- Home
- AI for business
- AI agents
AI agents that act in your ERP and your store.
An agent carries out the whole task: it reads the order, checks stock, posts it to the ERP and notifies the customer. Whatever needs judgment goes back to a person, with the context ready.
An assistant answers questions. An agent finishes the work: it queries the systems, decides within rules the company sets and records what it did. Both the value and the risk sit in the integration with operations, and that is where CCX works.
How an agent works, from request to record
The model is only one piece. The rest is queries, rules, permissions and the handoff to a person when the rule does not cover the case.
- ChannelWhatsApp, email or store
- AgentReads and identifies the task
- ContextDocuments and policies
- QueriesStock, price and customer
- Rules and limitsDecide who acts
- System actionPosts to the ERP or store
- Human reviewWhat needs judgment
- Customer noticeConfirmation in the channel
- RecordEvery step kept
Reference flow. Channels, systems and limits change with each company's operation.
Where an agent tends to make sense
Tasks with volume, repeated steps and clear rules, where someone today copies data from one screen to another.
Orders and after-sales
Reads an order received by WhatsApp or email, checks stock and price, posts it to the ERP and confirms with the customer. Exchanges and returns follow company policy.
Customer service with system access
Answers about delivery, invoice copies and deadlines by querying the systems, not a script. When needed, it hands the case to the team with the history ready.
Back office
Checks invoices against orders, reconciles statements and gives the analyst only the discrepancy, instead of asking them to look for it.
Sales and catalog
Prepares proposals from price lists and history, fills in product sheets and keeps records consistent between the store and the ERP.
When an agent is not the answer
Before building, it pays to rule out the simpler options. Recommending fewer agents is part of the assessment.
What keeps an agent under control
Permissions, limits and records are part of the agent's design, along with LGPD requirements.
Minimum permissions
The agent enters the systems with its own user and can only do what that user is allowed to do.
Action limits
Value, quantity and type of operation set how far it acts alone. Beyond that, it asks for approval.
Human review
The case reaches the person with what the agent read, queried and proposed, so they can decide without redoing the work.
A record of every step
Input, query, decision and result are kept, with personal data masked, for audit and improvement.
From assessment to an agent in production
- 1
Map the task
We pick a workflow with volume and clear rules and measure how it works today: time, errors and cost.
- 2
Design the agent
We define what it does alone, what goes through a person and which systems it needs to access.
- 3
Build and evaluate
We integrate it with the systems and compare the result with the human baseline, on real cases.
- 4
Go to production
We release it gradually, with monitoring, a rollback plan and the team trained to review.
Frequently asked questions
What is an AI agent for enterprises?
It is software that uses a language model to carry out a task end to end: it understands the request, queries systems, decides within rules set by the company and records what it did. Unlike a chatbot, an agent acts in the systems instead of only answering.
What is the difference between an agent and traditional automation?
Automation follows a fixed path and stops when the input falls outside what was planned. An agent handles free text and varied cases, but it needs limits, evaluation and human review. If the process has a fixed rule and standard input, automation is the better choice.
Can an AI agent post orders to SAP or Protheus?
Technically, yes: through an API or the integration layer the ERP already uses, with a restricted user. What shapes the project is data quality and business rules. That is why the assessment looks first at the integrations between systems.
How do you stop an agent from doing something wrong?
With minimum permissions, limits on value and type of operation, validation against system rules before saving and human review of low-confidence cases. Every step is recorded, and important actions require approval or can be undone.
VTEX, Salesforce and SAP already have their own agents. Do I need another one?
Not always. Native agents, such as Agentforce and Joule, work well inside their own platform. The work usually lies in connecting them to the rest of the operation: the ERP, the store, the data and customer service. We use the native agent when it does the job and build only what is missing.
Where should an AI agent project start?
With a single high-volume task, measured before any prototype. That first workflow, delivered end to end, shows what changes in the operation and makes the case for the next ones.
Talk to an AI specialist.
Thirty minutes with a CCX engineer, at no cost, to assess whether your task calls for an agent and what needs to be ready first.