This glossary defines the terms Indian teams actually meet when they deploy agents, send email and WhatsApp, handle consent, invoice a customer or try to be visible in AI answers.
It is a working reference, not a keyword list. Each long entry states what the term is, what it is not, how it works, what it means in India and how to measure it.
Terms are grouped by theme so you can read a lane end to end: agentic systems, data protection, email deliverability, Indian messaging, voice and latency, money and tax, and AI discovery.
Every entry names its near-neighbours, because most operational mistakes come from confusing two terms that sound alike.
Citations carry a date and a source. Definitions are reviewed with the product, legal and finance pages they link to, and corrected when those pages change.
If a definition here conflicts with a contract, a statute or a notification, the contract, statute or notification wins.

What this glossary is for

A glossary is useful when it settles an argument. Teams in India adopt agent platforms, email infrastructure, WhatsApp templates and GST invoicing at the same time, often with different vendors and different lawyers in the room. The same word can mean three things across those conversations. A run trace is a log to one person and an audit artefact to another. A suppression list is a hygiene chore to one team and a legal control to another. This glossary exists to make the differences explicit before they become a build decision.

It is not a dictionary of marketing buzzwords and it is not a substitute for professional advice. The definitions are written for operators: the person who has to configure the DNS record, approve the template, sign the invoice or answer the auditor. That is why each entry spends time on mechanics and measurement rather than on origin stories. The test of an entry is whether it changes what you do on Monday.

Figure 1. The lanes in one view. Most real questions sit at the boundary between two of them, which is why the near-neighbour notes matter.

It is also not neutral about quality. Where a term is commonly misused, the entry says so. Where a number is often quoted without context, the entry explains the context or declines to quote it. Where the answer depends on a contract or a notification, the entry says that instead of inventing a rule. A glossary that is confidently wrong is worse than no glossary at all.

How to use it without reading all of it

Start with the lane that matches the decision in front of you, not with the alphabetical list. If you are buying an agent platform, read the agentic systems lane and then the data protection lane, because the second determines what you can deploy. If a campaign is landing in spam, read the email lane in order: authentication, placement, bounce classification and suppression. If a customer is disputing an invoice, read the tax lane before you read the pricing page.

Inside an entry, read the short version first. It is written to be enough for a meeting. Then read the definition and the mechanics, and skip to the worked example if you want to see the term applied. The near-neighbour paragraph is the one to read when two teams are disagreeing, because it usually names the distinction that resolves the argument. The FAQ is for the questions that come up after the meeting.

Figure 2. The five-minute path. The near-neighbour step is optional until two people in the room are using the same word differently.

Cite with the date. The glossary index carries a review date, every page links to primary sources where they exist, and a definition that was right last year may be wrong after a notification, a platform policy change or a product change. If you are quoting a definition in a policy document, quote the source, not this page. The glossary is a map to the source, not a replacement for it.

The standard every long entry follows

Every glossary term has a one-paragraph definition with links to the pages that depend on it. Some terms also have a long-form page, and those pages follow a fixed shape. The short version gives five or six sentences that can be read aloud. The definition comes first, before any history or opinion. Mechanics follow: how the thing actually works, step by step, including the parts that fail.

The India lens comes next. That means DPDP duties where personal data is involved, DLT and DoT rules where SMS is involved, GST and place of supply where money is involved, and INR prices stated exclusive of 18 percent GST where a product price appears. The implementation section describes what a team configures, stores or reviews. Measurement follows: the two or three numbers that tell you whether the thing is working, and the ones that mislead.

Figure 3. The standard shape. The India lens is not a decoration at the end; it often changes the definition itself.

Each page ends with common mistakes, a worked example and sources. The worked example is labelled as a sample where the numbers are illustrative, because a plausible number presented as a benchmark is a lie with a chart. The sources are primary where possible: an RFC, a regulator, a platform policy or an official portal. The page is reviewed on a schedule and when the underlying source changes.

Short version that fits a meeting.
Definition before history.
Mechanics including failure modes.
India lens: DPDP, DLT, GST and INR.
Implementation and measurement.
Common mistakes and a worked example.
Primary sources and a review date.

Theme one: agentic systems and the runtime

This lane covers the machinery of agents: the runtime that executes them, the capability that makes them agentic, the record they leave, the tests that gate a release and the rules that constrain an action. Start with agent OS if you are choosing a platform, agentic AI if you are explaining the category to someone who has only seen chatbots, and run trace if you are the person who will be asked to prove what happened.

The related terms are where the operational work lives. An eval is a repeatable test, not a demo. A guardrail is a rule enforced in code, not a sentence in a prompt. RAG is a way to ground answers in a controlled corpus, not a guarantee of accuracy. MCP is a tool protocol, not an authorisation model. A human gate is a control with an owner and a timeout, not a queue that nobody checks.

Figure 4. The agentic lane in one picture. The evidence layer is what turns a demo into something an auditor can review.

The common mistake in this lane is buying one term and expecting another. A framework is not a runtime. A model is not an agent. An observability tool is not an audit record. When a vendor uses these words interchangeably, ask for the trace of a failed run and the policy decision that allowed the action. The answer tells you which term they actually sell.

Theme three: email deliverability

This is the largest lane because email has the most acronyms and the most expensive misunderstandings. It starts with authentication: SPF authorises the envelope domain, DKIM signs the content, DMARC aligns a passing result with the visible From domain, and BIMI displays a verified logo only after enforcement. These are not interchangeable, and none of them is a spam filter.

Placement comes next. A seed test samples a panel of inboxes; inbox placement is the share of delivered mail that reaches the inbox; a complaint rate is complaints divided by delivered mail. Each has a denominator and a method, and a number without them is not a measurement. Then comes list quality: a hard bounce is permanent, a soft bounce is temporary, a suppression list is a durable prohibition, and IP warm-up is the gradual building of sending reputation on a new address.

Figure 6. Three layers, three kinds of evidence. A passing SPF check says nothing about placement, and a good panel says nothing about consent.

The lane rewards reading in order. Teams that start with copy and subject lines, before authentication and list hygiene, spend weeks optimising the wrong variable. Teams that start with authentication, then suppression, then placement, then content, usually find the problem in the first two steps. The glossary entries are ordered to support that sequence rather than to follow the alphabet.

Theme four: WhatsApp, SMS and Indian messaging

Indian messaging has a regulatory layer that most global documentation ignores. DLT is the registration system for commercial SMS: the principal entity, the six-character header, the approved template and the consent record. DND is the preference set with the operator. A BSP is a Meta-approved provider for the WhatsApp Business Platform. An HSM template is a pre-approved WhatsApp message with a category and variables. Opt-in is the permission that both systems require, captured and evidenced.

The near-neighbour traps are everywhere. DLT is not WhatsApp template approval; clearing one does not clear the other. A BSP is not a carrier. WhatsApp template categories are not marketing labels to be chosen by price. An opt-in for order updates is not an opt-in for a festive offer. The glossary treats these as separate entries because conflating them produces silent delivery failures and, occasionally, a regulatory notice.

Figure 7. Registration, approval and consent are three gates. A message can clear one and fail the next, which is why the error is often silent.

The India lens also includes language. A template that reads well in English may fail in Hinglish or in a regional script, and the variables may render differently on a low-end handset. That is a product quality issue, not a policy issue, but it shows up in the same complaint queue. Test the rendered message, not just the approved text.

Theme five: voice and real-time agents

Voice has a different physics from text. A caller expects a response in well under a second, and a pause that feels fine in chat feels broken on a call. That is why the lane centres on latency and turn taking. Barge-in is the ability to interrupt the agent mid-sentence and be heard. p50 and p95 latency are the median and the slow tail, and the tail is what callers remember. Hinglish ASR is speech recognition for code-mixed Hindi and English, which is the normal register for a large share of Indian callers.

The near-neighbour distinction is between a voice demo and a voice system. A demo plays a clean recording with a quiet room and a cooperative caller. A system handles background noise, names, numbers, dates, interruptions, silence and a caller who switches language mid-sentence. The glossary defines the terms around that gap: the latency percentile, the interruption behaviour and the recognition quality on real audio.

Figure 8. Sample latency by stage. The median looks comfortable; the p95 is where callers start talking over the agent.

The India lens includes telephony quality, regional accents and the cost of a connected minute. A voice agent that works in a quiet office in Bengaluru may fail on a noisy street in Kanpur. Test with recordings from the actual customer base, not with studio audio, and measure the tail rather than the average.

Theme six: money, tax and billing

This lane covers the paperwork that surrounds a subscription. A GSTIN is the fifteen-character registration number that must be validated and printed on a tax invoice. A SAC code classifies the service and sits with the description, the taxable value and the rate. GST is charged in addition to the published price, and LoopNow’s plans are quoted exclusive of 18 percent GST, so a ₹7,999 Growth plan becomes ₹9,438.82 with tax when the supply is intra-state and the split is equal.

The near-neighbour trap is treating the SAC as a rate. The code describes the service; the rate comes from a notification; the place of supply decides whether the tax is central and state or integrated. Two invoices can carry the same code and different treatment if the facts differ. That is why the entry refuses to pick a code for the reader and instead explains the questions an accountant or an advance ruling must answer.

Figure 9. Sample arithmetic for a ₹7,999 plan at 18 percent GST. The published price is not the invoice total, and the tax heads depend on place of supply.

The India lens is the invoice itself. A customer’s accounts payable team reads the description, the code, the tax heads and the GSTIN together. A wrong head with a correct total is still a rejected credit. The glossary links the tax terms to the pricing page so that the number a buyer sees and the invoice they receive can be reconciled without a support ticket.

Theme seven: discovery in AI answers

GEO, AEO and AI visibility are the newest lane and the one most likely to be sold as magic. GEO is the practice of making a brand visible and citable in generative answers. AEO is the overlapping discipline of structuring content so an answer engine can extract a direct answer, a table and a source. AI visibility is the measurement: mention rate, citation rate, share of voice and answer accuracy across a prompt panel.

The near-neighbour distinction is between being mentioned and being cited. A model can mention a brand because it appeared in training data, without linking to the page. A citation is stronger evidence that the page was retrieved and used. Neither is the same as a ranking, and neither is stable across providers or over time. A single screenshot of a favourable answer is an anecdote, not a measurement.

Figure 10. Four measures, four questions. A mention without a citation is weaker evidence, and neither is a ranking.

The India lens is language and source quality. An answer in English may cite global sources while the customer asking in Hindi gets a different set. Local sources, original data and clear entity information matter more than keyword density. The glossary treats GEO as a content and measurement discipline, not as a trick for inserting a brand into a model.

Near-neighbour pairs that cause real mistakes

Some pairs deserve to be memorised. Opt-in and suppression: permission versus prohibition. Hard bounce and soft bounce: permanent versus temporary. SPF and DMARC: envelope authorisation versus alignment with the visible From domain. SAC and HSN: services versus goods. Data fiduciary and data processor: decides the purpose versus processes on instruction. Run trace and audit log: one agent decision versus a stream of access events. AEO and GEO: content structure versus the broader visibility practice.

The pattern behind most of these is a category error between a record and an action, or between a permission and a prohibition. A team stores an action in a record field, or treats a permission as if it were evidence. The glossary entries name the pair explicitly so the distinction can be quoted in a design review, which is usually where it matters more than in a blog post.

Figure 11. Six pairs, six distinctions. Most of them are a record versus an action, or a permission versus a prohibition.

When a pair is genuinely ambiguous in your context, write down the definition your team is using and put it in the design document. A shared definition that is slightly imprecise is better than two precise definitions in conflict. The glossary can supply a starting point, not a verdict on your contract.

How terms are added, reviewed and kept honest

A term earns an entry when it appears in a real decision: a procurement question, a build review, a customer dispute or an audit. That test keeps the glossary from becoming a list of every acronym in the industry. A term earns a long-form page when getting it wrong has operational consequences, which is why SPF and suppression have long pages while a passing mention of a vendor category may only have a paragraph.

Each entry has an owner in the function that depends on it. Product owns the agentic terms, legal and security own the data protection terms, the email team owns deliverability, finance owns tax. The owner reviews the definition when the underlying source changes and at least once a year. A correction is published with a date rather than silently edited, because someone may have cited the old version.

Figure 12. The lifecycle of an entry. The annual review is the minimum; a source change or a correction restarts the cycle sooner.

Retirement is part of the standard. When a term stops being used, when a regulation is replaced or when a platform changes its vocabulary, the entry is marked as historical and kept for reference. Deleting it breaks links and loses the explanation of why an older document says something different. A glossary is a record of usage as well as a definition of terms.

The editorial limits are part of the same discipline. The glossary will not invent a benchmark, a customer logo, a certification or an uptime number to make an entry feel authoritative. If a figure is illustrative, it is labelled as a sample, and if a rate depends on a notification, the entry says so. It will not give legal, tax or medical advice either: the India lens explains how a rule applies in practice, while the application to your facts needs your counsel or your accountant. Nor will it stay finished, because products change, regulators clarify and vendors rename their categories. If you find something wrong, the correction is more valuable than the agreement.

Frequently asked questions

Who writes these definitions?

They are written by the LoopNow team with the function that owns the term, then reviewed against primary sources. Legal and security review the data protection entries, finance reviews the tax entries and the email team reviews deliverability. The agent card on each page is a starting point for questions, not a substitute for the named owner.

Can I cite a definition in a policy document?

You can cite it with the page URL and the review date, but cite the primary source where one exists. A policy should rest on the RFC, the statute, the notification or the platform policy, not on a vendor glossary. The glossary is useful for explaining the term to colleagues and for finding the source quickly.

Why do some terms have long pages and others only a paragraph?

A long page is justified when getting the term wrong has operational consequences: a failed authentication setup, a leaked suppression, a rejected invoice or an unauditable agent action. A short entry is enough when the term is descriptive and the decision does not depend on a mechanism. The standard is consequence, not popularity.

How often are entries reviewed?

At least once a year, and immediately when the underlying source changes. A regulator clarification, an RFC update, a platform policy change or a product change can all invalidate a definition. The glossary index carries a review date so a reader can judge how current the set is. Corrections are published with a date rather than silently edited.

Does the glossary cover Indian languages?

It covers terms that matter when you operate in India, including Hinglish ASR, DLT and regional deliverability. It does not attempt to be a translation dictionary. Where language quality affects a channel, the entry describes the test to run, such as real call recordings or rendered WhatsApp templates, rather than offering a phrasebook.

Why are prices shown without GST?

Because published plan prices are exclusive of 18 percent GST, and mixing the two produces confusion on an invoice. Where a price appears, the entry says it is exclusive of GST and shows the arithmetic separately. The invoice total depends on the plan and the place of supply, so a single inclusive number would be wrong for some readers.

What is the difference between the glossary and the blog?

The glossary defines a term and gives the minimum mechanics needed to act on it. The blog goes deeper into a situation, such as a deliverability programme or a migration. A glossary entry should remain useful when the blog post is outdated. The two link to each other, but they have different jobs and different review cycles.

Can I suggest a term or a correction?

Yes. The most useful suggestions come with a source and a sentence about the decision the term affects. A correction that names the page, the claim and the better source can be reviewed quickly. Suggestions without a source are still read, but they take longer because someone has to verify the claim before it is published.

Does a definition here override a contract?

No. A contract, a statute, a notification or a platform policy takes precedence over anything on this page. The glossary explains how a term is commonly used and how it works in practice. If your contract defines a term differently, use the contract for that relationship and use the glossary to understand the surrounding mechanics.

Definitions you can cite.

4 Nov 2026

The runtime layer that plans work, routes models, calls tools, stores memory and records a run trace across voice, WhatsApp and email.

AI that plans a sequence, calls tools, observes the result and keeps going until the task is done — an action loop, not a chat bubble.

A sales development agent that researches accounts, scores ICP fit, runs sequences and logs every touch to the CRM before it hands off.

A workflow pause that waits for a named person to approve a high-value or high-risk agent action, and leaves an audit entry behind.

The audit artefact for one agent run: the plan, every tool call, latency, token and rupee cost, policy decisions and human approvals.

A repeatable test set with expected outcomes, scoring whether the agent completes the task, escalates correctly, records consent and refuses cleanly.

A code-level limit on what an agent may do: block a tool call, redact PII, force a human gate or refuse a topic outright.

Retrieval-augmented generation fetches relevant passages from a controlled knowledge base at answer time instead of trusting model memory.

A standard way for an agent to discover and call tools; LoopNow exposes workspace tools over MCP with scoped keys for Claude and Cursor.

India’s data protection law of 2023. Consent, notice, data-principal rights and breach reporting each need a working system, not a policy page.

What a person actually agreed to: notice version, purpose, timestamp, collection point and withdrawal path. A checkbox alone is not evidence.

Under the DPDP Act, whoever decides the purpose and means of processing; LoopNow sits as a processor for customer marketing data.

The individual a record is about. Their access, correction, erasure and nomination requests need a workflow that can find and delete the data.

Data collected for support cannot be reused for outbound marketing without a new lawful basis, so each workflow declares its purpose.

A scheduled deletion that runs after the stated retention period and is logged, covering primary storage, logs, analytics and backups.

A DNS TXT record naming the servers allowed to send mail for a domain. On its own it is weak, so pair it with DKIM and DMARC.

A cryptographic signature on outgoing mail that lets receivers confirm the message was not altered and came from an authorised domain.

The policy telling receivers what to do when SPF or DKIM fails. Start at p=none and monitor, then move to quarantine or reject.

A verified brand logo shown in supported inboxes, which needs strong DMARC enforcement and a verified logo certificate before it renders.

One-click unsubscribe headers that let a mailbox provider act on the user’s behalf, cutting complaints and expected of bulk senders.

A send to a controlled panel at Gmail, Outlook, Yahoo and Indian providers to check placement, rendering and authentication before launch.

The share of delivered mail that reaches the inbox rather than spam, measured with seed panels and always quoted with panel and date.

Spam complaints divided by messages delivered. It damages reputation fast, so segment by engagement and suppress non-openers first.

A permanent failure where the address does not exist or will never accept mail. Suppress it at once, because repeats erode sender reputation.

A temporary failure such as a full mailbox or a briefly unavailable server. Retry with a limit, then suppress after repeated attempts.

Unsubscribes, hard bounces and complaints in one file that must be imported before the first campaign on any new platform.

Building sending reputation by starting with small, engaged volumes and adding more over days; a cold dedicated IP loses to a shared one.

A WhatsApp template Meta has approved, with fixed category and variables, required for anything sent outside the customer service window.

India’s registration system for commercial SMS entities, headers and templates; OTP and SMS still depend on it even where WhatsApp is approved.

A Meta-approved WhatsApp Business Solution Provider. Number migration, template approval and display-name changes all run through this partner.

Specific, informed and recorded permission before you message someone, with an opt-out path at least as easy as the way in.

Letting a caller interrupt a voice agent mid-sentence; the agent has to stop speaking, listen and resume without losing the thread.

Median and tail response time. Voice quality hinges on p95 rather than a best-case demo, so always state the percentile and region.

Speech recognition for Hindi-English code-switching inside one sentence, tested on real names, numbers and dates rather than studio audio.

The AWS Mumbai region code. A data residency claim should say what is stored there and which sub-processors receive data outside India.

India’s 15-character GST registration number. Validate it before invoicing and print it with the place of supply on every tax invoice.

The service accounting code that classifies a taxable service; SAC 998314 is usual for IT design and development. Confirm it with your accountant.

Generative Engine Optimization: making a brand visible and citable in AI answers through entity clarity, original data and citation-friendly pages.

Structuring content so an answer engine can lift a direct answer, a table and a source; Google treats this as SEO, not a separate discipline.

Measured with a prompt panel: mention rate, citation rate, share of voice, citation share and answer accuracy, tracked as absolute counts.

A per-contact record of grant, withdrawal and notice versions, forming the evidence layer behind every DPDP response and every campaign.

The short version