Reference
AI automation glossary
101 terms defined in plain English, for the moment a vendor, a developer or a documentation page uses a word you were not going to admit you did not know. Every entry explains what it is, how it works, what it costs and what to watch for - with a link to the primary source and the month we checked it.
How to use this glossary
Most of the confusion in AI automation is vocabulary rather than difficulty. A vendor says their product does tool calling and you cannot tell whether that is a real capability or a rebranded webhook. A platform quotes a price per operation and a competitor quotes per execution, and the two numbers are not comparable. Someone asks whether your automation has guardrails and you would like to answer with something more specific than yes.
These pages are written for that. Each one gives a one-sentence answer at the top, then the detail underneath: how the thing actually works, what it costs, where it goes wrong, and what a business should do about it. They are longer than a dictionary entry on purpose, because the short version is the part that leaves you unable to ask a follow-up question.
Use it in reverse before a sales call. Read the vendor's page, list every term you could not define crisply, and look those up first - they are exactly where scope confusion and overselling hide. If you would rather someone else did the translating, the free AI automation audit maps your workflows and says plainly which of them are worth automating, and our automation services cover the build and the handover.
Anything that moves - prices, throughput limits, plan mechanics, regulation - is dated on the page it appears on, and every entry links to the specification, vendor documentation or regulator it came from so you can check it yourself. Where a figure is genuinely unknowable from someone else's data, the page says so and shows you how to measure your own instead.
What to learn first, if you are starting from zero
An alphabetical list is the right shape for looking something up and the wrong shape for learning. If you are new to this, read these in order. Each one is hard to understand until the one above it makes sense, and together they cover about eighty percent of what appears on a vendor call.
- 1. Workflow automation, then AI agent. Start with the older idea. Workflow automation is a fixed sequence someone wrote down; an agent decides the sequence at runtime. Almost every argument about what AI can do for a business is really an argument about which of these two you need, and the boring one wins more often than the industry admits.
- 2. Tool calling. This is the mechanism that turns a model that writes text into something that changes the state of your business. Once you understand that the model is only choosing which function to run and with what arguments, most of the mystique disappears and the right questions get much easier to ask.
- 3. Token and context window. The unit your model bill is denominated in, and the limit on how much the model can consider at once. These two explain both why the invoice moved and why the agent forgot something you told it earlier.
- 4. RAG. How an agent uses your documents. Read embedding, chunking and vector database straight after, as one sitting. They only make sense together.
- 5. Hallucination, guardrails and human in the loop. What goes wrong, what you do about it, and where a person stays in the chain. This is the trio that decides whether an automation is safe to point at a customer.
- 6. Execution versus task versus operation. The commercial one. Learn this before you compare two prices, because until you have, you are not comparing two prices. The AI agent platform directory has a full billing-unit table built on it.
- 7. ROI and payback period. Last, not first, because they need the six above to be filled in honestly. Once you can estimate the running cost, these two tell you whether to proceed.
Read as prose rather than as definitions, that sequence is roughly the learn hub, starting with what an AI agent is. If you would rather learn it with something being built alongside, the courses cover the same ground in order, and the build guides assume you already have it.
Agent concepts
What an agent is, how it decides what to do next, how it reaches your systems, and the controls that stop it doing something expensive.
This is the group where the marketing pressure is highest, so it is the group worth reading properly. The terms here divide into three jobs. Some describe how the model thinks: context window, token, temperature, system prompt, ReAct, planner-executor. Some describe how it reaches the outside world: tool calling, function calling, MCP, A2A. And some describe how you stop it being wrong in a way that costs money: guardrails, evals, observability, human in the loop, prompt injection. A vendor who is fluent on the first group and vague on the third is selling a demo.
The retrieval terms - RAG, embedding, chunking, reranking, vector database - belong together and are almost always encountered as a set. They are the answer to one question: how does an agent use documents it was never trained on. If a vendor tells you their agent knows your business, this is the machinery they are describing, and knowing the vocabulary lets you ask which part of it they actually built.
37 terms
A2A
Also: Agent2Agent, Agent to Agent protocol
A2A, the Agent2Agent protocol, is an open standard that lets independently built AI agents discover each other's capabilities and delegate tasks ac...
Agentic workflow
Also: agentic automation, agent workflow
An agentic workflow is an automation where a model decides some of the steps at runtime, sitting between a fully scripted workflow and a fully auto...
AI agent
Also: autonomous agent, LLM agent
An AI agent is a language model wired to tools and a goal, so it can decide which actions to take, take them, read the results, and keep going unti...
Chunking
Also: text splitting, document chunking
Chunking is splitting documents into passages small enough to embed and retrieve individually, while keeping each passage self-contained enough to...
Claude Haiku 4.5
Also: claude-haiku-4-5, claude haiku 4.5 pricing
Claude Haiku 4.5 is Anthropic's fast, cheapest tier, with a 200,000-token context window and a published API price of $1.00 per million input token...
Claude Opus 5
Also: claude-opus-5, claude opus 5 pricing
Claude Opus 5 is Anthropic's frontier model, with a 1,000,000-token context window and a published API price of $5.00 per million input tokens and...
Claude Sonnet 5
Also: claude-sonnet-5, claude sonnet 5 pricing
Claude Sonnet 5 is Anthropic's mid tier, with a 1,000,000-token context window and a standard API price of $3.00 per million input tokens and $15.0...
Context window
Also: context length, context size
The context window is the maximum amount of text, measured in tokens, that a model can consider in a single request - the system prompt, the conver...
DeepSeek V4 Flash
Also: deepseek-v4-flash, deepseek v4 flash pricing
DeepSeek V4 Flash is the cheaper of DeepSeek's two published model IDs, sharing the 1M-token context and 384K output ceiling of the Pro model at $0...
DeepSeek V4 Pro
Also: deepseek-v4-pro, deepseek v4 pro pricing
DeepSeek V4 Pro is DeepSeek's flagship API model ID, with a 1M-token context window and time-of-day pricing that halves off-peak: $0.66 to $1.32 pe...
Distillation
Also: knowledge distillation, model distillation
Distillation trains a small model to reproduce the outputs of a larger one on a specific task, so you get most of the quality at a fraction of the...
Embedding
Also: vector embedding, text embedding
An embedding is a list of numbers representing a piece of text's meaning, arranged so that texts with similar meanings sit close together in the sa...
Eval
Also: evaluation, AI evals
An eval is a repeatable test that runs your AI system against a fixed set of example inputs and scores the outputs, so you can tell whether a chang...
Fine-tuning
Also: model fine tuning, supervised fine-tuning
Fine-tuning is further training of an existing model on your own examples so it reliably adopts a format, tone or narrow skill - it teaches behavio...
Function calling
Also: structured function output, OpenAI function calling
Function calling is the original provider name for tool calling: the model returns a structured JSON request naming one of your declared functions...
Gemini 2.5 Flash-Lite
Also: gemini-2.5-flash-lite, gemini flash lite pricing
Gemini 2.5 Flash-Lite is the cheapest model ID on Google's Gemini API pricing page at $0.10 per million input tokens and $0.40 per million output t...
Gemini 2.5 Pro
Also: gemini-2.5-pro, gemini 2.5 pro pricing
Gemini 2.5 Pro is Google's Pro-tier model ID with tiered pricing that changes with prompt length: $1.25 per million input tokens and $10.00 output...
Gemini 3.7 Flash
Also: gemini-3.7-flash, gemini 3.7 flash pricing
Gemini 3.7 Flash is Google's current Flash-tier model ID, priced at $0.75 per million input tokens and $3.75 per million output tokens under promot...
GPT-5.6 Luna
Also: gpt-5.6-luna, gpt 5.6 luna
GPT-5.6 Luna is OpenAI's cost-optimised model ID in the GPT-5.6 family, at $0.20 per million input tokens and $1.20 per million output tokens, with...
GPT-5.6 Sol
Also: gpt-5.6-sol, gpt 5.6 sol
GPT-5.6 Sol is OpenAI's frontier model for complex professional work, with a 1,050,000-token context window, 128,000 max output tokens and a publis...
GPT-5.6 Terra
Also: gpt-5.6-terra, gpt 5.6 terra
GPT-5.6 Terra is OpenAI's balanced tier in the GPT-5.6 family, priced at $2.00 per million input tokens and $12.00 per million output tokens with t...
Guardrails
Also: AI guardrails, agent safety controls
Guardrails are the automatic rules around an AI system - input filters, output validation, permission limits, spend caps - that constrain what it c...
Hallucination
Also: confabulation, AI making things up
A hallucination is output that is fluent, confident and wrong - a language model producing plausible detail that has no basis in its input or in fact.
Human in the loop
Also: HITL, human approval
Human in the loop means an automation pauses at a defined point and waits for a person to approve, edit or reject an action before it goes ahead.
MCP
Also: Model Context Protocol, MCP server
MCP, the Model Context Protocol, is an open standard for connecting AI applications to tools and data sources, so one integration can be reused by...
Multi-agent system
Also: multi-agent, agent orchestration
A multi-agent system splits work across several specialised agents that hand tasks to each other, usually coordinated by an orchestrator that decid...
Observability
Also: LLM observability, agent tracing
Observability for AI systems means recording every prompt, tool call, result and cost for each run, so you can reconstruct exactly what happened wh...
Planner-executor
Also: plan and execute, planner executor agent
Planner-executor is an agent architecture that splits the work in two: one model call produces an explicit plan, then a cheaper executor carries ou...
Prompt injection
Also: indirect prompt injection, LLM01
Prompt injection is an attack where instructions hidden in content the model reads - an email, a web page, a document - override what you told it t...
RAG
Also: retrieval-augmented generation, retrieval augmented generation
RAG, or retrieval-augmented generation, is the pattern of searching your own documents for the passages relevant to a question and putting them in...
ReAct
Also: reason and act, ReAct prompting
ReAct is an agent pattern in which the model alternates between writing a reasoning step and taking an action, feeding each observation back into t...
Reranking
Also: reranker, cross-encoder reranking
Reranking is a second retrieval pass that takes the candidates returned by fast search and reorders them with a slower, more accurate model that re...
System prompt
Also: system message, agent instructions
A system prompt is the standing instruction block sent at the top of every request, defining an AI system's role, rules, tone and output format bef...
Temperature
Also: sampling temperature
Temperature is the sampling setting that controls how much randomness a model uses when picking each next token - low values make output predictabl...
Token
Also: tokens, LLM tokens
A token is the unit a language model reads and writes - roughly a short word or word fragment - and it is the unit every model provider bills you in.
Tool calling
Also: tool use, agent tools
Tool calling is the mechanism that lets a language model ask your code to run a specific function with specific arguments, so the model can look th...
Vector database
Also: vector store, vector search engine
A vector database stores embeddings alongside their text and metadata and finds the nearest ones to a query vector quickly, which is what makes sem...
Automation mechanics
How automation platforms actually work: triggers, integrations, the billing units they charge in, and what keeps a workflow reliable when a step fails.
Nothing here is glamorous and all of it decides whether an automation survives its first bad week. Two clusters matter most commercially. The first is metering: execution versus task versus operation is the term that makes two quotes comparable, and without it a $9 plan and a $50 plan tell you nothing. The second is reliability: error handling, retry and backoff, idempotency, queues and rate limits are the difference between a workflow that recovers from an outage and one that quietly sends the same invoice four times.
The connection terms - API, REST API, GraphQL, webhook, polling versus webhook, OAuth, API key, JSON, data mapping - are what an integration is actually made of. You do not need to write any of it, but knowing whether a platform reaches your system by webhook or by polling every fifteen minutes tells you how fast your automation can possibly be, and that is a question worth asking before you buy rather than after.
38 terms
Agentforce Flex Credit
Also: flex credits salesforce, agentforce credits pricing
A Flex Credit is Salesforce's consumption unit for Agentforce, sold at $500 per 100,000 credits, with one Agentforce action consuming 20 Flex Credi...
API
Also: application programming interface, API integration
An API is the defined set of requests one piece of software will accept from another, and it is what makes automation possible - if a system has no...
API key
Also: API token, secret key
An API key is a secret string that identifies and authorises an application when it calls a service - simple to use, and dangerous because it usual...
Bardeen credit
Also: bardeen credits, bardeen pricing credits
A Bardeen credit is charged per row of output rather than per step - one credit for each row a scraper, web search or AI tool produces, three credi...
Copilot Credit
Also: copilot studio credits, copilot studio messages
A Copilot Credit is the common billing currency across Microsoft Copilot Studio, which replaced the old message unit on 1 September 2025, and it is...
Cron trigger
Also: schedule trigger, cron job
A cron trigger starts a workflow on a fixed schedule - every ten minutes, every weekday at 8am - using the cron expression syntax that has describe...
Data mapping
Also: field mapping, data transformation
Data mapping is connecting the output fields of one step to the input fields of the next, converting formats along the way - the unglamorous work t...
Error handling
Also: failure handling, error workflow
Error handling is everything a workflow does when a step fails - retrying, routing to a human, logging, alerting - and it is usually more of the bu...
Execution vs task vs operation
Also: automation billing units, credits vs tasks vs operations
Executions, tasks, operations and credits are the different units automation platforms bill in - and because one workflow run can be one execution...
Fin resolution
Also: fin outcome, intercom fin resolution
Intercom's Fin bills per outcome - the unit formerly and still widely called a resolution - at $0.99, charged once per conversation when the custom...
GraphQL
Also: GraphQL API
GraphQL is an API query language where the client specifies exactly which fields it wants in a single request, instead of accepting whatever a fixe...
Gumloop credit
Also: gumloop credits, gumloop pricing credits
A Gumloop credit is a fixed slice of dollar cost - Gumloop documents one credit as $0.005 - and a workflow run costs one base credit plus whatever...
Idempotency
Also: idempotent, idempotency key
An operation is idempotent when doing it twice has the same effect as doing it once - which is what stops a retried request from creating a duplica...
iPaaS
Also: integration platform as a service, integration platform
iPaaS - integration platform as a service - is hosted software that connects your applications and moves data between them, providing pre-built con...
JSON
Also: JavaScript Object Notation, JSON payload
JSON is the text format almost every API and automation platform uses to exchange structured data, built from objects with named fields, arrays, st...
Latenode CPU second
Also: latenode credits, latenode execution credits
Latenode does not bill in execution credits at all - it bills in CPU seconds, charging for the processing time a workflow actually uses while it ru...
Make operation
Also: make credits, make.com operations
A Make operation is one module run against one bundle of data, and since Make renamed its billing unit to credits at a baseline of one operation to...
n8n execution
Also: n8n executions, what is an execution in n8n
An n8n execution is one full run of a workflow from trigger to finish, counted once no matter how many nodes it contains or how much data it moves...
No-code
Also: low-code, no code automation
No-code means building software by configuring a visual interface rather than writing source code, with low-code describing the same thing plus sma...
OAuth
Also: OAuth 2.0, OAuth2
OAuth is the standard that lets you grant an application limited access to your account in another service without giving it your password, using s...
Polling vs webhook
Also: polling trigger, push vs pull
Polling means your automation asks a system on a schedule whether anything changed; a webhook means the system tells you the moment it does. Webhoo...
Power Automate licensing units
Also: power automate premium, power automate process
Power Automate is not metered in runs or credits - it is licensed, either per user with Power Automate Premium or per unattended process with Power...
Queue
Also: message queue, job queue
A queue holds work items until something is ready to process them, decoupling the thing that creates work from the thing that does it so bursts are...
Rate limit
Also: API rate limiting, throttling
A rate limit is the cap a service puts on how many requests you may make in a period, enforced by rejecting excess calls - usually with HTTP status...
REST API
Also: RESTful API, REST
A REST API is the most common style of web API, addressing resources by URL and acting on them with HTTP methods - GET to read, POST to create, PUT...
Retry and backoff
Also: exponential backoff, retry policy
Retry and backoff is the practice of re-attempting a failed request after a wait that grows with each attempt, with randomness added, so transient...
Self-hosting
Also: self-hosted, on-premise automation
Self-hosting means running automation software on infrastructure you control rather than using the vendor's cloud, so your data and credentials sta...
ServiceNow assist
Also: now assist consumption, servicenow assists
An assist is the unit ServiceNow meters Now Assist in - roughly one per skill invocation, with third-party LLM calls charging one assist per call a...
Stack AI run
Also: stack ai runs, stackai run
A Stack AI run is the unit Stack AI's plan allowances are counted in, with the free plan listing 500 runs a month and Enterprise on a custom run co...
UiPath Agent Unit
Also: agent units uipath, uipath platform units agents
A UiPath Agent Unit is the consumption unit for agentic automations under Flex Licensing, charged per LLM call at 2 for premium models, 1 for stand...
UiPath Robot Unit
Also: robot units uipath, uipath RU consumption
A Robot Unit is UiPath's consumption unit for running automations, charged per minute of serverless runtime at rates from 1 to 10 RU per minute dep...
Voiceflow credit
Also: voiceflow credits, voiceflow pricing credits
A Voiceflow credit is the unit Voiceflow bills agent usage in, and Voiceflow's own live rate configuration puts one interaction at 1 credit, one mi...
Webhook
Also: web hook, HTTP callback
A webhook is a URL you give another system so it can push you an HTTP request the moment something happens, instead of you repeatedly asking whethe...
Workato recipe
Also: workato recipes, workato pricing
A Workato recipe is one automated workflow - a trigger plus the actions it runs - and it is the unit Workato's commercial packaging is built around...
Workflow automation
Also: business process automation, process automation
Workflow automation is software that runs a defined sequence of steps across your systems whenever a trigger fires, so a repetitive process happens...
Zapier task
Also: zapier tasks, what is a task in zapier
A Zapier task is one successful action step, counted every time Zapier completes a unit of work for you - triggers, filters, paths and failed actio...
Zendesk resolution allowance
Also: zendesk automated resolutions, zendesk resolution tiers
A Zendesk resolution allowance is the pool of AI agent resolutions included in your plan, and since the resolution tiers introduced on 18 May 2026...
Zendesk verified resolution
Also: verified resolution zendesk, zendesk verified vs contained resolution
A Zendesk verified resolution is the one outcome tier that draws down your resolution allowance: Zendesk defines it as one where the AI agent succe...
Voice and messaging
The terms that appear the moment an automation starts phoning or texting customers, including the US compliance vocabulary.
Voice has its own physics. Speech to text, text to speech and latency budget describe the chain a spoken sentence travels through, and barge-in and DTMF describe the two things callers do that break naive implementations: talking over the agent, and pressing keys. If a demo never shows an interruption, it has not shown you the hard part.
The messaging half is compliance rather than engineering. A2P 10DLC registration, opt-in and STOP handling are legal requirements in the United States, not features, and they apply to you rather than to your vendor. Missed-call text-back is the single highest-return pattern in this group for most service businesses, and it needs all three of those terms understood before it goes live.
12 terms
A2P 10DLC
Also: 10DLC, A2P messaging registration
A2P 10DLC is the US carrier standard for sending application-to-person text messages from ordinary 10-digit numbers, requiring you to register your...
Barge-in
Also: interruption handling, talk-over
Barge-in is the ability for a caller to interrupt an automated voice while it is still speaking and be heard, instead of having to wait for the pro...
DTMF
Also: touch tone, keypad input
DTMF is the tone system behind touch-tone keypads, and in automated phone systems it is the reliable way to collect numbers and confirmations that...
IVR
Also: interactive voice response, phone menu
IVR - interactive voice response - is the automated phone menu that plays options and routes callers based on keypad or spoken input, the technolog...
Latency budget
Also: response time budget, voice latency
A latency budget is the total response time a system is allowed, divided into an allowance for each stage, so you can see which component to fix ra...
Missed-call text-back
Also: missed call automation, call back text
Missed-call text-back automatically sends a text to anyone whose call you did not answer, usually within seconds, so the enquiry is captured instea...
Opt-in
Also: consent, express written consent
Opt-in is the recorded permission a person gives before you may text, call or email them, and the standard of consent required depends on what you...
SMS vs RCS
Also: RCS messaging, rich communication services
SMS is the universal plain-text messaging standard that reaches every mobile phone; RCS is its richer successor, supporting verified sender brandin...
Speech to text
Also: STT, ASR
Speech to text, also called ASR, converts spoken audio into written text, and it is the first stage of any voice AI agent - everything downstream d...
STOP handling
Also: opt-out handling, unsubscribe handling
STOP handling is processing a recipient's opt-out reply so they stop receiving messages immediately and permanently, across every system you own th...
Text to speech
Also: TTS, speech synthesis
Text to speech, or TTS, converts written text into spoken audio, and in a voice AI agent it is the component that turns the model's reply into some...
Voice AI agent
Also: AI phone agent, AI receptionist
A voice AI agent answers or makes phone calls, understanding what the caller says and taking real actions like booking a job or updating a record...
Business and buying
The numbers an owner uses to decide whether an automation is worth having, plus the contract and data questions that come up before anything touches live data.
These are the terms that decide whether a project happens. ROI, payback period and total cost of ownership are the frame; cost per booked job, lead response time, speed to lead and no-show rate are the operational numbers that actually move, and they are the ones worth measuring for a fortnight before you automate anything so you have a baseline to argue from afterwards.
The rest of this group is the paperwork that appears once something is going to touch live data: DPA, data residency, zero data retention, SLA, vendor lock-in, proof of concept and handover. Handover is the one most often skipped and most expensive to skip. If you cannot say who owns the workflow, the credentials and the documentation at the end of a project, you have not bought an automation, you have rented one.
14 terms
Cost per booked job
Also: cost per booking, cost per acquisition
Cost per booked job is everything you spent to win a job divided by the number of jobs booked, and it is the metric that tells you whether an autom...
Data residency
Also: data localisation, data sovereignty
Data residency is the requirement that data be stored and processed in a specified country or region, and it constrains which AI and automation pro...
DPA
Also: data processing agreement, data processing addendum
A DPA is the contract required between a data controller and a processor that sets out what personal data may be processed, for what purpose, under...
Handover
Also: handoff, project handover
Handover is the point at which a built automation becomes yours to own and operate, and it is complete only when you hold the credentials, the docu...
Lead response time
Also: first response time, response SLA
Lead response time is the measured interval between an enquiry and your reply, tracked as a distribution across channels and hours rather than as a...
No-show rate
Also: missed appointment rate, DNA rate
No-show rate is the share of booked appointments where the customer does not turn up and does not cancel in time to rebook the slot, making it one...
Payback period
Also: break-even period, time to payback
Payback period is how long an investment takes to recover its own cost from the savings or revenue it produces, and for a small business it is usua...
Proof of concept
Also: POC, pilot
A proof of concept is a deliberately small build that answers one specific question about whether an automation will work, run against real data wi...
ROI
Also: return on investment, automation ROI
ROI is the net gain from an investment divided by its cost, and for an automation it means the value it produced minus what it cost to build and ru...
SLA
Also: service level agreement, uptime guarantee
An SLA is a written commitment about service performance - typically uptime, response time and support hours - together with what happens when the...
Speed to lead
Also: lead speed, first response speed
Speed to lead is how quickly a business makes its first meaningful contact after an enquiry arrives, and it is one of the few operational numbers t...
Total cost of ownership
Also: TCO, true cost
Total cost of ownership is everything a system costs across its life - build, subscriptions, usage, maintenance, training and eventual replacement...
Vendor lock-in
Also: platform lock-in, switching cost
Vendor lock-in is the cost and difficulty of moving away from a supplier once you depend on them, which grows quietly with every workflow you build...
Zero data retention
Also: ZDR, no data retention
Zero data retention is a provider commitment not to store your API inputs and outputs after processing them, which is distinct from - and stronger...
Concepts that get confused
Four pairs account for most of the misunderstanding in this field. In each case the two things are used interchangeably in marketing copy and are not interchangeable at all, and in each case the confusion has a price attached.
Agent vs workflow vs automation
Workflow automation runs a sequence a human wrote down: when this happens, do these five things, in this order, every time. An AI agent is given a goal and a set of tools and decides the order itself, which is why the same input can produce different paths on different runs. An agentic workflow is the middle ground most real systems land on: a fixed skeleton with a model making judgement calls at two or three specific points.
Commercially this is the most expensive distinction on the page, because non-determinism is a cost, not a feature. A fixed workflow can be tested once and trusted. An agent has to be evaluated continuously, needs observability to debug at all, and will occasionally do something reasonable that you did not want. Pay for that flexibility where the work genuinely varies, and refuse to pay for it where it does not. A large share of disappointing AI projects are agents doing a job a workflow would have done more cheaply and more reliably.
RAG vs fine-tuning
RAG retrieves relevant text at the moment of the question and puts it in front of the model, so the model has never seen your data before and does not need to. Fine-tuning changes the model's own weights by training it on examples, which shifts how it behaves rather than what it knows.
The rule of thumb worth remembering: RAG for knowledge, fine-tuning for behaviour. If the complaint is that the agent does not know your pricing, your policies or your product catalogue, that is retrieval, and retraining a model will not fix it while costing considerably more. If the complaint is that the agent knows the right answer but says it in the wrong voice, at the wrong length or in the wrong format, fine-tuning is the lever, though a better system prompt usually gets most of the way there for nothing. Distillation is a third thing again: making a smaller, cheaper model imitate a larger one, which is a cost play rather than a capability one.
MCP vs function calling vs plugins
Function calling is a model capability: the model emits a structured request naming a function and its arguments, and your code decides whether to run it. It is the underlying mechanism, and it is provider-specific in its details. MCP, the Model Context Protocol, is a layer above that: an open standard for how an AI application connects to external tools and data sources, so that a connector written once works across the clients that support it. Its own documentation describes it as a USB-C port for AI applications, which is a fair summary of the ambition. Plugins, in the older sense, were the vendor-specific version of the same idea, and they did not travel between products.
The distinction that matters when buying: function calling ties your integration to one model provider's conventions, while an MCP server is portable across the clients that speak it. That is a vendor lock-in question rather than a technical one. Worth knowing alongside it, A2A solves a different problem again: MCP connects one agent to its tools, A2A lets separate agents discover each other and delegate work across frameworks. They are complementary rather than competing, and a vendor who treats them as alternatives has not read either specification.
Orchestration vs choreography
Both describe how several moving parts coordinate, and they are opposites. In orchestration there is a conductor: one component holds the plan, calls each step and knows the overall state. That is what a planner-executor setup does, and what most visual automation platforms are. In choreography there is no conductor: each component reacts to events it cares about, usually through a queue or a webhook, and the overall behaviour is emergent. Most multi-agent systems are a mix.
The commercial consequence is debuggability. An orchestrated system has one place to look when something goes wrong and one place to enforce guardrails. A choreographed one scales better and fails in ways that are much harder to trace, because no single component knows what was supposed to happen. For a business automating a process rather than building a platform, orchestration is almost always the right answer, and the distinction is worth raising with anyone proposing the other.
Common questions
- Which terms should I learn first?
- Six of them, in this order: AI agent, workflow automation, tool calling, token, RAG and human in the loop. Together they let you read almost any vendor page and tell what is actually being sold. Everything else in this glossary is detail hanging off one of those six. If you would rather have it as a narrative than as definitions, the learn hub covers the same ground as prose, and the courses walk through it with something being built alongside.
- Is an AI agent the same thing as a chatbot?
- No, and the difference is what it can do rather than how it talks. A chatbot produces text. An agent produces actions: it calls tools, writes to your systems and changes the state of something. A chatbot that answers a question about your refund policy is a chatbot. One that looks up the order, checks the policy and issues the refund is an agent, and the second one needs guardrails, an audit trail and usually a human review point that the first one does not.
- Do I need to understand any of this to buy automation?
- Not all of it, but the billing vocabulary pays for itself immediately. Knowing the difference between an execution, a task and an operation is the difference between a quote you can compare and a quote you cannot. So does knowing what a webhook is, because it decides whether your automation reacts in seconds or in fifteen-minute polls. The rest you can safely look up when a vendor uses it and you want a follow-up question ready.
- How current are these definitions?
- Every entry carries the month it was checked and links to the specification, vendor documentation or regulator it came from, so you can verify it rather than trust it. The fast-moving parts are prices, plan mechanics, throughput limits and messaging regulation, and those are dated on the page they appear on. Where a figure genuinely cannot be known from someone else's data, the entry says so and shows you how to measure your own instead.
- What is the difference between the glossary and the learn hub?
- The glossary answers what a word means when you hit it mid-sentence. The learn hub answers how the ideas fit together and why they matter, at the length of an article rather than a definition. Use the glossary in the middle of a vendor call or a documentation page, and use learn when you have an hour and want the shape of the whole thing. The build guides are the third layer: they assume both and get on with wiring something up.
- Why do the same words mean different things on different vendor sites?
- Because most of them are unregulated marketing terms rather than defined ones. Agent, workflow, orchestration and credit are all used loosely, and a credit in particular means whatever the vendor decides it means: a row of output on one platform, a second of compute on another, an action weighted by complexity on a third. Where a term has a real specification behind it, such as MCP or A2A, the entry links to it. Where it does not, the entry says so plainly.
- Which of these terms actually affect the price I pay?
- Six do most of the work. Execution versus task versus operation decides how your volume is metered. Token decides your model bill. Polling versus webhook decides how many wasted runs you pay for. Rate limit and retry and backoff decide how much duplicated work a bad afternoon produces. Self-hosting decides whether you pay a subscription or a server. Total cost of ownership is the one that ties them together, and it is the number worth building before you sign anything.
- Can I use these definitions in my own documentation?
- Yes, with a link back. They are written to be quoted: each entry opens with a one-sentence answer that stands alone, and the sources are cited so whoever reads your document can check them too. If you need the same material shaped around your own systems rather than in general, the free automation audit produces exactly that as part of its write-up.
