AI terms, in plain English.

A

Action boundary

The list of things an agent may do, beyond reading or suggesting. Sending a message, posting an entry or creating a ticket are actions, and each one can be allowed, blocked or set to need a person's approval.Example: An invoice agent may prepare a Tally voucher but may not post it.Read more

Agent

Software that uses an AI model to work through a task in steps: it reads inputs, decides what to do next, uses tools such as a database lookup, and produces a result. A good agent works within limits you set and stops for a person where needed.Read more

API

Application programming interface. A defined way for one system to ask another for data or to do something. Most modern ERPs, CRMs and helpdesks have one, and agents use them to read and write.Read more

Approval step

A point in a workflow where the agent stops and waits for a named person to accept, edit or reject its work before anything happens. Also called a human checkpoint.Read more

Audit log

A record of every run: what came in, which sources the agent used, what it produced, and who approved it. You need it to investigate mistakes and to show a regulator or auditor what happened.Read more

C

Chunking

Splitting long documents into smaller passages before they are indexed for search. Good chunking keeps a clause or a table together, so the agent finds the whole answer in one place.

Citation

A pointer from an answer to the document, record or policy it came from. Citations let a person check an answer in seconds before relying on it.

Confidence score

A number or label that says how sure the system is about a value it extracted or a decision it made. Low confidence items go to a person to check before they are used.Example: A smudged invoice total read with low confidence goes to accounts for a check.

Context window

The amount of text a model can consider at once, measured in tokens. It includes the instructions, the documents you pass in and the conversation so far. Larger windows cost more per call.

Copilot

An assistant that works next to a person. It gathers, drafts and checks, and the person decides and sends. Compare with an automation agent, which runs a task on its own within limits.Read more

D

Data Fiduciary

Under India's Digital Personal Data Protection Act, 2023, the person or organisation that decides why and how personal data is processed. If your agent processes customer or employee data, your organisation is usually the Data Fiduciary.Read more

Data residency

Where data is physically stored and processed. Some organisations need certain data to stay in India or inside their own servers, which affects which models and deployment options they can use.Read more

Document extraction

Reading a document, such as an invoice, claim form or delivery note, and pulling out specific fields into structured data. Modern extraction combines OCR with a language model and checks the result against rules.Read more

DPDP Act

India's Digital Personal Data Protection Act, 2023, with the DPDP Rules notified in November 2025 and applying in phases. It sets duties for anyone processing digital personal data, including notice, consent, security safeguards and breach reporting. Check current dates with a qualified adviser.Read more

E

Embedding

A list of numbers that represents the meaning of a piece of text. Passages with similar meaning get similar numbers, which lets a system find the right paragraph even when the question uses different words.

Evaluation

Testing an agent against a set of real past cases with known right answers, before launch and after every change. Evaluation tells you how often it is right, where it fails, and whether a change made things better or worse.Read more

F

Few shot prompting

Including a handful of worked examples in the instructions so the model sees what a good answer looks like. Often more effective than long written rules.

Fine tuning

Further training a model on your own examples so it adopts a format, style or narrow skill. Most business agents do not need it; good instructions and retrieval usually come first.

Foundation model

A large model trained on broad data that can be adapted to many tasks. The models from OpenAI, Anthropic, Google and open model makers are foundation models.

Function calling

A way for a model to ask the surrounding software to run a defined tool, such as look up a PO in Tally, and get the result back. Also called tool use. The surrounding software decides whether each call is allowed.

G

Grounding

Making an agent answer from specific sources you supply, such as your policies or your order data, so each answer can be checked against your own records.

Guardrail

Any rule that keeps an agent within safe limits: what data it may read, which actions it may take, what tone it uses, which topics it refuses, and when it must hand over to a person.Read more

H

Hallucination

When a model produces something that sounds right but is not supported by any source, such as an invented clause number or a wrong delivery date. Grounding, citations and approval steps reduce the risk; they do not remove it entirely.

Human in the loop

A design where a person reviews or approves the agent's work at chosen points. It keeps responsibility with people for decisions that carry money, legal weight or customer impact.

I

Inference

Running a trained model to get an answer. Each inference call has a cost, usually based on tokens, and a delay, usually a second to a few seconds.

K

Knowledge source

A set of documents or records an agent is allowed to search and quote, such as the HR policy folder or the product FAQ. Each source can have its own access rules.Read more

L

Large language model (LLM)

A model trained on very large amounts of text that can read, write, summarise and reason in natural language. It is the engine inside most copilots and agents.

Latency

How long it takes to get an answer. For a WhatsApp reply a few seconds is fine; for a voice call it needs to be much shorter, which affects the choice of model.

Low code and no code

Building software mostly by arranging blocks and settings in a visual editor, with little or no hand written code. Business teams can build much of an AI app this way; hard integrations still need engineers.Read more

M

Model Context Protocol (MCP)

An open standard for connecting AI models to tools and data sources in a consistent way. It makes it easier to give an agent access to a system, though access still needs the same care.

Model drift

When an agent's results change over time because the model was updated, the inputs changed, or the business rules moved on. Regular evaluation catches drift before users do.

Model provider

The company or project that supplies the AI model, such as OpenAI, Anthropic, Google, or an open model you host yourself. Providers differ in cost, speed, language quality and where data is processed.Read more

O

OCR

Optical character recognition. Turning images of text, such as scans and photos, into machine readable text. It is usually the first step of document extraction.

On premise

Running software on servers in your own building or data centre, so data does not leave your control. AksharAI can be installed this way, with models you host yourself where needed.Read more

Open weight model

A model whose trained weights are published, so you can run it on your own hardware. Useful when data must not leave your environment, at the cost of running the servers yourself.

Orchestration

Coordinating the steps of a task: which model is called, which tools run, in what order, and what happens when a step fails. In AksharAI this is designed in logic and workflows.Read more

P

Personal data

Any data about an individual who can be identified by or in relation to it. Names, phone numbers, PAN, salary and health details all count. Agents that touch personal data need notice, purpose limits and security.

Private cloud

Running the platform inside your own cloud account, such as AWS, Azure or Google Cloud, under your network rules and keys.Read more

Prompt

The instructions and context given to a model for a task. In a business agent, prompts are versioned, tested and owned like any other piece of logic.

Prompt chaining

Breaking a task into several smaller model calls, each with its own prompt, where the output of one feeds the next. Easier to test and fix than one long prompt.Read more

Prompt injection

An attack where text inside an input, such as an email or a document, tries to give the agent new instructions. Defences include separating instructions from data, limiting tools and requiring approval for actions.Read more

R

Reranking

A second pass that sorts search results by how well they answer the question, so the most useful passages reach the model first.

Retrieval augmented generation (RAG)

A pattern where the system first searches your documents for relevant passages, then gives them to the model to answer from. It is how agents answer from your policies without retraining a model.

S

Speech to text

Turning spoken audio into text, for voice notes and phone calls. Quality varies by language and accent, so test it on how your customers actually speak.

Structured output

Asking a model to answer in a fixed format, such as fields with set names and types, so the result can be checked and written into another system.

System prompt

Standing instructions that apply to every conversation with an agent: its role, its tone, its limits and what it must never do.

T

Temperature

A setting that controls how varied a model's answers are. Low temperature gives consistent answers, which suits extraction and policy questions. Higher values suit drafting and ideas.

Token

The unit models read and write in, roughly three quarters of an English word. Indian language scripts often use more tokens per word. Model pricing and context limits are counted in tokens.Read more

Trigger

The event that starts an agent: a new email, a WhatsApp message, a file in a folder, a scheduled time or a call from another system.Read more

V

Vector database

A database built to store embeddings and find the ones closest in meaning to a question. It powers search over documents in most retrieval setups.

W

Webhook

A message one system sends to another when something happens, such as a new order. Agents can start work the moment a webhook arrives, with no need to check every few minutes.