StepInsight

AI & Technical Terms

A plain-English primer — enough to hold your own in a CTO conversation.

Reference
Prepared by StepInsight
6 July 2026

The aim isn't to make you technical — it's to make you fluent. Enough of the AI vocabulary to follow a technical conversation, ask the right questions, and not get caught out. Grounded in real analytics work: margin and trade-spend modelling — reconciliation, elasticity, scenarios — plus an AI narrative written over the finished numbers.

How to read this

Essential. Everyday language, or it matters directly to how we pitch and secure this kind of work. Learn only these and you're covered.

Good to know. Useful depth — reach for it when a conversation goes deeper.

The three that matter most — if you remember nothing else: deterministic vs probabilistic, data residency, and hallucination / human-in-the-loop. They're the spine of "the numbers are auditable, your data stays put, and a person signs off."

Tier 1

Entry level — what AI is at all

11 terms · 9 essential, 2 good to know

Term	In plain English	Why it comes up
Artificial Intelligence (AI)	Software that does things we'd normally say need human judgement — reading text, spotting patterns, answering questions. It's an umbrella word.	Everyone uses it loosely; useful to know it's a category, not one thing.
Machine Learning (ML)	A type of AI that learns patterns from examples/data instead of being hand-coded with rules.	The elasticity/forecasting side of our work is ML, not "chatbot AI."
Model	The trained "engine" that takes an input and gives an output — a number, a prediction, or some text. When people say "the model," this is the thing doing the work.	You'll hear "the model" constantly; it just means the working part.
Algorithm	The step-by-step method behind the scenes — the recipe. Often used loosely to mean "the logic that does X."	CTOs say it a lot; it's rarely as mysterious as it sounds.
Training / training data	The examples you feed a model so it learns the patterns. Good data in → good answers out; rubbish in → rubbish out.	Lets you ask the right question: "what was it trained on?"
Large Language Model (LLM)	A model trained on huge amounts of text that can read and write language — ChatGPT, Claude, Gemini. This is what most people now mean when they say "AI."	The part that would write the plain-English narrative over our numbers.
Generative AI	AI that creates new content (a summary, a paragraph, an image) rather than just scoring or sorting.	Distinguishes the "writes you a summary" part from the "does the maths" part.
Prompt	The instruction or question you give an LLM. "Prompting" is just how you ask.	Demystifies the interaction — it's typing a good brief.
Structured vs unstructured data	Structured = tidy rows and columns (a spreadsheet, a database). Unstructured = free text, PDFs, emails. Different tools handle each.	Their remittance/settlement data is structured — a point in our favour.
Input / output	What goes in (your data, your question) and what comes back (the answer, the number, the summary).	Simple framing for "where does my data go and what do I get."
Automation	Using software to do a repetitive task without a person each time. Not always "AI."	Useful to separate plain automation from actual AI.

Tier 2

Intermediate — how it works, and how you'd use it here

18 terms · 10 essential, 8 good to know

Term	In plain English	Why it comes up
Inference	The moment a trained model is actually used to produce an answer (as opposed to being trained). "Running inference" = using the model.	You'll hear "inference cost" — it's the cost of using it, not building it.
Token	The unit LLMs read and write in — roughly a word-piece. Usage and cost are measured in tokens.	Explains why AI is priced "per token" and why big documents cost more.
Context window	How much text a model can hold in mind at once (your data + your question + its answer). Too much and it won't all fit.	Relevant to "can it read our whole dataset in one go?" (usually: summarise first).
Hallucination	When an LLM says something plausible but wrong or made-up. The central reliability risk of generative AI.	The thing a CTO will raise. Our answer: the numbers are done by maths, not the LLM.
Deterministic vs probabilistic	Deterministic = same input always gives the exact same answer (like a spreadsheet formula). Probabilistic = an LLM might word things differently each time.	The crux of our pitch. Reconciliation & elasticity are deterministic and auditable; the LLM only writes the narrative.
Regression	A long-established statistical method that fits a relationship between things — e.g. how volume moves when price moves.	This is the elasticity maths. Decades old, explainable, not AI hype — a strength.
RAG (Retrieval-Augmented Generation)	Instead of retraining a model, you let it look things up in your documents/data at answer-time so it stays grounded in your facts.	How you keep an LLM honest with company data without handing the data over to train on.
API	The "plug" that lets one piece of software call another (e.g. our app calls Claude's API). How most AI actually gets integrated.	Comes up in any "how would this connect to our systems?" chat.
Guardrails	The rules and limits put around an AI so it stays in-bounds — won't touch certain data, won't answer certain things.	Reassures on control; pairs with the security story.
Human-in-the-loop (HITL)	A person reviews and approves the AI's output before it's used or sent.	The main answer to "how do we stay accountable?" Keep this one.
Parameters / weights	The internal numbers a model learns during training — roughly, its "knowledge." Bigger models have more.	Only if a CTO gets specific about model size.
Fine-tuning	Taking a general model and training it a bit further on your specific data or task.	Worth knowing it exists, and that RAG is often the lighter alternative.
Classification	Sorting things into categories — on-promo vs off-promo, valid vs disputed.	Names a common, boring, reliable use of ML.
Supervised vs unsupervised learning	Supervised = trained on labelled examples with known answers. Unsupervised = the model finds structure on its own.	Only if the conversation goes into how a model learns.
Embeddings	Turning text into numbers that capture meaning, so a computer can find "similar" things.	The engine under RAG / semantic search; skip unless they dig in.
Vector database	Where those embeddings live, so you can search by meaning rather than exact keywords.	Same — only if RAG comes up in detail.
Agent	An AI set up to take a few steps on its own toward a goal (fetch data, use a tool, then answer) rather than replying once.	Buzzword of the moment; good to recognise, not central to our pilot.
Temperature	A dial for how "creative" vs "predictable" an LLM's wording is. Low = more consistent.	Minor, but shows control over consistency if asked.

Tier 3

Advanced — what a CTO might actually probe

17 terms · 9 essential, 8 good to know

Term	In plain English	Why it comes up
Foundation / frontier model	The big general-purpose base models (Claude, GPT, Gemini) that everything else is built on. "Frontier" = the most capable current ones.	A CTO will ask which model; good to know it's a small set of these.
Open-weight vs closed model	Open-weight = you can download and run it on your own kit. Closed = you can only reach it through the vendor's service.	Directly tied to control and where the data goes — our security story.
Managed inference / model hosting (e.g. AWS Bedrock)	Running a model through a cloud provider's service instead of your own servers. Bedrock, for instance, lets you call Claude inside your own AWS account.	Central to "your data stays in your environment" — Bedrock in the client's own account.
Self-hosting	Running the model on infrastructure you fully control — for maximum data control or an air-gap.	The strongest-control option; know it exists as the far end of the dial.
Data residency	Which country / jurisdiction your data physically sits and is processed in.	A top compliance question — and a place we're strong (AU soil, in-boundary).
Confidential computing / secure enclave (TEE)	Hardware that keeps data encrypted even while it's being processed, so even the operator can't see it.	Underpins "not even we can see your data" — our sealed-enclave option.
Evaluation / "evals" / benchmarks	How you measure whether a model is good enough for the job, with defined tests.	Answers "how do you know it's accurate?" — shows rigour.
Explainability / interpretability	Being able to show why a model gave an answer.	Deterministic maths (regression) is inherently explainable; a real advantage to name.
Prompt injection	A security risk where hidden instructions buried in data trick an AI into misbehaving.	A sharp CTO's security question; good to recognise the term.
Transformer	The underlying architecture behind modern LLMs. You don't need the maths — just to recognise the word.	Only so it doesn't land as a mystery if a CTO says it.
Multimodal	A model that handles more than text — images, audio, tables — in one system.	Nice-to-know; not central to our numbers work.
MLOps	The discipline of deploying, monitoring and maintaining models in production — "DevOps for models."	Comes up if they ask how it's run and kept healthy over time.
Model drift	When a model's accuracy quietly degrades over time as the world moves away from its training data.	Pairs with evals — "how do you keep it accurate as things change?"
Bias / fairness	Models can inherit skews from their training data; matters most where decisions affect people.	Less central to margin maths, but a common governance question.
Chain-of-thought / reasoning	Prompting or model designs that make the model work through steps before answering; helps on harder problems.	Buzzy; recognise it, don't lead with it.
Quantization / distillation	Techniques to make models smaller, cheaper and faster (compress a model; or train a small one to mimic a big one).	Only if cost/performance gets deep.
Model Context Protocol (MCP)	An emerging standard for plugging tools and data sources into AI assistants.	Very current; optional unless integrations come up.