Skip to content
Victoria AI
Concept

An AI agent does not answer questions: it does the work.

It reads the systems where your company keeps its information, decides within the rules you set and carries out the action, leaving a record of what it did. This page covers what it is, how it decides, what it can carry out today and where its autonomy ends.

Real limits included: what it does not do out of the box, and what depends on your plan and integrations.

Definition

What an AI agent for business is.

An AI agent for business is a system that takes a request in plain language, reads the systems where that company keeps its information —ERP, CRM, calendar, product catalog—, decides what to do within a set of rules the company itself defines, and carries out the action in those same systems.

Unlike an assistant that only drafts replies, an agent has tools: it can read a price, issue a quote, create an opportunity, book a visit or register an order. Every action is recorded together with its result. When a request goes beyond its scope, or a piece of data is missing, it does not improvise: it hands the case over to a person, complete.

Three conditions, no exceptions

Connected tools

With no access to your systems, a language model can only write. An agent reads and writes where the data actually lives: the ERP, the CRM, the calendar.

Explicit rules

Autonomy is not freedom. Scope is declared action by action: what it carries out alone, what waits for approval and when it hands over.

A record of what it did

If there is no written trace of what it did, with what data and with what result, there is no way to audit it or to correct it. The record is part of the agent.

Agent and chatbot

The difference is not how it writes. It is what it can do.

Both of them reply to a message. What changes is what happens after the message, and it shows up in four situations that come up every day.

A customer asks about stock and price

A chatbot

Returns the text someone loaded earlier, or asks them to wait for a rep.

An agent

Reads stock and price from your management system and answers with today’s figure.

The customer decides to buy

A chatbot

Leaves the message on file for a person to pick up later.

An agent

Builds the quote, sends it as a PDF and leaves the opportunity in the CRM with its owner.

The data is not in any system

A chatbot

Improvises an answer or shows the menu of options again.

An agent

Says what it does know, records the case and hands it over with the full thread.

Once the conversation ends

A chatbot

You are left with a message history someone will have to read.

An agent

You are left with an updated record, the opportunity in its stage and every action logged.

An agent is still a language model: what changes the outcome is the tools it has connected, the rules you gave it and the record of what it did.

How it decides

Four steps between the message and the action.

The case changes, the sequence does not. Every request follows the same route, and at each step you can see what it read and why it decided what it decided.

  1. 01

    Identifies and puts in context

    It recognises who is writing, which channel they came from and what happened before: history, source of the contact, previous purchases and open opportunities.

  2. 02

    Reads your systems

    It asks for the data where the data lives: stock and price in the ERP, record and stage in the CRM, availability in the calendar. Never from memory or from a stale copy.

  3. 03

    Decides against your rules

    It matches the request against the scope you granted: what it can settle alone, what needs human approval and which cases it hands over without even trying.

  4. 04

    Carries it out and records it

    It performs the action in the right system and logs it with the result that system returned, so anyone can review it afterwards.

Integrations depend on what each system allows: if your ERP does not expose an operation, the agent does not invent it — it hands the case over. We map that before we start.

See how it connects to your systems
What it carries out

A catalog of actions, not of answers.

Every function is an operation against a real system. They are grouped by module: these are the ones that exist in the agent’s catalog today.

6 functions

Catalog and pricing

Searches products with price and stock, checks availability before confirming, calculates the order total, sends product photos and issues the quote as a PDF.

4 functions

Orders

Closes the order and alerts the person in charge, modifies or cancels it when the customer changes their mind, and reports on its current status.

7 functions

Calendar and bookings

Creates, moves and cancels events on the team calendar, and generates booking links so the customer picks a slot without going through anyone.

5 functions

Field visits

Logs the visit from a voice note or a text sent by the rep, edits it, lists it and deletes it, with permissions by role.

4 functions

Sales reports

Returns reports on sales, tickets and items as text or as a file, and pulls up reports already saved — all requested in plain language.

4 functions

Shipping and files

Creates the shipment with the integrated logistics provider and returns its status, and sends the customer company files: spec sheets, manuals or price lists.

On top of this come the industry functions —today, parts by licence plate or VIN, trade-in valuation and savings plans for the automotive sector—, which take the catalog past eighty. Which ones stay active in your account depends on the plan, the industry and the integrations you contract.

An example

What it looks like in an ordinary conversation.

A routine order at an auto parts shop, with closing approval set to manual. It shows what the agent does, what a person does and what gets written down.

  1. 10:42Customer

    The question arrives on WhatsApp

    “Do you have brake pads for a 2019 Polo?”

  2. 10:42Agent

    Identifies the customer

    It recognises the number, opens their CRM record and sees they have bought twice and came in through an ad.

  3. 10:42Agent

    Pins down the vehicle and checks stock

    It resolves make, model, version and year, finds the compatible part and reads price and availability from the management system.

  4. 10:43Agent

    Quotes and records

    It sends the quote as a PDF with tax broken out and leaves the opportunity created in the pipeline, with the quote attached.

  5. 10:51Customer

    The customer confirms

    “Great, I’ll take it. Can you ship it to the city centre?”

  6. 10:52Agent

    Proposes the close

    It prepares the order close and leaves it pending approval, because that action is configured as manual.

  7. 11:04Person

    A person approves

    The lead owner reviews the proposal from the CRM and confirms it. Only then is the order closed and the shipment arranged.

What was left on record
  • Lead #4821 · Martín Gómez · WhatsApp
  • Source: Meta ad · Sales pipeline
  • Carried out: stock check · quote #2093 · opportunity created
  • Approved by: Lucía Fernández · 11:04
  • Pending: arrange the shipment

An illustrative example with fictional data: it does not describe any customer. The real scope of each agent is set by the company using it.

Scope and control

Autonomy is not a switch: it is configured.

Scope is defined action by action. What the agent carries out on its own, what waits for a confirmation and what never leaves without a person looking at it.

What the agent carries out

  • Answering with data read from your systems, not with hand-written text
  • Building quotes, creating opportunities and moving leads along the pipeline
  • Booking, rescheduling and cancelling appointments, visits and reservations
  • Closing, modifying or cancelling an order, if you gave it that tool
  • Handing the case over to the right owner with everything already loaded

Approvals are configured action by action, and allow exceptions by time slot and by day of the week.

What stays in a person’s hands

  • Deciding which actions run automatically and which ask for confirmation
  • Approving or rejecting what the agent proposes, proposal by proposal
  • Authorising exceptions on price, payment terms or commercial conditions
  • Settling the cases the agent flagged as out of scope
  • Changing the rules when the business changes

When a lead has no owner yet, that rule is defined too: ask for confirmation, assign by rotation or apply directly.

Every agent action and every approval is audited inside the platform.

See security and data
Limits

What it does not do out of the box.

Almost all of this can be solved, but none of it comes solved. Saying it up front costs less than finding out after signing.

Not every function is active in every account

The set depends on the plan, the industry and the integrations you contract. Before promising a specific action we confirm it is enabled in your scope: that is why the catalog above says what exists, not what you will get.

It does not charge through every gateway

The integrated payment gateways are Mercado Pago, Payway and Rebill. If your company charges through another one, that is custom development, with its own timeline and validation.

It does not ship with every carrier

The integrated logistics provider is PuniYa: creating the shipment and checking its status. Another carrier, or your own fleet, is assessed as a custom integration.

It does not invent what your system does not expose

If the ERP does not allow an operation from the outside, the agent does not fake it: it hands the case over. What can be read and what can be written is mapped before we start, not along the way.

If your system has no API, we build the connection. The integration is done once and every agent uses it.

See the available integrations
Frequently asked questions

What we always get asked.

Next step

Tell us what you would like it to carry out.

In a 30-minute meeting we review which systems you have, which actions would make sense to automate first and which ones are better left to a person.

Victoria AI · Córdoba, Argentina

WhatsApp