AI automation glossary: the words on this site, defined.

Terms

AI agent

Software that takes a goal, decides the steps, uses tools and data to carry them out, and stops to ask a person when a rule says it must. The difference from a chatbot is the last clause and the middle one: an agent acts on systems, and it has rules about when not to. See AI support agent vs chatbot.

AI workforce

Aldenebai's name for a set of agents that each take over one whole job: support tickets, QA, engineering tickets, phone calls, bug fixing, documentation. Not assistants that help a person do the job; systems that do the job, with a person on the exceptions. See the systems.

Confidence threshold

The score below which an AI system will not act on its own. Above it, the agent replies, closes, or opens the pull request. Below it, nothing is sent and the case goes to a person with the agent's notes. The threshold is a dial your team controls; raising it trades autonomy for safety.

Human-in-the-loop

A design where a person approves, corrects, or takes over specific decisions inside an otherwise automated process. In Aldenebai's systems the loop is placed on irreversible actions, such as merging code, issuing a refund, or changing an account, and on anything below the confidence threshold. Full guide →

Escalation

Handing a case from the machine to a person. Done well, the person receives the machine's work so far: the intent it assigned, the sources it found, its draft, and the reason it stopped, so they start from a summary rather than from scratch.

Allowed-action list

The fixed set of actions an agent may take, for example reply, close, tag, open a pull request. Anything not on the list, such as a refund or a deletion, requires a person. It is the difference between an agent with admin rights and an agent with a job description.

Audit trail

A log of every read, decision, and action a system took, with the reason, reviewable afterwards. In production this is what makes an AI system arguable with: when a result looks wrong, you can see exactly what it saw and why it did what it did.

Kill switch

A control that stops an automated system instantly, whatever it is doing. After the switch, work routes back to people exactly as it did before the system existed. Every Aldenebai system ships with one.

Self-healing test

An automated test that repairs itself when the product changes on purpose, a renamed button or a moved field, while still failing when the product actually breaks. Telling the two apart is the job. See AI test automation.

Build gate

An automated compile-and-test step that must pass before a pull request can be reviewed or merged. In Aldenebai's dev agent, a red build triggers an automatic retry in which the agent reads the failure and fixes its own change, up to a limit, before a person is asked.

Ticket-to-PR

A pipeline that turns an engineering ticket into a review-ready pull request with code and tests, for a human to approve. The human never stops approving; what disappears is the setup and the boilerplate. See the AI dev agent.

Automation Map

A one-page inventory of the repetitive work in a company, scored by volume, time, error cost, and reversibility, and ranked by return and risk, with a written "not yet" column. It is what you receive after the Automation Audit. Blank template →

Automation Audit

Aldenebai's free, 30-minute remote call in which the repetitive work in your company is mapped live and the written Automation Map is sent within 48 hours. No pitch; the map is yours either way. Book it →

Pilot

The first version of a system on a slice of real traffic, for example the top five ticket types or one repository, together with its monitoring and its tests, and a target named before it starts. It proves the number on your data before anything is widened. See how pricing works.

Scale model

An animated demonstration on this site of how a production system behaves, using simulated traffic. Scale models show the shape of the flow; they are not live data and are labelled as such wherever they appear.

Watchdog

A scheduled check that looks for work stuck past its time limit and flags it. On the ticket-to-PR pipeline at the marina-software client it runs every 30 minutes and flags anything in flight for more than an hour, which is what stops a silent failure from looking like an idle queue.

Orchestrator

The component that decides what runs next, holds the state while it waits, retries what failed, and tells someone when it cannot proceed. n8n, Zapier, and Make are orchestrators; so is a small service you write yourself. Keeping the orchestrator thin, with the heavy work in separate workers, is what makes a pipeline debuggable.

n8n

An open-source workflow orchestrator that can be self-hosted, so credentials and data stay on your own infrastructure. Aldenebai uses it as the orchestration layer for the AI dev agent and the bug fixer. See n8n consulting and n8n vs Zapier vs Make.

Guardrail

Any limit that makes an autonomous system safe to run: a confidence threshold, an allowed-action list, a build gate, a per-queue cap, an audit trail, a kill switch. A guardrail is enforced by permissions or code, not by an instruction in a prompt, which is the difference between a guarantee and a wish.

Deflection

The share of visitors who did not open a ticket after a chatbot conversation. It rises when a widget is hard to escape, so it cannot tell you whether anyone was helped. Resolution with no human touch is the honest counterpart. See agent vs chatbot.

Retrieval (RAG)

Giving a model the specific documents it needs at the moment it answers, instead of hoping the answer is in its training data. Every Aldenebai support answer is retrieved from approved sources first; where no source exists, the agent escalates rather than inventing one.

Regression suite

The set of automated tests that check the things that already worked still work, run before a release. Its value is entirely in whether people trust it: a suite that is often red for irrelevant reasons gets ignored, which is worse than having none.

Precision gate

The accuracy floor a data pipeline must clear on a hand-labelled test set before it is allowed to write anything. On the dealership enrichment pipeline it is 95%: answers above it are written to the CRM, everything else is flagged for a person.

Self-hosted

Running software on infrastructure you control rather than a vendor's cloud, so credentials, payloads, and logs stay inside your perimeter. It costs upkeep and buys data residency, which is why Aldenebai self-hosts the orchestrator for anything touching tickets, code, or customer records.