The words, one at a time

Lesson 2 of 4 in The Words People Use — Agent, Agentic, Copilot, Assistant.

Here is the field guide — not as law, but as the usage you are most likely to meet, with the part each word reliably hides. Read the third column hardest. Every one of these terms leaves something important unsaid, and knowing exactly what it leaves unsaid tells you which follow-up question to ask.

Three of these words commit to control flow (workflow, RPA, and usually chatbot) — and the RPA row shows how even a firm commitment gets eroded once model-driven decisions are bolted on. The rest are compatible with almost any architecture.

Ten words, what they usually mean, and what they never tell you
Word	What it usually means in practice	Who decides the next step	What the word does NOT tell you
Agent	An LLM with tools, run in a loop, pursuing a goal. Also, very often, any AI feature at all.	Usually the model — when the word is used precisely. When used loosely, anyone.	Whether the loop is real. Plenty of “agents” are two chained prompts with a fixed order.
Agentic AI	Contested. Sometimes “built as an agent,” sometimes “multi-agent,” sometimes “more autonomous than last year’s product,” sometimes just an adjective on a product line.	Unspecified by the term.	Almost everything — including, crucially, whether more than one agent is involved. Lesson 3 is entirely about this word.
Copilot	Historically: help offered alongside a human doing the work. Currently also a brand family spanning many products with very different autonomy.	Varies widely. Some copilots suggest; others act on your behalf.	Whether a human actually approves anything. The metaphor implies a second pilot; the product may fly alone.
Assistant	A conversational front door to capabilities. Often the umbrella term when a vendor is describing the user-facing surface rather than the machinery.	Usually the human, turn by turn — but assistants increasingly call tools between turns.	Whether it has tools with side effects. “Assistant” describes the interface, not the authority.
Chatbot	Turn-taking text interface. Historically intent-and-rule based; now usually LLM-backed.	The human in the conversation.	Whether it can act. A chatbot wired to a refund API is doing far more than chatting — and is often still called a chatbot.
Workflow	A sequence of steps fixed in code or drawn in a designer. Model calls may sit inside steps.	A human, at design time. The most honest word on this list.	How fuzzy the individual steps are. A frozen graph of LLM steps still needs evals per step.
RPA (robotic process automation)	Software that drives existing applications the way a person would — clicking screens, filling fields, moving files — to automate clerical work.	A human, at design time. The “robot” follows a recorded or scripted path.	Whether a model has been added on top. Many vendors now bolt model-driven decisions onto RPA runners, which changes the risk profile entirely.
Orchestration	The coordination layer: which component runs, in what order, with what state — across models, tools, or agents.	Whoever wrote the orchestrator. That may be your code (fixed) or a model acting as router (runtime).	Which of those two it is — the single most important thing about an orchestration layer, and the thing the word omits.
Autonomous	A position on the autonomy spectrum: how much the system does without a human confirming.	Implies the model, for the actions it covers.	Which actions, and what oversight exists. Autonomy is set per action; the adjective is applied per product.
AI employee / digital worker	A packaging and pricing metaphor: a bundle of capabilities sold as a role (“an AI analyst”) rather than as a feature, often priced per seat rather than per token.	Unspecified. The metaphor is about the org chart, not the control flow.	Anything technical at all. It is a commercial frame — which does not make it dishonest, but it answers no engineering question.

Four of these do most of the damage in real conversations.

Copilot is the one that fools careful people, because the metaphor is so good. A co-pilot sits beside a pilot; the implication is that a human remains in command. But the word is also a product-line brand, and brands span architectures. Some products carrying it suggest completions a human accepts one at a time; others read your data, call tools and take actions between your turns. The word is doing branding work and capability work simultaneously, and only one of those is load-bearing for you. So: what can this copilot do without me clicking anything?

Orchestration hides exactly the fact you need. Coordinating steps is what an orchestrator does — but if your code holds the order, you have a workflow; if a model decides which component runs next, you have an agent with a fancier name. Both ship under the same word, sometimes in the same product on different pages. So: is the orchestrator code, or is it a model?

Autonomous is an adjective applied to products for a property that lives on actions. You already know from the previous module that one agent typically runs several dial positions at once — reading freely, editing with an audit trail, executing the risky call only with approval. A product page has one adjective and no room for that nuance, so “autonomous” gets applied to the whole system on the strength of its most autonomous action. So: autonomous for which actions, and what happens when one of them is wrong?

AI employee and digital worker are commercial frames rather than technical claims, and it is worth being fair about why they exist: buyers find it much easier to reason about “a role we are filling” than “a token budget and a tool allowlist.” The frame communicates scope and price well. It carries no information about control flow, so treat it as the beginning of the conversation. So: which of this role’s tasks does the model decide, and which are fixed?

Myth: “Copilot” means a human approves every action

Not reliably. It was the original implication and it still holds for some products — inline code completion, where every suggestion is accepted or rejected by a person, is a genuine human-in-the-loop design.

But the word now also names product families whose members act between your turns: retrieving documents, calling tools, updating records. The honest read is that copilot tells you the product is positioned as assistance, not that a gate exists. Gates are a runtime fact — check for them directly.

Myth: “RPA is dead, agents replaced it”

No. Screen-driving automation is still deployed at scale, for a good reason: when a target system has no API, driving its interface is sometimes the only integration available, and for a stable, high-volume, fully specified process a fixed path is better than a model choosing — cheaper, faster, and testable by assertion.

What has changed is that model-driven decision points are increasingly added on top of those runners. That hybrid is where the interesting risk lives, and it is why “is this RPA or an agent?” is often the wrong question. The right one is which specific steps became model-decided.

Myth: “Autonomous means no human is involved”

It names a position on a spectrum, not a fact about your oversight. A system can act without per-action approval and still be sampled, audited, rate-limited, and stopped — that combination is normal and healthy.

The reverse trap matters more: a system with a human gate can drift into effective autonomy when reviewers approve reflexively. Neither the presence of the word nor the presence of a gate tells you what oversight actually happens. Ask what stops it, and who would notice.

Myth: “Assistant is the beginner tier, agent is the advanced tier”

Tempting, and wrong often enough to hurt. The two words describe different axes: assistant is mostly about the interface (you converse with it), while agent is about control flow (the model chooses steps). A product can be both, either, or neither.

Some of the most capable agentic systems in production present as a chat box, because chat is a good front door. Some products called agents are a single templated prompt. Interface tells you nothing about authority.

Myth: “If it uses an LLM to decide something, it is an agent”

The previous module’s test still holds: a model inside a step is a component; a model choosing the next step is an agent. A classifier that routes a ticket to one of five queues, after which a fixed pipeline runs, is a workflow with a smart branch.

This matters because the boundary is where testing strategy changes. A fixed graph with fuzzy nodes gets per-node evals plus end-to-end assertions. A model-chosen path gets pass rates and trace review, because there is no fixed sequence to assert against.

Tool: Is It an Agent? — Practise the control-flow test on real systems — including several whose marketing name points the wrong way — in the Is It an Agent? simulator.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.