent and plausible by design.
- Treating AI output as a measurement rather than a hypothesis.
Rule of thumb
Every AI output enters Scrum as an unvalidated artifact until a human checks it against a source of truth. Fluency means knowing which questions to ask of the output, not memorising model internals.
Worked example
A Scrum Master asks a model to summarise retrospective notes twice and receives two different summaries. The correct reading is not "the tool is broken" but "output is probabilistic"; the response is to compare both against the raw notes and keep only claims the notes support.