AI automation terms. In plain English.

The written conditions a result must meet before the workflow owner accepts it: what must be correct, what may be uncertain, and what must stop the process.

Software that uses a language model to work toward a goal: it reads instructions, gathers context, chooses tools, and prepares or takes actions within limits you set. A good agent stops for a person when it is unsure.

A defined point where work stops until a named person accepts, revises, or rejects it. Sending, filing, paying, and deleting usually sit behind one.

A record of who asked for what, which sources were used, what was proposed, and who approved it. Required wherever money, records, or customers are affected.

The measured behavior of the current process or system before a change. Without one, an improvement claim cannot be checked.

The broader practice of redesigning a process so that software handles the predictable parts end to end. AI adds judgment for messy inputs; it does not replace the rules.

The amount of text a model can consider at once, measured in tokens. Long documents must be split or retrieved in parts.

Everything spent to produce accepted results (model calls, tools, infrastructure, retries, review labor) divided by the number of results a reviewer accepted.

How long inputs, outputs, and logs are kept, by whom, and where. Set per system in the data path, including the model provider.

Turning the fields in a document (supplier, amount, dates, clauses) into structured data with a link back to the passage each value came from.

The combination of extraction, comparison, and checklist review that gives a reviewer evidence rather than a summary.

Gradual change in results as sources, models, or inputs change while the workflow stays the same. Caught by regression tests and monitoring, not by hope.

Comparing an incoming record with existing ones under explicit rules, keeping the matching references so a reviewer can decide.

A numeric representation of text that lets software find passages with similar meaning. The mechanism behind semantic search and most retrieval systems.

Repeatable tests that compare a workflow’s outputs with expected results on a fixed set of cases. The way to know whether a change helped or hurt.

The place uncertain or failed items wait for a person, with the source and the reason attached. A workflow without one hides its failures.

Training a model further on your own examples to change its style or behavior. Rarely the first step; retrieval and better instructions solve most business cases.

A curated collection of inputs with known correct outputs, including the awkward cases, used to evaluate every version of a workflow the same way.

Checks around a model that constrain inputs, outputs, and actions: allowed tools, blocked topics, format validation, and stop rules.

A confident answer that is not supported by the sources. Reduced by retrieval, citations, and stop rules; never eliminated, which is why review exists.

The point where the workflow and its operating record pass to the team that will run it: instructions, checks, credentials, owners, and what to do when it should stop.

A workflow design where a person checks or approves specific steps before they take effect. The amount of review depends on risk and reversibility.

Designing an action so that running it twice has the same effect as running it once. Prevents duplicate records, emails, and payments after a retry.

The approved set of documents, policies, and records an assistant may answer from. Its ownership, versioning, and access rules matter more than its size.

A model trained on large amounts of text that can read, summarize, classify, extract, and draft language. It predicts likely text; it does not know your data unless you supply it.

How long a step takes from request to result. Matters for interactive use; matters less for overnight batches, where cost and correctness dominate.

Giving software and people only the access a task needs, for only as long as it needs it. The default for any connection an agent uses.

What an assistant retains between steps or sessions: the current task, recent decisions, and approved facts. Decide what is remembered, for how long, and who can see it.

An open standard for connecting assistants to tools and data sources through a common interface, so integrations can be reused across models and products.

The company whose models the workflow calls. Its terms decide training use, retention, and processing location, so the choice is recorded in the scope.

Several specialized agents that hand work to each other, for example a router, a researcher, and a drafter. More capable and harder to test than a single agent; use when one role is not enough.

Logs, traces, and metrics that show what a workflow did, which sources it used, and where it failed. The basis for monitoring and for investigating incidents.

The written instructions, examples, checks, and stop rules that keep the next run on the rails. Code can be regenerated; the record cannot.

Converting scanned images into text so the content can be read by software. Quality depends on the scan; unreadable pages should stop a check, not be guessed.

The code or platform that sequences steps, calls tools and models, handles retries, and records what happened. Where most of the reliability of a workflow lives.

The explicit list of what an agent may read, propose, and do, per system. Read, propose, and write are separate grants; writes that matter wait for approval.

A small, real scope built to answer a business question: does this workflow produce acceptable results at an acceptable cost? Not a demo, and not yet a rollout.

The instructions and context given to a model for one request. Good prompts state the goal, the sources, the format, and what to do when information is missing.

Instructions hidden in content the model reads (an email, a web page, a document) that try to make it ignore its rules. Treat all retrieved content as data, never as commands.

Rerunning the golden set after a change to confirm nothing that used to work has broken. Essential when models, prompts, or sources change.

A pattern where relevant passages are fetched from approved sources and given to the model with the question, so answers cite real documents instead of guessing.

Whether an action can be undone cheaply. Reversible actions can run with lighter review; irreversible ones wait for a person.

Scripts that imitate clicks and keystrokes in existing screens. Useful when there is no API; brittle when screens change. Often a stopgap rather than a foundation.

The standing instructions that define an assistant’s role, limits, and style for every request. Part of the operating record that keeps a workflow on the rails.

The unit models use to count text, roughly three-quarters of an English word. Providers price model calls per token, so tokens drive the model line of operating cost.

The ability of a model to request an action, such as looking up a record or drafting an email, through a defined interface. Which tools exist and what they may do is a scope decision.

A store optimized for finding embeddings that are close to each other. One of several ways to power retrieval; not required for every project.

Connecting the steps of a repeated business process (intake, checks, approvals, follow-ups) so routine work moves by rule and exceptions reach a person.