AI microservices
AI agents that work inside your company.
Text, voice, and documents. Deployed as microservices on your cloud or your own servers.
on-prem or cloud deployment · traceability per response · integrates with your ERP
Example conversation. The user asks whether invoice FV-4471 from North Cement is already paid. The agent queries the ERP and answers: yes, paid on March 12 for $4.280.000 COP, receipt RC-9982. 2 open invoices remain for this client. Source: ERP, Accounts Receivable module, queried 1 second ago.
source: ERP · Accounts Receivable module · queried 1 s ago
Capabilities
Four capabilities, one microservice architecture.
Each tab runs on real data and tool calls: the source the agent cites, the field sent for review, the step waiting on approval.
- Connects to your systems by API: ERP, CRM, database, documents.
- Every response cites its source: system, module, and time of query.
Example conversation. The user asks how much stock there is of SKU AC-2290 and whether it's enough for tomorrow's order. The agent queries the WMS and the ERP: North Warehouse has 96 units, South Warehouse 18, total 114. Order PED-8843, for 90 units, ships tomorrow from South Warehouse and is short by 72 units. The agent proposes transferring 75 units from North Warehouse to South Warehouse today; without the transfer, the order won't ship complete tomorrow. Source: WMS Inventory module and ERP Orders module, queried 1 second ago.
source: WMS · Inventory module + ERP · Orders module · queried 1 s ago
- Connects to your systems by API: ERP, CRM, database, documents.
- Every response cites its source: system, module, and time of query.
Integrations
Connects to what your company already uses.
It doesn't replace your systems. Agents read and write directly to your ERP, your CRM, your channels, and your documents.
ERP & accounting
- SAP
- Oracle
- Dynamics 365
- Siigo
- World Office
CRM & sales
- Salesforce
- HubSpot
- Zoho
Channels
- WhatsApp Business
- Slack
- Microsoft Teams
- SIP telephony
Data & documents
- PostgreSQL
- SQL Server
- Amazon S3
- SharePoint
- Google Drive
Not on the list? If it has an API, a database, or files, it connects.
Architecture
The full path of a request. No black box.
This is how a request moves from your application to the response — and you decide where the model runs.
Where the model runs
Your application
Where the request originates: your website, your app, or WhatsApp.
API gateway
Authenticates, rate-limits, and audits every call.
Agent orchestrator
Chooses the tool and applies your policies.
Local model
Runs on GPU in your datacenter; it never leaves your network.
Connectors
Talks to your ERP, CRM, and your documents.
Observability
Every request is traced, with cost and quality.
Security
The questions in your security review.
The six that come up in every vendor assessment, with the technical answer.
Where is the model executed?
In your VPC or on your own servers, according to your data residency policy. In a local deployment, inference happens inside your perimeter: neither the prompt nor the document crosses it.
How is data protected in transit and at rest?
Both states are encrypted. Key management stays on the client side; Andes AI keeps no copy of the key material.
Can an individual response be audited?
Yes. Each response records the tools invoked, the data queried and the model version that produced it. A sample trace is shown below.
Is our data used to train models?
No. Neither conversations nor documents from a client feed the training of base models or of third-party deployments.
- request_id
- req_8f3a1c92e01d
- timestamp
- 2026-08-09T14:32:07Z
- user
- usr_4821 · rol: cartera
- tools_called
- erp.facturas.buscar, erp.pagos.consultar
- model
- andes-agent-2.4 · inferencia local
- decision
- answered — source cited
Companies already running on Andes AI
How we work
From diagnosis to production
Three stages, each with a clear condition to move to the next.
Diagnosis
1 to 2 weeksDelivers: A map of your processes, your data, and the use case with the best return.
Needs: Process documentation and a session with the team that runs it.
Pilot
4 to 6 weeksDelivers: An agent in scoped production, with success metrics agreed on.
Needs: Test access to a system and someone responsible for validating results.
Production
Full deploymentDelivers: Full deployment, SLA, monitoring, and continuous improvement.
Needs: Security and IT sign-off, and the operating channel.
In production
FREQUENTLY ASKED QUESTIONS
What gets asked before the first call.
What is an AI microservice and how does it differ from a chatbot?
An AI microservice is a service with its own versioned, observable API that performs work inside a company systems. Unlike a chatbot, it does not just answer: it queries the ERP, writes to the CRM, extracts fields from a document or triggers a workflow, and logs every step. A chatbot converses; a microservice operates.
Where does the model run, and does our data leave the network?
It is deployed in the client VPC or on their own servers, according to their data residency policy. In a local deployment, inference happens inside the client perimeter: neither the prompt nor the document crosses it. In cloud mode, the model runs on provider-managed infrastructure and data travels encrypted.
Which systems does it integrate with?
ERP and accounting (SAP, Oracle, Dynamics 365, Siigo, World Office), CRM (Salesforce, HubSpot, Zoho), channels (WhatsApp Business, Slack, Microsoft Teams, email, SIP telephony) and data and document sources (PostgreSQL, SQL Server, Amazon S3, SharePoint, Google Drive). If a system exposes an API, a database or files, it can be connected.
How long does it take to get an agent running in production?
Diagnosis takes one to two weeks and ends with a process map and the case with the best return. The pilot takes four to six weeks and leaves an agent in scoped production, with success metrics agreed before it starts. Full production depends on scope and on the client security and IT approvals.
Is our data used to train models?
No. Neither conversations nor documents from a client feed the training of base models or third-party deployments. Each tenant runs in an isolated space with its own credentials and storage, and there are no read paths between tenants.
Can we audit what the agent did on a specific response?
Yes. Each response logs the tools it invoked, the data it queried, the model version that produced it and the final decision. The agent also operates with the permissions of the user making the query: it inherits access control from the client systems and keeps no parallel access model.
Need the technical detail of each capability?
See servicesNext step
Book a demo with your use case
We'll show you how an agent solves a real process from your operation, in a 30-minute call.
30 minutes · video call