Skip to main content
Menu
Take the Diagnostic
AI agents

We build AI agents that work inside your systems

AvantIT designs, builds and runs AI agents for businesses: assistants that read email, calendars, documents, your ERP and CRM — on Microsoft 365, Google Workspace, SAP, Salesforce or your own systems — apply your rules and do the follow-up work, leaving people with only the decisions.

Talk to us about an agent See the five agents

What an AI agent is (and why it is not a chatbot)

An AI agent is a system that is given a goal, looks up the data it needs, decides the steps and takes actions in real tools — filing an email, updating a spreadsheet, booking a meeting, sending a reminder. A chatbot answers questions; an agent does work.

We build agents on the platform the company already uses — Microsoft 365, Google Workspace, ERP, CRM or in-house applications through APIs — and pick the right model for each task: OpenAI, Anthropic Claude, Google Gemini or open-source models. The agent uses its user’s permissions, data stays in your infrastructure, and every action with an external effect goes through human approval until the team decides to give it more autonomy.

Where it works
Microsoft 365, Google Workspace, SAP and other ERPs, Salesforce, HubSpot, databases and in-house applications via API
Built with
Copilot Studio, Azure AI Foundry, Google Vertex AI, LangGraph, MCP and n8n — with OpenAI, Claude, Gemini or open-source models
Control
Drafts for approval, an audit log of every action, data in your infrastructure and in the EU
Who we are
Microsoft Partner headquartered in Matosinhos (Porto), working with companies across Portugal and the EU

Chatbot, automation or agent?

All three have their place. The difference is who decides the next step.

ChatbotAutomation (Power Automate, n8n)AI agent
What it does Answers questions Runs a fixed sequence of steps Pursues a goal and chooses the steps
Handles exceptions No Only those built into the flow Yes: asks for what is missing or escalates to the right person
Reads unstructured context Partly No Yes: emails, minutes, Word and Excel documents
Example “What is the holiday policy?” “When an invoice arrives, file it in the right folder” “Prepare the month-end close and tell me what looks off”

Five agents we are designing for management, finance, operations and quality teams

Every agent has an owner, defined data sources, rules written by the business and success criteria measured from day one. These are the five we are designing now — and they are a starting point for others.

01

Executive Assistant Agent

For: Board members, directors and managers with overloaded inboxes and calendars.

Goal: Take inbox, calendar and follow-up work off the manager, leaving only the decisions.

Data sources

  • Email and calendar (Outlook or Gmail)
  • Chats and channels (Teams, Slack or Google Chat)
  • Documents (SharePoint, OneDrive or Google Drive)
  • Meeting transcripts (Teams, Google Meet or Zoom)

What the agent does

  • Email triage and filing by folder and priority, following the company’s rules
  • Morning briefing with the day’s agenda, critical emails, open items and deadlines
  • Meeting preparation with context, history and relevant documents
  • A running list of commitments and open items taken from emails and meetings
  • Alerts on deadlines at risk and replies still owed
  • Reply drafts in the manager’s voice, for approval
  • Calendar management: booking, moving and cancelling meetings
  • Follow-ups with third parties and a weekly summary

How we measure success

  • Hours saved per week
  • Share of drafts approved without edits
  • No critical email left unanswered

Technology

  • Microsoft Graph or Gmail API
  • Copilot Studio or LangGraph
  • OpenAI / Claude
  • Outlook or Gmail

02

Financial Controlling Agent

For: Finance departments and management-control teams, especially at month-end close.

Goal: Keep controlling files up to date, consistent and distributed on time, and flag anything outside the expected range.

Data sources

  • Excel or Google Sheets files, including those linked to other files or data sources
  • ERP extracts: SAP (CO/FI), Primavera, PHC or others
  • Emails with data sent in by departments
  • Distribution lists and the month-end close calendar

What the agent does

  • Checks on incoming files: structure, periods, totals, broken formulas and links
  • Refreshing linked Excel files
  • Reconciling values between files and sources, with a list of differences
  • Variance analysis against budget and prior period, with deviations above a threshold, trends and an analysis note
  • Reminders to departments about missing data and deadlines
  • A checklist of what has arrived and what is missing, with a log of checks and changes
  • Distributing reports once validated
  • At each month-end: updated files, variance analysis and the report sent out

How we measure success

  • Days and hours saved per close cycle
  • Errors caught before distribution, not after
  • Reports delivered on time

Technology

  • Azure AI Foundry or Vertex AI
  • Python
  • Excel / Google Sheets
  • SAP and ERP connectors
  • n8n or Power Automate

03

Team Operations Agent

For: Department heads and teams with recurring meetings and shared action plans.

Goal: Make sure every meeting ends with minutes, decisions and assigned actions, and that no action is left without an owner or follow-up.

Data sources

  • Calendar and meeting transcripts (Teams, Google Meet or Zoom)
  • Team chats, channels and emails
  • Shared documents (SharePoint, Google Drive or another repository)
  • The team action plan (Excel, Planner, Lists, Jira, Asana or monday.com) and list of owners

What the agent does

  • Minutes for every meeting with topics, decisions and actions with owner and due date, sent to attendees
  • Updating the action plan: new actions, due dates, status and completions
  • The next meeting’s agenda with the open items
  • Reminders to owners before due dates and escalation of overdue actions
  • A weekly summary of completed, in-progress and overdue actions, and blockers

How we measure success

  • Minutes delivered on time
  • Share of actions with an owner and due date
  • Actions completed on time
  • Time spent per set of minutes

Technology

  • Teams, Meet or Zoom
  • Planner, Jira or monday.com
  • MCP
  • n8n or Power Automate

04

Capex Management Agent

For: Investment committees, controlling and finance teams that review Capex requests.

Goal: Follow every Capex request from Project Summary to close, ensuring complete and compliant documentation, critical review and follow-through on actions.

Data sources

  • The Project Summary in Word and attachments in the document repository
  • Sponsors’ emails and Teams
  • Calendar and transcripts of Capex meetings
  • The Capex procedure, template and compliance criteria
  • The action plan and request register

What the agent does

  • Compliance checks on each request (mandatory sections, attachments, consistency of figures) and completing the checklist
  • Critical review of assumptions, return, strategic fit and risks, in a fixed-structure draft note for review
  • Requesting missing information from the sponsor, with a deadline
  • Meeting minutes, action plans, follow-up and sending opinions
  • Escalating delays
  • A weekly Capex portfolio dashboard: requests, status, amount and delays

How we measure success

  • Review time per request
  • Share of requests returned for missing information
  • Time from submission to decision
  • Nothing missed that a human review would have caught

Technology

  • RAG over the Capex procedure
  • Claude / GPT
  • Azure AI Foundry or Vertex AI
  • Power BI

05

Quality Operations Agent

For: Quality and production departments in manufacturing, running certified quality management systems.

Goal: Continuously watch system data and quality manuals, apply defined rules and alert the right person, with context, when something is off.

Data sources

  • Quality and production systems: process data, inspection, complaints, non-conformities and audits
  • Quality management system manuals and procedures
  • Work instructions and specifications
  • Emails and messages (Teams, Slack) about incidents and complaints

What the agent does

  • Ingesting and normalising data from each source on a fixed schedule
  • Applying rules (thresholds, trends, tolerances) to detect deviations
  • Matching each deviation to the applicable manual, citing the procedure and the prescribed action
  • Looking for patterns by line, product, shift or supplier
  • Alerts classified by severity, with the supporting data, in an auditable log
  • A daily exceptions report and a weekly trends report

How we measure success

  • Alert precision (relevant versus false positives)
  • No real deviation missed
  • Time from occurrence to alert
  • Hours saved on manual analysis

Technology

  • RAG over the quality manuals
  • Python
  • Fabric, BigQuery or SQL
  • Alerts via Teams, Slack or email

Other agents we build on the same foundation

The pattern repeats in any process with clear rules, lots of documents and lots of follow-up.

Customer support and internal helpdesk

Answers from company documentation, cites its sources and escalates what it does not know.

Onboarding and HR

Welcomes new hires, answers policy questions and tracks onboarding tasks.

Approvals and documents

Reads requests and invoices, checks them against policy and prepares the approval in Teams, Slack or email.

How we build an agent

We start small, measure, and only then give the agent more autonomy.

  1. 01

    Discovery

    We choose the process, write the rules with the team, confirm the data sources and fix the success metrics.

  2. 02

    Pilot

    One agent, a group of real users and “draft for approval” mode switched on. We measure the agreed metrics from week one.

  3. 03

    Production

    Security, permissions, audit logging, monitoring and training. The agent becomes part of day-to-day operations.

  4. 04

    Run and improve

    We review the metrics, refine the rules and widen autonomy — or add the next agent.

Security, control and compliance

Human in the loop

Emails, opinions and reports go out as drafts for approval. Autonomy only grows when the numbers justify it.

User permissions

Through Microsoft Graph, Google Workspace APIs or each system’s OAuth, the agent only sees what its user can already see.

Data in your infrastructure

Agents run in your cloud (Azure, Google Cloud or AWS) or on your servers, with EU data residency and no model training on your data.

Audit log

Every check, alert and action is logged: what, when, with which data and who approved it.

EU AI Act and GDPR

We classify each agent’s risk, document human oversight and apply GDPR-aligned retention policies.

No model lock-in

Model-agnostic architecture: we switch models — OpenAI, Claude, Gemini or open-source — when another is better or cheaper, without rewriting the agent.

Technology

We use the platform the company already has — Microsoft, Google or other — and pick the right model and framework for each agent.

  • Microsoft Copilot Studio
  • Azure AI Foundry and Azure OpenAI
  • OpenAI
  • Anthropic Claude
  • Google Gemini and Vertex AI
  • Amazon Bedrock
  • Open-source models (Llama, Mistral)
  • LangGraph
  • Semantic Kernel
  • Model Context Protocol (MCP)
  • n8n
  • Power Automate
  • Microsoft Graph and Google Workspace APIs
  • SAP, Primavera and PHC
  • Salesforce, HubSpot and Dynamics 365
  • Python and .NET

Frequently Asked Questions

Get answers about this service

It identifies the processes where an agent saves the most time, writes the rules the agent must follow together with the team, connects it to the data sources (email, Teams, SharePoint, Excel, ERP), builds it, measures it and keeps it running. AvantIT covers that whole path, on Microsoft 365, Google Workspace, ERP, CRM or in-house applications.
No. Microsoft is our specialty, but we build agents for Google Workspace, ERPs (SAP, Primavera, PHC), CRMs (Salesforce, HubSpot, Dynamics), databases and in-house applications through APIs or the Model Context Protocol (MCP). We choose the platform and the AI model to fit the process, not the other way round.
Microsoft 365 Copilot, ChatGPT or Gemini are general-purpose assistants: they help each person write, summarise and search. A custom agent knows one specific process — your month-end close, your Capex procedure, your quality manual — and runs it with your rules, connected to your systems. The two often work together: the agent is published inside Copilot, Teams or Slack.
Only if the company decides so. By default anything that goes outside — replies, follow-ups, opinions, reports — stays as a draft for approval. Internal, low-risk actions such as filing email or updating an action plan can be automatic from the start.
Yes. The agent can read SAP extracts (for example CO/FI) or another ERP, connect through APIs or connectors (Power Platform, n8n), and it works with Excel or Google Sheets files — including checking for broken formulas and links before refreshing values.
Data stays in the company’s infrastructure — Microsoft 365 and Azure, Google Workspace and Google Cloud, AWS or its own servers — with EU data residency. The agent respects existing permissions: it only accesses what its user can see. Data is not used to train models.
It depends on the process and the data sources. The path is always the same: a short discovery to fix rules and metrics, a pilot with real users, and only then production. A first agent is typically in production in 6 to 8 weeks — our fixed-scope “First AI agent” offer.
It depends on the platform and the agent. On Microsoft, some agents run on Copilot Studio, licensed by capacity, without Microsoft 365 Copilot for every user. Custom agents pay for the model on consumption (Azure OpenAI, Anthropic, Google or open-source models hosted by the company). We do the maths during discovery and pick the most cost-effective option for the expected volume.
By classifying each agent’s risk, documenting its purpose and the data it uses, ensuring human oversight on decisions with impact and keeping a log of its actions. The productivity agents on this page do not make decisions about people and generally fall outside the high-risk categories.
AvantIT is a Microsoft Partner headquartered in Matosinhos, Greater Porto, Portugal. We work with SMBs and mid-sized companies across Portugal — Porto, Lisbon, the Algarve — and the European Union, in Portuguese and English.

Further reading

What would your team’s first agent be?

In a 30-minute call we pick the process with the best return, confirm the data sources and tell you how we would measure success.

Talk to us about an agent Quick Diagnosis