Edison AI
AI Glossary

The Edison AI Glossary.

47 AI terms every Australian SMB operator should understand. Each entry: a plain-English definition, an Australian-context paragraph (regulators, common SMB tool stack, real implementation patterns), and a list of related terms. Use it to brief a board, calibrate a vendor conversation, or train a team.

ABCDEFGHIJKLMNOPQRSTUV

A

4 terms

  • AI Agent

    A software entity built on a large language model that takes a goal, decides what to do next, calls external tools (CRM, email, calendar, documents), and reports back. Unlike a chatbot, an agent acts on systems rather than just answering questions.

    Australian Context

    Most Edison AI agents shipped for Australian SMBs sit inside the existing stack (HubSpot, Xero, Microsoft 365, Cin7, Salestrekker) and handle administrative or coordination work. Intake, follow-up, document collection, reporting. Under ACSC-aligned data boundaries.

    Related Terms

  • Agentic AI

    The class of AI systems that plan multi-step work, invoke tools, observe results, and re-plan. Agentic AI is the practical layer above generative AI: less about producing text, more about doing the job.

    Australian Context

    For Australian operators, agentic AI is the difference between paying for ChatGPT seats and actually moving work off the team's plate. Edison AI's agentic implementations typically replace 8 to 16 hours of weekly administrative coordination per role.

    Related Terms

  • AI Readiness Audit

    A structured diagnostic of a business's workflows, tools, data, team capability, and AI opportunities before any build commitment is made. Produces a costed roadmap rather than a vendor pitch.

    Australian Context

    Edison AI's two-week fixed-fee audit is the standard entry point for Australian SMBs. The output is a sequenced 12-month plan plus quick-win identification, aligned with ACSC Secure AI guidelines and OAIC privacy principles.

    Related Terms

  • ACSC (Australian Cyber Security Centre)

    The Australian government agency responsible for cyber security guidance and incident response. Publishes the Secure AI guidance that Australian businesses use as the baseline framework for safe AI adoption.

    Australian Context

    ACSC's Secure AI guidance sets the practical floor for any AI rollout in Australia. Data classification, vendor due diligence, prompt hygiene, incident response. Edison AI builds and trains to ACSC-aligned standards by default.

    Related Terms

B

2 terms

  • Bias (AI)

    Systematic skew in an AI model's outputs caused by skew in its training data, evaluation, or prompting. Bias can be benign stylistic skew or harmful demographic skew that produces unfair outcomes.

    Australian Context

    For Australian SMBs, the most common bias issue is not demographic. It is recency and US-centric bias. Models often default to American spelling, US tax codes, and US business norms unless prompted otherwise. Edison AI prompt libraries lock to Australian English and AU context.

    Related Terms

  • Benchmark

    A standardised test used to compare AI models on a specific capability. Coding, reasoning, language understanding, multi-step task execution. Public benchmarks include MMLU, GPQA, HumanEval, and SWE-bench.

    Australian Context

    Benchmarks matter less for Australian SMBs than the brutal practical question. Does the model actually do this team's work well? Edison AI evaluates models against representative client tasks (your real invoice chases, your real client briefs) rather than public benchmarks alone.

    Related Terms

C

5 terms

  • Chain-of-Thought

    A prompting technique where the model is asked to reason step by step before giving a final answer. Improves accuracy on multi-step problems but increases token cost and latency.

    Australian Context

    Useful for compliance-sensitive Australian workflows. Quote calculations, file note review, claim assessment, where the trail of reasoning needs to be auditable. Edison AI templates include chain-of-thought scaffolds for high-stakes drafting tasks.

    Related Terms

  • ChatGPT

    OpenAI's consumer-facing chat interface, layered on top of GPT models. The most widely-adopted entry point for non-technical AI use globally, with Free, Plus, Team, and Enterprise tiers.

    Australian Context

    Most Australian SMBs first encounter AI through ChatGPT. The free tier is fine for personal experimentation; ChatGPT Team is the minimum tier for any client data so data isn't used for training. Edison AI advises ChatGPT Team or Enterprise for any team handling Australian PII.

    Related Terms

  • Claude

    Anthropic's family of large language models. Opus, Sonnet, Haiku. Known for longer context windows, careful refusal behaviour, and strong performance on writing and coding tasks.

    Australian Context

    Edison AI defaults to Claude for client-facing drafting and document analysis because of its tone control and long-context behaviour. Anthropic offers an Australian data residency commitment via AWS Sydney for Enterprise customers.

    Related Terms

  • Context Window

    The maximum amount of text (measured in tokens) a model can read in a single prompt. Including system instructions, conversation history, attached documents, and the user's question.

    Australian Context

    Practical Australian use cases that benefit from large context windows.200k tokens and above. Include reviewing a whole client file, a tender document, or a year of compliance logs in one pass. Edison AI uses Claude Sonnet and Opus for these workflows.

    Related Terms

  • Compliance (AI)

    The discipline of running AI in line with regulatory and contractual obligations. Data residency, privacy, sector-specific rules (ASIC, AHPRA, APRA), and internal AI use policy.

    Australian Context

    For Australian SMBs the compliance floor is OAIC privacy principles plus ACSC Secure AI plus any sector regulator. Edison AI's responsible-AI training maps these into a one-page AI Use Policy any team can actually follow.

    Related Terms

D

2 terms

  • Dataset

    A collection of data used to train, fine-tune, or evaluate an AI model. Quality, breadth, and labelling of the dataset are the largest determinants of model behaviour.

    Australian Context

    Most Australian SMBs do not need to train a model. They need to ground existing models in their own dataset (their SOPs, client files, product catalogue) via RAG. That is cheaper, faster, and keeps data under your control.

    Related Terms

  • Deepfake

    Synthetic audio, image, or video generated by AI that depicts a real person doing or saying something they did not. Used in fraud (executive voice cloning) and disinformation.

    Australian Context

    Voice-cloning fraud is the live deepfake risk for Australian SMBs. Fake CFO call requests, fake invoice approvals. Edison AI's responsible-AI training includes voice-verification protocols for any payment or sensitive-data request.

    Related Terms

E

2 terms

  • Embedding

    A numerical representation of text, image, or audio that captures its semantic meaning. Embeddings let you search by meaning rather than keyword, and are the storage layer underneath RAG.

    Australian Context

    Embeddings power Edison AI's SOP knowledge layers, document-Q&A agents, and CRM-grounded responses for Australian businesses. Vector databases (Pinecone, pgvector, Weaviate) hold the embeddings; the LLM reads the most-relevant matches.

    Related Terms

  • Enterprise AI

    AI tooling licensed under enterprise terms. Explicit no-training-on-customer-data, contractual data residency, SSO, admin controls, audit logs. Includes ChatGPT Enterprise, Claude Enterprise, Microsoft Copilot Enterprise, Google Gemini Enterprise.

    Australian Context

    For most Australian SMBs, ChatGPT Team or Claude Team meets the bar without enterprise-grade overhead. Enterprise tier becomes worth it above ~50 seats or when sector regulation (APRA, AHPRA) demands explicit data controls.

    Related Terms

F

2 terms

  • Few-shot Prompting

    Including 2-5 example input/output pairs inside a prompt so the model picks up the desired format, voice, or reasoning pattern without fine-tuning. Cheaper and faster to iterate than training.

    Australian Context

    Few-shot is the right tool for Australian SMBs that want consistent on-brand drafting. Give the model three approved past examples of a client email, ask it to draft the next one. Edison AI prompt libraries are mostly structured few-shot patterns.

    Related Terms

  • Fine-tuning

    Continuing the training of a pre-existing model on a smaller, task-specific dataset so it specialises in a particular voice, format, or domain.

    Australian Context

    Rarely the right first move for Australian SMBs. Prompt engineering plus RAG plus a clean prompt library covers 90% of what fine-tuning would deliver, at a fraction of the cost and risk. Edison AI recommends fine-tuning only after one or two prior iterations.

    Related Terms

G

3 terms

  • GEO (Generative Engine Optimisation)

    The practice of making content findable and citable by generative AI answer engines. ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews. Differs from SEO in that the target is a single extractable answer, not a ranked link.

    Australian Context

    GEO matters most for Australian B2B SMBs that show up in considered-purchase research (legal, financial, healthcare, professional services). Edison AI's AI Marketing & Search Visibility service combines GEO with traditional SEO and AI-assisted content workflows.

    Related Terms

  • GPT

    Generative Pre-trained Transformer. OpenAI's family of large language models (GPT-4, GPT-4o, GPT-5). Also commonly used as shorthand for any large language model, though strictly GPT is OpenAI-specific.

    Australian Context

    Most Australian SMBs encounter GPT through ChatGPT or the OpenAI API. For commercial deployments Edison AI uses the API tier so data isn't retained for training, with the model choice driven by task. GPT-4o for speed, GPT-4-class for hard reasoning.

    Related Terms

  • Guardrails

    Programmatic and procedural constraints around an AI system. Input filters, output filters, prompt-level boundaries, human-in-the-loop checkpoints, scope limitations.

    Australian Context

    Edison AI implementations ship with guardrails as a default, not an add-on. Agents never see PII unless required, never auto-send external communications without review, never execute payments. Maps to ACSC Secure AI principle of human oversight.

    Related Terms

H

2 terms

  • Hallucination

    When an AI model produces confident-sounding output that is factually wrong or invented. Fake citations, fabricated client details, non-existent product features. A function of how language models predict text.

    Australian Context

    Hallucination is the main reason Australian professional services (legal, financial, allied health) cannot let unsupervised AI go straight to the client. Edison AI mitigations: RAG grounding, file-note review patterns, output evaluation training for staff.

    Related Terms

  • Human-in-the-loop

    A design pattern where AI drafts or proposes and a human reviews or approves before an action is taken. The standard safety pattern for any AI workflow with external or financial consequences.

    Australian Context

    Every Edison AI implementation for an Australian client includes explicit human-in-the-loop checkpoints. Draft client comms reviewed before send, invoice follow-ups approved before chase, file notes verified before save.

    Related Terms

I

2 terms

  • Inference

    The process of running a trained model to produce an output. Generating a response, classifying a document, scoring a lead. Distinct from training. Inference is what gets billed per token in API pricing.

    Australian Context

    Inference cost is what Australian operators actually feel. The monthly Claude or OpenAI bill. Edison AI rights-sizes model choice per workflow (cheap, fast model for triage; premium model for high-stakes drafting) so inference cost stays predictable.

    Related Terms

  • Implementation Sprint

    A time-boxed delivery cycle (typically 3 to 6 weeks) that takes one workflow from audit to in-production. The unit Edison AI uses to ship the first agent, automation, or dashboard.

    Australian Context

    Australian SMBs benefit from sprint-based delivery because it forces a single workflow choice, removes scope creep, and produces a paying-back system before the next sprint starts. Three sprints typically take a business from pilot to embedded.

    Related Terms

J

1 term

  • Jailbreak

    A prompt or input crafted to make an AI model bypass its safety guardrails. Produce restricted content, reveal system prompt, or execute disallowed actions.

    Australian Context

    For Australian SMBs the practical jailbreak risk isn't a hacker. It's a staff member experimenting and accidentally pushing client data into an unapproved tool. Edison AI's responsible-AI training covers the staff behaviours that matter, not abstract red-team scenarios.

    Related Terms

K

1 term

  • Knowledge Graph

    A structured representation of entities (people, products, clients) and the relationships between them. Used to ground AI outputs in factual structure rather than text similarity alone.

    Australian Context

    Knowledge graphs sit one layer up from RAG and matter for Australian businesses with complex client structures. Wealth management, multi-entity ownership, advisory groups. Edison AI builds knowledge-graph backed RAG for clients where relationships matter.

    Related Terms

L

2 terms

  • LLM (Large Language Model)

    A neural network trained on a vast corpus of text to predict the next token given previous tokens. Modern LLMs (GPT-4, Claude 4, Gemini 2) can reason, code, summarise, draft, and follow multi-step instructions.

    Australian Context

    Every Edison AI build sits on top of one or more LLMs. Claude for drafting and analysis, GPT for general tasks, smaller models for routing and triage. Choice depends on cost, latency, context length, and Australian data-residency requirements.

    Related Terms

  • Latency

    The time between sending a prompt and receiving a response. Driven by model size, prompt length, output length, and provider load. Critical for voice agents and live customer interactions.

    Australian Context

    For Australian voice AI receptionists, end-to-end latency under 800ms is the threshold for the conversation to feel natural. Edison AI's voice stacks pair streaming LLMs with low-latency speech-to-text and text-to-speech to hit that bar.

    Related Terms

M

2 terms

  • Multi-agent System

    A coordinated set of AI agents where each has a specialised role and they communicate to complete a complex task. Typical pattern: planner agent decomposes, worker agents execute, reviewer agent validates.

    Australian Context

    Multi-agent systems make sense for Australian SMBs only when one workflow genuinely needs more than one specialised step. Shopify merchandising, RFP response, complex underwriting. Edison AI builds single-agent systems first and only adds agents when warranted.

    Related Terms

  • MCP (Model Context Protocol)

    An open protocol from Anthropic that standardises how AI applications connect to data sources and tools. Lets an agent talk to your CRM, file storage, or database without bespoke per-system integration.

    Australian Context

    MCP matters for Australian SMBs because it lowers the integration cost for new tools. Edison AI's recent agentic implementations increasingly use MCP servers for Notion, Salesforce, SharePoint, and Xero where official connectors exist.

    Related Terms

N

2 terms

  • NLP (Natural Language Processing)

    The broader field of getting computers to understand, generate, and act on human language. Modern NLP is dominated by LLMs but also includes classical tasks. Named-entity recognition, sentiment, classification.

    Australian Context

    Edison AI uses classical NLP under the hood of agentic systems for cheap, fast tasks (routing, classification, extraction) where invoking a full LLM would be wasteful. The right tool for the smallest job.

    Related Terms

  • Notion AI

    AI features built into the Notion workspace platform. Q&A across your workspace, writing assistance, summary generation, project triage.

    Australian Context

    For Australian SMBs already running Notion as their wiki, Notion AI is often the fastest path to AI-searchable SOPs without a custom build. Edison AI configures Notion AI as the entry-level capability layer for many Australian clients.

    Related Terms

O

3 terms

  • OpenAI

    The AI lab behind GPT, ChatGPT, DALL-E, and the Whisper speech model. Operates a developer API (priced per token) and consumer products (ChatGPT Free, Plus, Team, Enterprise).

    Australian Context

    OpenAI's API is one of two default model providers Edison AI deploys for Australian clients (the other is Anthropic). API tier matters. Paid API and Enterprise have no training on customer data; free ChatGPT does.

    Related Terms

  • Output Evaluation

    The discipline of systematically judging AI output quality. Accuracy, tone, completeness, safety. Across a set of representative tasks. The foundation for trusting AI in production.

    Australian Context

    Edison AI's AI Workshops for Teams train Australian staff to evaluate AI output critically, when to trust, when to verify, when to escalate. The single most under-invested skill in Australian SMB AI rollouts.

    Related Terms

  • OAIC (Office of the Australian Information Commissioner)

    The Australian regulator for privacy and freedom of information. Publishes guidance on AI and privacy, including specific guidance on the Privacy Act's application to generative AI training and use.

    Australian Context

    Edison AI rollouts respect OAIC's Australian Privacy Principles. Particularly APP 6 (use and disclosure), APP 11 (security), and the OAIC's generative AI guidance on transparency, consent, and data minimisation.

    Related Terms

P

3 terms

  • Prompt

    The input given to a language model. Instructions, context, examples, and the actual task. The prompt is the primary interface for controlling what a model does.

    Australian Context

    For Australian SMBs the difference between mediocre AI output and excellent AI output is almost always the prompt. Edison AI ships every implementation with a vetted prompt library, role-tagged and tone-locked to Australian English.

    Related Terms

  • Prompt Engineering

    The craft of writing, testing, and iterating prompts to get reliable, high-quality output from a language model. Combines clear instruction, examples, role-setting, and output-format specification.

    Australian Context

    Edison AI treats prompt engineering as a team capability, not a specialist role. Workshops train Australian staff to write, test, and improve prompts for their own workflows. The skill that compounds across every AI tool the business adopts.

    Related Terms

  • Prompt Library

    A curated, version-controlled collection of vetted prompts for a team. Role-specific, task-specific, brand-locked. The institutional IP that turns personal AI know-how into team capability.

    Australian Context

    Most Australian SMBs leak the value of their AI investment because each staff member is reinventing prompts in private. Edison AI builds shared prompt libraries (in Notion, SharePoint, or a dedicated tool) as part of every team training engagement.

    Related Terms

Q

1 term

  • Quantisation

    Reducing the numerical precision of a model's weights (e.g. from 16-bit to 4-bit) to shrink memory footprint and speed up inference, with modest accuracy trade-off. Enables LLMs to run on commodity hardware.

    Australian Context

    Relevant when an Australian client wants on-premise inference for data-residency or cost reasons. Quantised open-weight models (Llama, Mistral) can run on a single GPU and keep data inside the organisation. Edison AI selects quantisation only when the trade-off is justified.

    Related Terms

R

2 terms

  • RAG (Retrieval-Augmented Generation)

    A pattern where the model retrieves relevant documents from your data store, then generates an answer grounded in those documents. The default way to make an LLM answer from your knowledge, not its training set.

    Australian Context

    RAG is the foundation of every Edison AI build that involves company-specific knowledge. SOPs, client files, product catalogues, historical comms. Keeps data inside the organisation and dramatically cuts hallucination on factual questions.

    Related Terms

  • ROI (Return on Investment, AI)

    The measurable financial return from an AI implementation, calculated as (annual benefit − annual cost) ÷ annual cost. For SMBs the dominant benefit is hours of staff time recovered.

    Australian Context

    Edison AI's published baselines for Australian SMBs: a well-scoped first agent typically returns 4-10x in year one through hours saved. Use the ROI Calculator at /resources/roi-calculator to model your own.

    Related Terms

S

2 terms

  • System Prompt

    The instruction layer set by the application (not the end user) that defines the model's persona, scope, format rules, and safety boundaries for every conversation.

    Australian Context

    Edison AI agents ship with locked system prompts that enforce Australian English, your brand voice, data-handling rules, and escalation triggers. Staff prompts run inside that envelope. They cannot override the boundaries.

    Related Terms

  • Streaming

    Returning model output token-by-token as it is generated, rather than waiting for the full response. Critical for perceived latency in chat and voice applications.

    Australian Context

    Edison AI voice receptionists and live-chat agents stream by default. The user hears or sees the start of the response while the rest is still generating. The difference between feeling 'AI is thinking' and feeling 'AI is talking to me'.

    Related Terms

T

3 terms

  • Token

    The unit of text an LLM reads and writes. Roughly three-quarters of an English word. Token count drives API cost, prompt length limits, and inference speed.

    Australian Context

    For Australian operators the token unit matters mostly for budgeting. Edison AI quotes per-workflow inference cost up front so finance teams know what they're approving. Typically $0.10 to $2.00 per agent action depending on model and length.

    Related Terms

  • Temperature

    A model parameter that controls output randomness. Low temperature (0-0.3) for deterministic factual tasks; higher temperature (0.7+) for creative or varied drafting.

    Australian Context

    Edison AI's compliance-sensitive Australian workflows (invoice chases, file notes, client comms) run at low temperature so two staff requesting the same draft get materially the same output. Marketing and creative workflows run hotter.

    Related Terms

  • Training Data

    The corpus of text, code, and other content used to train an AI model. The composition and recency of training data determine what a model knows and where its biases sit.

    Australian Context

    Training-data composition matters for Australian SMBs because off-the-shelf LLMs over-index on US content. Edison AI grounds models in client-specific Australian context via RAG and prompt scaffolds rather than retraining.

    Related Terms

U

1 term

  • Use Case

    A specific, named workflow where AI is being applied. Invoice Follow-Up Agent, Lead Intake Agent, Voice AI Receptionist. The unit of value Edison AI ships to clients.

    Australian Context

    Australian SMBs win or lose on the choice of first use case. Edison AI's published gallery at /use-cases catalogues 46 deployable use cases across sales, finance, ops, support, marketing and operations, every one shipped or shippable for a real Australian client.

    Related Terms

V

2 terms

  • Vector Database

    A database optimised for storing and searching high-dimensional vector embeddings. The storage layer underneath RAG. Examples: Pinecone, Weaviate, Qdrant, pgvector (Postgres extension).

    Australian Context

    Edison AI's default for Australian SMB RAG builds is pgvector on a managed Postgres (Supabase or RDS). Keeps the vector layer inside the existing database, simplifies backups and compliance, avoids a new vendor relationship.

    Related Terms

  • Voice AI

    An AI system that takes voice input and produces voice output. Speech-to-text, LLM reasoning, text-to-speech, with sub-second latency. Used for AI receptionists, outbound calling, and voice agents.

    Australian Context

    Voice AI Receptionists are the single highest-leverage AI build for many Australian SMBs. They cover after-hours, qualify enquiries, and book consults with no human time. Edison AI's voice stack uses Australian-English voices and integrates with HubSpot, GoHighLevel, and most Australian PMS systems.

    Related Terms

From definitions to deployment

Know the vocabulary? Time to ship the first agent.

One 20-minute call is enough to identify the workflow with the biggest weekly drag and scope the first agent. ACSC-aligned by default. No vendor pitch.

Run the ROI calculator