October 17, 2026· Filippo Pietrantonio

AI Strategy

AI Automation Glossary: 25 Terms Leaders Keep Getting Wrong [2026]

"Agent," "copilot," "autonomous," "eval" — the words in your AI proposals are doing a lot of work, and most of it is vendor work. Here are the 25 terms that actually change what you buy, build, and sign.

The short answer

Most AI vocabulary confusion is expensive, not academic. "Agent" usually means a scripted workflow — Gartner estimates only about 130 of the thousands of self-described agentic vendors are genuine. "Pilot" means nothing until something runs in production, where MIT found roughly 95% of enterprise GenAI pilots deliver no measurable P&L impact. These 25 terms are the ones that change what you buy, what you build, and what you refuse to sign.

Nearly half of companies now generate meaningful value from AI, up from 5% a year earlier, according to BCG's September 2026 survey of 1,330 executives. The other half are still paying for things they can't define.

That's not a coincidence. AI procurement is unusually vulnerable to language drift, because the vendor writes the glossary. When "agent" can mean anything from a GPT wrapper to a system with real tool access and decision authority, a six-figure proposal becomes impossible to compare against an alternative — or against doing nothing.

If you're a COO or VP of Ops signing off on AI spend, you don't need to learn machine learning. You need about 25 words well enough that nobody can hide behind them. Here they are, grouped by the decision each one affects.

What do the foundational terms actually mean?

These five get swapped for each other constantly, and the swap is usually what makes a quote look cheap or expensive. The distinction that matters is whether the system follows rules, generates content, or decides what to do next.

AI automation. Software that completes a task using a model's judgment, not just a fixed rule — reading an unstructured invoice, routing an ambiguous ticket, drafting a reply. The judgment is the point. If the work can be fully described as if-this-then-that, it isn't AI automation and shouldn't be priced like it.

Workflow automation. Deterministic task-chaining: triggers, conditions, and actions across systems (Zapier, Make, n8n). It's cheaper, more reliable, and more auditable than AI automation. A surprising share of "AI projects" should be this. Don't let anyone tell you that's a downgrade.

RPA (robotic process automation). Bots that mimic clicks and keystrokes in a UI, usually to bridge systems without APIs. Rule-bound and brittle — change the screen, break the bot — but genuinely excellent at high-volume, stable, structured work. It isn't obsolete; it's narrow. (Full comparison here.)

Intelligent automation. The umbrella term for combining the above: RPA for the mechanical steps, models for the judgment steps, orchestration to hold it together. Useful as a category, useless as a product claim. Ask which parts are which.

Generative AI. Models that produce new content — text, code, images, audio. It's a capability, not a solution. "We're adding GenAI" describes the ingredient, not the dish.

Which AI terms are mostly vendor marketing?

These five are where language drift costs real money. None of them are meaningless — they've just been stretched until they stopped constraining anything, which is exactly when you should start asking for specifics.

Agentic AI / AI agent. A system that pursues a goal across multiple steps, chooses its own tools, and adapts when a step fails. The test isn't autonomy in the demo — it's whether it recovers from something the builder didn't anticipate. Gartner expects agentic AI in 33% of enterprise software applications by 2028, up from under 1% in 2024, which means almost everything you're shown today is early. (Our working definition.)

Agent washing. Gartner's own term for rebranding chatbots, assistants, and RPA as "agentic" without the underlying capability. Their estimate: of thousands of agentic AI vendors, roughly 130 are real. That ratio should change how you read a deck.

Copilot. A model in a human's workflow that suggests, drafts, or explains — the human still acts. This is the highest-success-rate category in AI right now and the lowest-status one in vendor positioning. Ignore the status. 80% of McKinsey's respondents who use AI in their own role report improved individual productivity, and most of that is copilot work.

Autonomous. Should mean "acts without human approval on a defined decision set." Usually means "the demo had no human in it." Always ask: which specific decisions does it make alone, and what's the escalation path? By 2030, 42% of companies expect to give agents real decision-making authority — but only 5% have the full set of controls in place today (BCG, 2026).

Digital worker / AI employee. A pricing frame, not an architecture. It invites you to compare the tool to a salary instead of to the alternative software — which is almost always flattering to the tool. Price it against the next-best system, not against a person.

What do the build terms mean when you're reading a proposal?

You don't need to implement these. You need to know what they imply about cost, timeline, and lock-in — because the difference between two of them can be a 10x difference in a quote.

RAG (retrieval-augmented generation). The model looks up relevant content from your own documents or data at question time and answers from it. This is the right default for "AI that knows our stuff." Cheap to update, easy to audit, no retraining.

Fine-tuning. Adjusting a model's weights on your data to change its behavior or style. Expensive, slow to update, and rarely the answer to "it doesn't know our policies" — that's a retrieval problem. If a proposal leads with fine-tuning for a knowledge use case, ask why RAG was ruled out.

Context engineering. Deliberately deciding what information enters the model's context window on each call — which documents, which history, which tools, in what order. This has quietly replaced prompt engineering as the actual skill. a16z's framing of the shift is that the hard problem moved from how to build to what to build; context engineering is where that gets decided in practice.

MCP (Model Context Protocol). An open standard for connecting models to tools and data sources, so an integration written once works across assistants. For a buyer, MCP support is a portability signal — it reduces the cost of switching models or vendors later.

Orchestration. The layer that decides which step runs when, retries what failed, and hands off between models, tools, and humans. It's where most of the real engineering lives and where most proposals are vaguest. "We use LangGraph" is a tool choice, not an orchestration design.

Which terms decide whether a pilot survives production?

These five are the difference between a demo and a system. They're also the terms most often missing entirely from a proposal — which is itself the finding.

Eval. A repeatable test set that scores the system's output against known-correct answers, run on every change. Without evals, you have no way to know whether a prompt tweak improved things or quietly broke them. Ask to see the eval suite. If there isn't one, you're buying vibes.

Human-in-the-loop. A human approves or corrects specific outputs before they take effect. Not a weakness — the design choice that makes automation deployable in anything consequential. The right question is which decisions are gated, not whether any are.

Hallucination. The model producing fluent, confident, wrong output. It's a property of how these systems work, not a bug awaiting a patch. You manage it with retrieval, constrained outputs, and review gates; you don't eliminate it. Any vendor claiming they've solved it is telling you something useful about themselves.

Guardrails. Enforced constraints on what the system can do or say — input filters, output validation, tool permissions, spend limits. Deloitte's 2026 survey of 3,235 leaders found only 21% have a mature governance model for agentic AI, meaning roughly four in five are scaling agents faster than the controls around them.

Observability. Logging what the system actually did — inputs, tool calls, outputs, costs, failures — in a form a human can inspect after the fact. Without it, you cannot debug, audit, or defend the system. This is the single most common thing missing from an otherwise good pilot.

Which terms decide whether your CFO renews the contract?

Cost and compliance language is where AI projects die quietly, twelve months in. These five come up in the renewal conversation whether or not they came up in the sales conversation.

Token. The unit models read and write in — roughly ¾ of a word. You're billed per token, in and out. Why it matters: long documents, long chat histories, and chatty agents multiply cost invisibly, and a single agent step can consume thousands.

Inference cost. What it costs to run the model in production, per request. Per-token prices keep falling; bills keep rising, because capability growth drives volume faster than unit price drops. 20% of McKinsey's 2026 respondents say AI operating costs have constrained their use of the technology. KPMG found 42% of companies have only partial visibility into what they're spending on AI. (How to get ahead of this.)

Pilot vs production. A pilot proves a model can do something; production means it runs on real volume, with owners, monitoring, error handling, and a budget line. MIT's Project NANDA found roughly 95% of enterprise GenAI pilots deliver no measurable P&L impact. Gartner separately expects over 40% of agentic AI projects to be canceled by the end of 2027. The gap between those two words is where the money goes. (The handoff most companies skip.)

Shadow AI. Employees using unapproved AI tools with company data. It's a signal, not just a risk — it tells you exactly which workflows people find painful enough to route around you. BCG found more than 80% of AI spending now sits outside the enterprise IT budget, which is the governance problem in one number. (More here.)

High-risk AI system. An EU AI Act classification covering uses like recruitment, performance evaluation, and worker monitoring, which carry documentation, oversight, and conformity obligations. The compliance date moved: the Digital Omnibus deferred high-risk obligations from 2 August 2026 to 2 December 2027, entering into force in July 2026. Deferred is not cancelled — and if you're automating anything in HR, you are in scope.

The vocabulary gap is a governance gap

Here's what we see repeatedly in the field: the companies that get AI working aren't the ones with the best vocabulary. They're the ones where the buyer can interrogate the vocabulary.

When we at Mesh Flow scope an automation, the most productive hour is usually the one where we replace every abstract noun in the brief with a specific one. "An agent that handles vendor onboarding" becomes "reads the W-9 and COI, validates against the vendor master, flags mismatches to AP, auto-approves below $10k." The second version can be estimated, tested, and killed if it doesn't work. The first version can only be believed.

That's also why the gap between leaders and laggards is widening rather than closing. BCG's "future-built" companies — 7.5% of the market — show 2.3x the shareholder return of laggards. McKinsey's picture is blunter: 37% of organizations attribute at least some EBIT impact to AI, but only 6% qualify as high performers, and that share didn't budge year over year. The technology got better. The buying didn't.

Precision in language is the cheapest diligence available. It costs one meeting.

Frequently Asked Questions

What's the real difference between AI automation and RPA?

RPA follows fixed rules and mimics a user interface; AI automation applies a model's judgment to unstructured or ambiguous inputs. RPA breaks when the screen changes, AI automation is probabilistic and needs evals and review gates. Most durable systems use both, with RPA on the mechanical steps.

Is "agentic AI" just a buzzword?

No, but most products using the label aren't agentic. Gartner coined "agent washing" for the practice and estimates only about 130 of thousands of agentic vendors are genuine. A real agent chooses its own tools and recovers from unanticipated failures — ask to see it handle something the builder didn't plan for.

Do we need fine-tuning for AI that knows our company's information?

Usually not. "It doesn't know our policies" is almost always a retrieval problem, solved with RAG, which is cheaper to build and far cheaper to keep current. Fine-tuning changes behavior and style, not knowledge, and it locks you into a retraining cycle every time your documents change.

Why do our AI costs keep rising if token prices keep falling?

Because volume and context length grow faster than unit prices fall. Longer documents, longer histories, and multi-step agents each multiply token consumption per task. McKinsey found 20% of organizations say AI operating costs have constrained their use of the technology, and KPMG found 42% have only partial visibility into AI spend.

What should I ask a vendor to prove a system is production-ready?

Ask for the eval suite, the observability logs, the list of decisions gated by a human, and the per-task cost at your expected volume. If any of the four doesn't exist, you're looking at a pilot with a production price tag. (Ten questions that separate builders from slide decks.)

Does the EU AI Act still apply to us in 2026?

Yes, though the timeline shifted. The Digital Omnibus deferred high-risk obligations from August 2026 to 2 December 2027, and it entered into force in July 2026. If you automate recruitment, performance evaluation, or worker monitoring and operate in the EU, plan for the classification now rather than in late 2027.

The bottom line

Vocabulary isn't pedantry — it's the only free diligence you have. Every vague noun in a proposal is a place where risk hides.
The three words that matter most: eval (can you tell if it works?), observability (can you tell what it did?), and production (does it run on real volume with an owner?).
"Agent" is the most inflated term in the category. Gartner's 130-out-of-thousands estimate is the number to keep in your head.
Deferred compliance is not absent compliance. December 2027 arrives faster than an HR automation rebuild.

If you want help turning a vague AI brief into something you can actually estimate, test, and hold a vendor to, that's most of what we do at Mesh Flow.