AI and WhatsApp automation, defined

The WhatsApp Business API is Meta’s official programmatic interface for businesses to send and receive WhatsApp messages at scale. Unlike the free WhatsApp Business app, it has no chat interface of its own — it is accessed through software, which is why businesses use it via a provider platform. Every legitimate WhatsApp automation product in India, including Skanda AI’s, runs on this same API, so message delivery is identical across providers.

WhatsApp platform →

A Business Solution Provider is a company approved by Meta to give businesses access to the WhatsApp Business API, usually through a dashboard the business logs into. Because all BSPs connect to the same Meta API, they compete on platform pricing, features, custom development and support rather than on deliverability. The practical questions to ask any BSP are whether they mark up Meta’s conversation fees, whether they charge per agent seat, and whether the WhatsApp Business Account will be in your name.

How to choose a provider →

A Meta Tech Provider is a company authorised to build and operate software on the WhatsApp Business Platform for other businesses, including running embedded signup so a client can connect their own WhatsApp Business Account. Skanda AI Solutions is an approved Meta WhatsApp Business Tech Provider, which is what allows its multi-tenant platform to onboard clients onto their own numbers rather than reselling shared infrastructure.

A WhatsApp Business Account is the Meta-side container that holds your business’s WhatsApp phone numbers, message templates and quality ratings. Ownership of the WABA determines how easily you can change providers: if it sits in your own Meta Business Manager, you can migrate your number to a different platform, and if the provider holds it on your behalf, leaving is considerably harder. Confirm which arrangement applies before signing with any provider.

A message template is a pre-written WhatsApp message approved by Meta in advance, required whenever a business starts a conversation rather than replying to one. Templates are submitted for review and approved, rejected or flagged, typically within minutes to a few hours. They are categorised as utility, marketing or authentication, and the category directly affects what Meta charges — utility costs less than marketing, so categorising honestly saves money as well as avoiding rejection.

Templates documentation →

Meta charges for WhatsApp messaging by conversation category rather than per message. Utility conversations cover transactional updates such as order confirmations and appointment reminders; marketing conversations cover promotional content; authentication conversations cover one-time passcodes; and service conversations are those the customer starts. Utility costs less than marketing and customer-initiated service conversations are often free, which makes conversation design a direct lever on cost.

Pricing explained →

The 24-hour window is the period following a customer’s message during which a business can reply on WhatsApp with free-form content rather than an approved template. Once it closes, re-opening the conversation requires a template message, which carries a charge. This is why well-designed WhatsApp automation resolves enquiries inside the window rather than letting them lapse into a paid template send.

WhatsApp Flows are structured, form-like interfaces that open inside a WhatsApp chat, letting a customer pick dates, choose options or submit details without leaving the conversation. They replace long back-and-forth question sequences with a single screen, which raises completion rates and reduces the number of billable messages needed to capture the same information.

Flows documentation →

WhatsApp assigns each business phone number a quality rating — green, yellow or red — derived from how recipients react to its messages, principally blocks and reports. A falling rating reduces how many customers the number may message per day and can ultimately restrict it entirely. Messaging only people who expect to hear from you protects the rating far more effectively than any technical remedy applied afterwards.

An AI voice agent is a system that answers or places phone calls and holds a spoken conversation, understanding what the caller says rather than making them navigate a keypad menu. It typically combines speech recognition, a language model and speech synthesis, wired into a telephony layer. In production it books appointments, qualifies leads, answers product questions and hands the call to a person when the request exceeds what it should handle alone.

AI voice agents →

Interactive Voice Response is the traditional phone menu system where a caller presses digits to route themselves — press 1 for sales, press 2 for support. It is cheap and predictable but limited to the options recorded in the tree, so any request the menu does not anticipate ends in a queue or a dead end. AI voice agents differ by letting the caller state the problem in their own words.

AI voice agent vs IVR →

Real-time agent assist keeps a human on the call and puts AI behind them. It integrates with an existing dialer, streams the customer’s speech through a language pipeline as they talk, and surfaces a recommended response on the agent’s dashboard, grounded in the organisation’s approved answers and policies. It is used where accuracy and compliance make full automation inappropriate, and it doubles as a training tool because new agents see the correct answer while they learn.

Real-time agent assist →

Asterisk is open-source telephony software that lets an organisation run its own phone system rather than routing calls through a cloud provider. Self-hosting on Asterisk becomes worthwhile at sustained high call volume, when operating many phone numbers that each need correct caller-ID presentation, or when call audio cannot legally leave the organisation’s own infrastructure. Below those thresholds, cloud telephony is cheaper and far less operational work.

Self-hosted vs cloud telephony →

PJSIP is the SIP protocol stack Asterisk uses to set up and manage phone calls, and ARI — the Asterisk REST Interface — is the API through which software controls those calls programmatically while they are in progress. Together they are what allows an AI agent to be placed inside a live call on self-hosted infrastructure, with full control over routing, transfer and caller-ID presentation per number.

Code-switching is moving between two languages inside a single conversation or even a single sentence, as in Hinglish (Hindi and English) or Tenglish (Telugu and English). It is the normal way most Indians speak, and it breaks speech systems built for one language at a time. Handling it correctly requires recognition tuned for mixed-language input and speech output rendered in native script rather than Latin transliteration, which is otherwise mispronounced.

A Large Language Model is a neural network trained on very large volumes of text to predict and generate language, which in practice lets it answer questions, summarise, classify and hold conversations. Well-known examples include the GPT, Claude and Gemini families. An LLM alone knows nothing about a specific business — making it useful commercially requires connecting it to that business’s data, rules and systems.

Retrieval-Augmented Generation connects a language model to a source of current information — documents, policies, catalogues, an internal wiki — by retrieving the relevant passages at question time and giving them to the model as context. It is the right approach whenever the underlying facts change, because updating an answer means re-indexing a document rather than retraining anything. It also allows the system to cite which source produced an answer, which matters for audit.

RAG vs fine-tuning →

Fine-tuning further trains an existing language model on examples specific to a domain, changing how the model behaves rather than what it can look up. It is the right choice for specialised vocabulary, a strictly consistent output format, a particular tone, or reducing inference cost at high volume by making a smaller model perform adequately. It is the wrong choice for information that changes, because retraining is slow and a model trained on last quarter’s policy is confidently out of date.

RAG vs fine-tuning →

A hallucination is a confident, fluent, factually wrong statement produced by a language model. It happens because the model generates plausible language rather than looking facts up, so it fills gaps rather than reporting them. Production systems reduce hallucination by grounding answers in retrieved sources, constraining what the model is permitted to assert, and escalating to a human when confidence is low — not by asking the model to be more careful.

Agentic AI describes systems that take actions rather than only producing text — calling APIs, updating a CRM, scheduling a job, moving a record through a workflow — and that decide which action to take based on the situation. The distinction from a chatbot is consequence: a chatbot tells the customer their order status, an agentic system reschedules the delivery. Building them safely means bounding what actions are permitted and logging every one.

Agentic AI systems →

A vector database stores text as numerical representations called embeddings, which capture meaning rather than exact wording, and retrieves passages by semantic similarity to a query. It is the retrieval half of a RAG system: it is what lets a question phrased one way find a document that answers it phrased another way. Common options include Pinecone, Weaviate, Qdrant and the pgvector extension for PostgreSQL.

The context window is the maximum amount of text a language model can consider at once, covering the instructions, the conversation so far and any retrieved documents. Exceeding it means something must be dropped, which is why long conversations lose earlier detail and why retrieval systems select passages rather than submitting entire documents. Larger context windows reduce but do not remove the need for careful selection, because longer prompts also cost more and can dilute accuracy.

Prompt engineering is the practice of writing the instructions given to a language model so it behaves reliably — specifying its role, the rules it must follow, the format to return and what to do when it does not know. In production systems the prompt is treated as source code: versioned, tested against real cases and changed deliberately, because a small wording change can alter behaviour across thousands of conversations.

Computer vision is machine learning applied to images and video — detecting objects, reading documents, spotting defects, counting items, verifying identity. In Indian industrial settings the most common uses are quality inspection on production lines and document extraction in back-office operations. Vision models can run at the edge on local hardware, which keeps images inside the facility and removes dependence on connectivity.

Computer vision →

Edge AI runs models on local hardware — a device, a camera, an on-premise server — rather than sending data to a cloud service. It is chosen when latency must be low, when connectivity is unreliable, or when the data legally or contractually cannot leave the site. The trade-off is that models must be small enough for the available hardware, which usually means optimising or distilling a larger model.

MLOps is the engineering discipline of running machine-learning systems in production — versioning models and data, deploying updates safely, monitoring for accuracy drift, and retraining on a schedule or a trigger. It is the difference between a model that worked in a notebook and one that still works in eighteen months, and it is the part most commonly omitted from AI projects that later fail quietly.

Model drift is the gradual decline in a model’s accuracy as the real world moves away from the data it was trained on — new products, changed customer behaviour, a shifted seasonal pattern. It is dangerous because nothing breaks visibly; predictions simply get worse. Detecting it requires monitoring live accuracy against outcomes rather than assuming a model that passed testing remains correct.

Generative Engine Optimisation is the practice of structuring content so AI assistants such as ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews can retrieve it, understand it and cite it when answering a question. It differs from traditional SEO in what it competes for: not a position in a list of ten links, but inclusion among the handful of sources a model draws on to compose one answer. In practice it requires permitting AI crawlers, writing self-contained factual statements, and marking up entities consistently.

SEO & GEO engineering →

Answer Engine Optimisation is structuring content so it can be lifted directly as the answer to a specific question, in featured snippets, voice results or AI-generated responses. The core technique is writing each answer to stand alone without surrounding context — no unexplained pronouns, no "as mentioned above" — so an extracted paragraph remains correct and complete on its own.

An llms.txt file is a plain-text document at a website’s root that summarises the site’s content and key facts in a form language models can read efficiently, analogous to robots.txt but for comprehension rather than permission. It is an emerging convention rather than a standard any model is obliged to honour, and it costs little to publish. Skanda AI generates its llms.txt from the same typed data that renders its pages, so it cannot drift out of date.

Structured data is machine-readable markup, usually JSON-LD following the schema.org vocabulary, that states explicitly what a page is about — that this is a service, this is its provider, this is the question and this is its answer. It removes the need for a search engine or model to infer meaning from prose. Using stable identifiers so every page references the same organisation entity is what lets a crawler resolve an entire site to a single business.

AI crawlers are automated agents that fetch web pages on behalf of AI systems, either to train models or to retrieve current information when answering a user’s question. Named examples include GPTBot and OAI-SearchBot from OpenAI, ClaudeBot from Anthropic, PerplexityBot, and Google-Extended. They are controlled through robots.txt, and blocking them removes a site from consideration entirely — a model cannot cite a page it was never permitted to read.

The Digital Personal Data Protection Act 2023 is India’s data protection law, governing how organisations collect, process and store the personal data of individuals in India. For AI systems it bears on consent for processing, limiting collection to what is necessary, securing data in transit and at rest, and being able to delete an individual’s data on request. Systems that record conversations or call audio need a defined retention position rather than an implicit one.

An on-premise deployment runs software, including AI models, on infrastructure the client controls rather than in a vendor’s cloud. It is chosen where data cannot leave the organisation for legal, contractual or competitive reasons. It is fully achievable for conversational, vision and fine-tuned models, at the cost of the client owning capacity planning and operations, and it removes the risk of a vendor changing terms on a system the business now depends on.

Human handover is the point at which an AI system stops handling an interaction and passes it to a person, carrying the context across so the customer does not repeat themselves. Designing it well matters more than raw model capability: a system that recognises its limits and escalates cleanly outperforms a more capable one that strands the customer when it fails. Good handover passes the transcript, the captured details and the reason for escalation.

A multi-tenant platform serves many client organisations from one deployment while keeping each client’s data, configuration and credentials strictly separated. For WhatsApp automation this means each client’s phone number, access tokens and conversations are isolated, with messages routed by phone number identifier and credentials encrypted per client. It is what makes a platform affordable — the infrastructure cost is shared — provided the isolation is genuine.

Idempotency is the property that repeating the same request produces the same result as making it once, rather than duplicating the effect. In messaging it is what prevents a customer receiving the same notification three times because a network timeout caused your system to retry. It is implemented by attaching a unique reference to each request so the server recognises a repeat and returns the original outcome instead of acting again.

API reference →