Terms, defined plainly

A2P 10DLC

Application-to-Person messaging over a 10-Digit Long Code: the US carrier registration system for business text messages sent from ordinary 10-digit phone numbers. Businesses register a brand (who they are) and a campaign (what they send and why); unregistered traffic is filtered or blocked by carriers rather than rejected with a clear error.

Registration typically takes days to weeks, so it should be started well before a launch date.

TCPA

The Telephone Consumer Protection Act, the primary US federal law governing automated calls and texts to consumers. It requires prior consent for many message types, mandates working opt-out handling, and carries statutory damages per message, which is why consent records and enforced opt-outs matter operationally, not just legally.

Opt-out (STOP)

A customer's instruction to stop receiving messages, sent by replying with a keyword such as STOP, STOPALL, UNSUBSCRIBE, CANCEL, END, or QUIT. Once received, further messages to that contact must not send until the customer opts back in (typically by texting START or UNSTOP). Well-built systems enforce this in the send path itself rather than relying on staff to remember.

Quiet hours

Time windows during which businesses should not send messages, based on the recipient's local time rather than the sender's. Some US jurisdictions regulate these windows directly; beyond compliance, sending overnight reliably damages engagement and reputation.

Grounding

Restricting an AI system's answers to a specific, approved source of truth rather than letting it generate text freely. A grounded AI receptionist answers a pricing question only from a price list the business approved, and defers to a human when no approved source matches, instead of inventing a plausible number.

The opposite failure mode is 'hallucination' - confident, fluent, and wrong.

Guardrail

A deterministic check that runs before or around an AI system's normal behavior to prevent a category of unacceptable output. Typical examples are refusing to answer clinical or legal questions, detecting prompt-injection attempts, and escalating to a human on specific keywords.

Prompt injection

An attempt to manipulate an AI system by embedding instructions in user input, such as a message telling the assistant to ignore its rules or reveal its configuration. In a customer-messaging context the safe response is to stop automated handling and route the conversation to a person.

Escalation

Handing a conversation from automated handling to a human, either because the customer asked for a person, because the topic is sensitive, or because the AI has no confident grounded answer. A conversation that has been escalated should stay with a human until someone decides otherwise.

Speed to lead

How quickly a business responds to a new inbound inquiry. It matters because a prospective customer contacting a service business is usually contacting several, and response time strongly influences who wins the job. Automated first responses exist mainly to compress this interval.

Missed call text-back

Automatically sending a text message to someone whose phone call went unanswered, typically within seconds, so the inquiry continues by text rather than ending in voicemail. In the US this is still subject to messaging rules including carrier registration and opt-out handling.

Unified inbox

A single queue where messages from every channel a business uses - SMS, WhatsApp, Instagram, Messenger, email, website chat - appear together, usually grouped per customer, so staff work one list instead of watching several apps.

Identity resolution

Recognizing that contacts arriving through different channels are the same person, and merging them into one record. For example, the same phone number messaging on SMS and on WhatsApp should resolve to a single contact with a single timeline rather than two unrelated records.

Lead scoring

Assigning a numeric score to a prospective customer to prioritize follow-up, based on signals such as the channel they came from, the value of the service they asked about, and stated urgency. A score is only useful if its breakdown is visible, so staff can tell whether they agree with it.

Approval mode

An operating setting where an AI system composes a reply but does not send it, leaving a draft for a person to approve, edit, or reject. It is the standard way to build confidence in automated messaging before allowing fully autonomous sending.

Structured facts

Discrete, business-specific data points an AI system may answer from - services and prices, opening hours, policies, staff - stored as records rather than free text. They are more reliable than prose documents because they can be verified, edited, and expired individually.

Longer treatments of several of these live in our guides and blog.