Glossary
AI automation, in plain words.
20 terms you will hear on any AI project, and what they actually mean for your operations.
A
- Agentic AI
- AI that takes actions toward a goal, such as updating a record or sending a reply, rather than only answering a question.
- AI agents
- AI agent
- A program that uses a language model to read a task, decide the next step and act in your tools. It asks a person when it is unsure.
- Resandhi’s Mitra
- Audit trail
- A record of everything an agent saw, decided and did, and why. What an auditor, or you, reads after the fact.
C
- Computer vision
- Models that read images. We use it to check installation photos and to score skin on a phone.
- Voltavic
E
- Evaluation
- A fixed set of real cases an agent must handle correctly before any change ships. Our answer to “how do you know it works?”
F
- Fine-tuning
- Training a model further on your own examples. Useful sometimes; usually good instructions and good data do more for less.
G
- Guardrails
- Checks that stop an agent doing something it should not. Ours are written in code, not just in the prompt, so they cannot be talked around.
- How Resandhi does it
H
- Hallucination
- When a model states something that is not true. Guardrails and grounding in your own data are how we stop it reaching a customer.
- Human in the loop
- A person approves the steps that need judgement, such as payments, refunds or a diagnosis. The agent prepares, the person decides.
I
- Intelligent document processing
- Reading invoices, forms and IDs, pulling out the fields and checking them against your records.
- Document automation
L
- Large language model (LLM)
- The model that reads and writes text, such as Claude, GPT or Llama. It is the reasoning part of an agent, not the whole agent.
M
- Maker checker
- One party prepares, another approves. In our builds the agent is often the maker and your team the checker.
- Finance automation
- MCP (Model Context Protocol)
- An open standard for connecting agents to tools and data. It lets an agent use your CRM or ERP without custom wiring each time.
- Multi-agent system
- Several specialist agents, each with one job, coordinated by a lead agent. Easier to test and more accurate than one agent doing everything.
- Multi-agent systems
O
- Orchestrator
- The lead agent in a team. It splits the work, hands parts to specialists and puts the answer back together.
- Crowta
P
- Prompt injection
- Text hidden in an email or document that tries to make an agent misbehave. Limited permissions and code-level checks defend against it.
R
- RAG (retrieval augmented generation)
- The agent looks up the relevant part of your documents before answering, so it answers from your facts, not its memory.
- RPA (robotic process automation)
- Scripts that click through screens following fixed rules. Fast, but they break when a screen changes and cannot read messy input.
- Scripts vs agents
T
- Tool calling
- How an agent acts: it asks to run a named function, such as “check stock”, and your system returns the real answer.
V
- Vector database
- A store that finds text by meaning rather than exact words. It is what lets an agent find “cotton saree” when asked for “something light for summer”.
Still not sure what you need? Ask us in plain words.
Describe the task. We will tell you which of these it needs, if any.