The terms, defined plainly.
Short, precise definitions of the AI and software terms that come up in our work. Written so a non-engineer can follow them and an engineer will not wince.
Food cost variance
What the kitchen spent, minus what the menu sold should have cost.
Read →
Prime cost (restaurant)
Food cost plus labor cost, as one share of sales.
Read →
Silent AI (insurance)
Policy wording that neither covers nor excludes AI losses. The adjuster decides.
Read →
AEDT (automated employment decision tool)
A score, a rank or a knockout that drives the hiring call.
Read →
e-Manifest
EPA's electronic tracking system for every hazardous waste shipment.
Read →
Key Data Element (KDE)
The facts FSMA 204 makes you record at every tracking event.
Read →
NEC 2026 load calculation
Code math that decides whether an EV charger fits the panel you have.
Read →
EPA Section 608 compliance
Certified techs, no venting, leaks fixed on a 30-day clock, records kept three years.
Read →
Roofing supplement
The documented request that closes the gap between the adjuster's estimate and the real scope.
Read →
Forward deployed engineer (FDE)
An engineer who builds inside your business, on your own systems.
Read →

Reconciliation
Two records of the same money. Prove they agree.
Read →

Settlement vs authorisation
Approved is a promise. Settled is the money.
Read →

Model risk management
A model that's wrong with confidence is riskier than one that fails loudly.
Read →

AI evals
The regression test suite for a model that has no fixed right answer.
Read →

AVM (automated valuation model)
A price with no confidence range is a guess with a decimal point.
Read →

Computer software assurance (CSA)
Test hardest where failure reaches the patient. Write less elsewhere.
Read →

Dataset shift
The model did not change. The world did.
Read →

PHI (protected health information)
Health data plus anything that points at a person.
Read →

Statistical inventory reconciliation
A leak conclusion, drawn from weeks of ordinary tank data.
Read →

Wetstock management
Track every gallon from delivery to sale. Explain the gap.
Read →

Work product doctrine
Protection depends on who directed the work, not who filed it.
Read →

BAA (business associate agreement)
The contract that makes a vendor legally answerable for patient data.
Read →

Audit trail
A log you can edit isn't evidence. An audit trail is a log you can't.
Read →

HIPAA
A federal law your whole system has to satisfy, not a feature a vendor sells you.
Read →

Data residency
Your database region means little if a prompt leaves it for a model API abroad.
Read →

Intelligent document processing (IDP)
OCR reads the page. IDP checks the numbers actually add up.
Read →

Agent washing
Old automation, new label, same demo.
Read →

Zero data retention
The provider keeps nothing after answering, and it is an approval, not a default.
Read →

Webhook
The event knocks on your door. You stop phoning to ask if anything happened.
Read →

Voice agent
Software that answers the phone, understands the caller, and speaks back inside a budget of a few hundred milliseconds.
Read →

Vibe coding
You describe the app. The model writes it. You never read the code.
Read →

Turn detection
Deciding whose turn it is to speak, beyond whether the room goes quiet.
Read →

Text-to-Speech (TTS)
The voice half of a voice agent. Late audio still sounds like a robot, even with a natural voice.
Read →

Technical debt
Every shortcut is a loan. The interest is paid in delivery speed.
Read →

Shadow AI
Your staff already adopted AI. Nobody told procurement.
Read →

Semantic search
It finds meaning, not letters. That's also how it fails.
Read →

RPA (robotic process automation)
Software that clicks through screens the way a person did, until the screen changes.
Read →

Reranking
The nearest passage and the right one are not always the same passage.
Read →

Regression evals
Run the whole eval suite on every prompt change, not one case.
Read →

RBAC (role-based access control)
Permissions belong to a role, not a person. RBAC is why offboarding doesn't need a spreadsheet.
Read →

Prompt Injection
Instructions hidden in the data a model reads, not in the prompt you wrote.
Read →

Production-grade AI
It works on a Tuesday when nobody is watching.
Read →

LLM observability
Your uptime graph is green. The answers went bad on Tuesday.
Read →

Multi-agent system
Several AI agents on one job. Usually one agent was enough.
Read →

Model deprecation risk
The model your product depends on has a retirement date.
Read →

LLM tracing
It is the flight recorder for a wrong answer nobody can explain.
Read →

LLM benchmark
A public exam every model takes, and a poor predictor of your task.
Read →

LLM-as-judge
One model grades another's answers, at a scale no human review team can match.
Read →

Label leakage
The model scored 99 percent because it was shown the answer.
Read →

Inference
Training happens once. Inference happens on every request, forever, and that is where the bill lives.
Read →

Idempotency key
One key. The server runs your request once, however many times you send it.
Read →

Hybrid search
Keywords catch the SKU. Vectors catch the paraphrase. Use both.
Read →

Grounding
RAG fetches documents. Grounding is proving the answer actually used them.
Read →

Golden set
The examples a model change is not allowed to break.
Read →

Foundation model
A foundation model is the base you build on. You almost never train one.
Read →

Eval-driven development
Write the test cases first. Then find out if the prompt works.
Read →

Endpointing
The millisecond guess that decides whether a voice agent cuts you off or leaves an awkward pause.
Read →

Chunking
Cut a document badly and retrieval hands the model half a rule.
Read →

Chain-of-Thought
The model writes out its steps before the answer, so you can see how it got there.
Read →

Barge-in
Letting a caller talk over the agent and having it stop, mid-sentence, without losing the thread.
Read →

Automatic Speech Recognition
The word error rate a vendor demos on a quiet mic is not the one your phone line gets.
Read →

Agent orchestration
Deciding which agent runs, with what context, and who gets the result.
Read →

Agentic workflow
A fixed path a developer wrote, with AI doing one or more of the steps.
Read →

Agentic RAG
A plain RAG pipeline searches once. Agentic RAG searches until it has enough.
Read →

Agentic AI
Software that decides its own next step, then takes it.
Read →

AI agent
An agent picks the next step. A workflow was told it.
Read →

Context window
The model's working memory. Fill it and something gets dropped.
Read →

Embedding
Text turned into numbers, so a computer can measure meaning.
Read →

Fine-tuning
Training a model on your examples changes how it answers, not what it knows.
Read →

Guardrails
Checks around the model, not inside it.
Read →

Hallucination (AI)
The model is not lying. It is finishing the sentence.
Read →

Human in the loop
A person who can still say no, and does.
Read →

Model Context Protocol (MCP) / MCP Server
Write the connector once. Every AI client can use it.
Read →

Prompt engineering
Writing the instruction well. It matters less each year.
Read →

RAG (retrieval-augmented generation)
Search your documents first. Answer only from what came back.
Read →

Tool calling
The model asks. Your code decides whether to obey.
Read →

Vector database
Stores meaning as numbers. Finds the nearest, not the right one.
Read →

Agent handoff
Passing the task on, without losing what it knew.
Read →

Agent harness
Swap the model and keep the harness. Nothing should break.
Read →

Blue-green deployment
Two live environments. One switch. Rollback is flipping it back.
Read →

Context engineering
The prompt was never the hard part. Deciding what it can see is.
Read →

Feature flag
Code ships Friday. The switch flips Monday, for one user at a time.
Read →

Large language model (LLM)
A large language model predicts the next word. It does not look anything up.
Read →

SLA (service-level agreement)
The written promise about response time, resolution time, and what happens if we miss it.
Read →

System prompt
The instructions the user never sees, and the ones that break first.
Read →

Canary deployment
A slice of traffic tests the new version before everyone gets it.
Read →

Computer use
An agent that clicks and types is the fallback, not the goal.
Read →

Dead-letter queue
Where a message goes after it fails too many times to keep retrying.
Read →

Dead man's switch
The alert that fires when a scheduled job goes quiet, not when it errors.
Read →

Embedding drift
The vectors you stored last year no longer match the model you run today.
Read →

Exponential backoff
Hit the same wall fast, or wait a little longer each time and get through.
Read →

GraphRAG
Chunks answer what a passage says. A graph answers how things connect.
Read →

Knowledge cutoff
The date a model's world stopped. Nothing after it is in there.
Read →

Minimum viable product
The smallest product that tests the business, not the smallest demo that survives a meeting.
Read →

Model fallback
The backup model answers differently, so it needs its own test pass.
Read →

PII (personally identifiable information)
Anything that points back at one person, wherever it sits.
Read →

Proof of concept
Built to answer one question, then thrown away. Keeping it is how a demo becomes production by accident.
Read →

Rate limiting
The 429 you'll hit eventually. Design for it, don't wait for the outage.
Read →

Reasoning model
A reasoning model spends tokens thinking before it answers, and you pay for every one.
Read →

Red teaming
You attack your own AI system before a stranger does it for free.
Read →

Rollback
Code rolls back in seconds. A sent email or a run migration does not.
Read →

SLO (service-level objective)
The internal target that pages an engineer before the SLA is ever breached.
Read →

Structured outputs
The model answers in a shape your code can act on directly.
Read →

Token
The unit every AI bill is counted in, and it is not a word.
Read →

Agent memory
Not the context window. What you choose to save and reload.
Read →

Call escalation
The moment a voice agent hands the phone to a person.
Read →

CI/CD
Every change tested and shippable the moment it's merged.
Read →

Circuit breaker
Stop calling the dependency that's down before it takes you down too.
Read →

High Availability
99.9% sounds close to perfect. It is still 8 hours and 46 minutes of downtime a year.
Read →

Message queue
The buffer that lets a fast system talk to a slow one without waiting on it.
Read →

Prompt caching
Stop paying to re-read the part of the prompt that never changes.
Read →

Temperature
The dial that controls how much a model is allowed to guess.
Read →

Disaster Recovery
Backups are a component. A rehearsed plan to bring the system back up is disaster recovery.
Read →

Multi-tenancy
One system, many customers, and none of them can see each other's data.
Read →

OCR (optical character recognition)
Turns a picture of text into text a computer can search and edit.
Read →

ReAct pattern
The model thinks out loud, acts, reads the result, then thinks again.
Read →

RPO (recovery point objective)
How much data you can afford to lose, measured in time since the last good backup.
Read →

Recovery time objective (RTO)
The longest a system can stay down before the business decides it can't.
Read →

Shadow deployment
The new model answers every request. Nobody sees its answer yet.
Read →
