Plain English

Every term, one sentence.

If a word in the course is doing more work than it should, it is defined here. No vendor language.

AI Nucleus: The four layers a store owns itself: clean data, documented workflows, AI intelligence loops and digital employees.
Clean owned data: Customer, vehicle and service records that are de-duplicated, defined the same way everywhere, and exportable by you today.
Operational-grade data: Data accurate enough to run a process on. Marketing-grade data is good enough to send an email, not to make a decision.
AI outward: AI pointed at customers: chat widgets, ad copy, follow-up text. Everyone can buy it, so it is a commodity.
AI inward: AI pointed at your own operation, running on data only you own. That is the part competitors cannot copy.
Intelligence loop: A cycle where the store captures a signal, acts on it, measures the result and feeds that result back in.
Retrieval: Making the AI look up your actual records before answering, so it quotes your data instead of inventing an answer.
Hallucination: A confident answer the AI made up because it had no real record to look at.
Digital employee: An AI agent that runs one documented workflow end to end, supervised by a person in your building.
Agent operator: The named person who builds, watches and corrects your digital employees, with protected hours to do it.
Gate: A checkpoint you must pass before spending: foundation before buying, pilot before deploying, results before scaling.
Data hygiene: The work of fixing your data: removing duplicates, standardizing formats, filling gaps. Without it, every AI tool you buy runs on garbage.
Data warehouse: A central store where your sales, service and finance data are copied and organized for reporting. It does not run your day-to-day, it answers questions.
Data lake: A holding tank for raw, unstructured data: call recordings, photos, notes. You keep it now and figure out what it means later.
Data lakehouse: A warehouse and a lake in one: structured records and raw files in the same place, so an AI agent can read both.
ETL (extract, transform, load): The pipeline that pulls data out of your DMS, cleans it, and writes it into your warehouse. It is the plumbing vendors charge for.
Data pipeline: The automated path data travels from one system to another. If it breaks, your dashboard quietly shows yesterday's numbers.
Structured vs unstructured data: Structured data fits in rows and columns: a repair order. Unstructured does not: the voice of the call about that repair.
Master data: The single, agreed version of a customer, a vehicle or a part across every system. If your CRM and DMS disagree, you have no master data.
Data governance: The rules for who owns which data, who may see it, and what good looks like. Without it, data quality is nobody's job.
Data lineage: A map of where each number came from and what changed it on the way. It is how you trust a dashboard enough to act on it.
Single source of truth: One system whose numbers win every argument. When sales and service report different gross counts, this is what settles it.
LLM (large language model): The engine behind tools like ChatGPT: software trained on huge amounts of text that predicts the next words, which is why it sounds confident even when it is wrong.
Model: One specific version of that engine (GPT, Claude, Gemini). Vendors swap models the way you swap engines, and the results change when they do.
Prompt: The instruction you give the AI. Vague instruction in, vague answer out.
System prompt: The standing instructions a vendor writes behind the scenes that set the AI's rules, tone and limits before your customer ever types.
Context window: How much text the AI can hold in mind at once. Past that limit it forgets the beginning of the conversation.
Token: A chunk of a word. It is the unit AI vendors bill you by, so long conversations literally cost more.
RAG (retrieval augmented generation): Making the AI look up your real records first and answer from them. It is the difference between an agent that quotes your inventory and one that invents it.
Vector database: The index that lets RAG find the right record by meaning rather than by exact wording.
Embedding: A numeric fingerprint of a piece of text that lets the system tell which records are about the same thing.
Grounding: Forcing every answer to point back to a real source in your data.
Fine-tuning: Retraining a model on your own examples. Expensive, and usually unnecessary if your retrieval is set up properly.
MCP (model context protocol): A shared standard that lets an AI agent plug into your systems (DMS, CRM, calendar) the way USB let any device plug into any computer.
Tool call / function call: When the AI stops talking and actually does something: books the appointment, writes the RO note, sends the text.
API: The doorway one system opens so another can read or write data. If your vendor has no API, your data is stuck.
Agent: An AI that plans and takes multiple steps toward a goal, rather than answering one question and stopping.
Orchestration: The layer that decides which agent or tool handles which step, and in what order.
Human in the loop: A person approves the action before it reaches the customer. Where every first agent should start.
Guardrails: The written limits on what an agent may say, promise or do, enforced by the system rather than by hope.
Prompt injection: An attack where text in an email, a web page or a note tricks your agent into ignoring its instructions.
PII: Personally identifiable information: names, phone numbers, SSNs, credit applications. Know exactly where your vendor sends it.
Inference: The act of the model producing an answer. This is what you are paying for per use.
Latency: How long the answer takes. Over a few seconds in a live chat and the customer leaves.
Multimodal: A model that also handles images, audio or video: damage photos, recorded service calls, walkaround videos.
Speech to text / text to speech: Turning phone calls into searchable text, and turning text into a voice that answers the phone.
Voice agent: An AI that answers or places calls. Judge it on booked appointments, not on how human it sounds.
Copilot: AI that helps a person work faster while the person stays responsible for the output.
Automation vs AI: Automation follows rules you wrote. AI makes a judgment call. Use automation wherever the rule is known, it is cheaper and it never invents.
Machine learning: The older, still useful branch that predicts numbers from history: which customer is likely to buy, which repair is likely to be declined.
Training data: What the model learned from. If a vendor trains on your data, ask in writing whether it benefits your competitors.
Model drift: Results quietly getting worse over time as the world, or the model, changes. This is why an agent needs a supervisor.
Benchmark: A vendor's test score. It tells you nothing about your store until you measure the same thing on your own data.
Token cost / cost per conversation: The real unit economics of an AI tool. Ask for it per conversation, not per month, before you sign.