# AI in day-to-day ERP, without giving up control.

> This is what I work on: agents, automations and tools around support, documentation, costs and integration. Sorted by the problem they solve. Perhaps you’ll recognise yours.

Deutsch: https://api-bi.dev/portfolio.md · Tiếng Việt: https://api-bi.dev/vi/portfolio.md

## One principle runs through everything

The agent does the work, a person signs it off. Nothing goes to customers, colleagues or into the system before someone has approved it. What the agent did can be traced.

## How I show what I show

Much of this came out of my professional practice. From that work I describe the pattern: which problem, which flow, which guardrails. Names, figures, code and screenshots stay where they belong. My own projects I show openly, and some you can try right away.

## Our support is drowning in the same questions

Key users and IT answer the same question for the tenth time, and the documentation falls behind with every update.

### Support that learns from every case (From practice)

**The situation:** A problem comes back and someone starts searching from scratch again. Last time’s fix sits in a closed ticket nobody can find.

1. **Question in the team chat**: Someone reports a problem in the support channel, just as before.
2. **Search with a source**: The assistant searches the ticket system and wiki, then the vendor’s documentation.
3. **Answer and follow-up**: It answers with a source and asks whether that helped. It doesn’t guess.
4. **Escalation when needed**: If it didn’t help, it opens a ticket with a description and the screenshots from the chat.
5. **Knowledge stays**: From solved cases it writes troubleshooting pages into the wiki. Where a guide is missing, a draft is created.
6. **First steps for the next case** (Next step): When a similar case comes in, it attaches the matching page straight away as a first course of action.

**Guardrails:** Answers only with a source, Escalate instead of guessing, Wiki pages as drafts for review

### Support radar (From practice)

**The situation:** Nobody sees which topics keep piling up.

A dashboard shows open cases by topic and age, so you can see where training or a guide is missing.

### Ticket sweep (From practice)

**The situation:** Tickets sit there until someone asks.

A regular run goes through stalled tickets and suggests the next step.

### Vendor docs kept in sync (From practice)

**The situation:** After every app update, the internal guide is out of date.

Once a month a run compares the documentation of the ERP extensions with your own wiki, creates what’s missing and updates what has changed.

### Wiki audit (From practice)

**The situation:** The wiki has grown for years and nobody finds anything.

An audit checks structure, naming, orphaned pages and duplicates. It fixes clear errors and proposes everything else.

## We want AI, but safely

The worry is justified: data leaks, an agent sends something out, nobody knows what it did. Guardrails help, bans don’t.

### Guardrails I use every day (Own project)

**The situation:** AI should help, but confidential names must never show up anywhere, and nothing may go out unchecked.

1. **Block list**: Confidential terms sit in a list outside the code. Every output is checked against it automatically, PDFs included.
2. **Check in the pipeline**: If the check finds a hit, nothing gets published. That applies to this website too.
3. **Approval before sending**: Emails, meeting replies and publications only go out after approval in the chat.
4. **Scheduled runs can’t change themselves**: Automated routines are not allowed to change their own rules.
5. **Goal check**: At the end of every task the agent checks whether the goal was really met, instead of just reporting “done”.

**Guardrails:** Redact, don’t hope, Approval before anything goes out, Traceable logs

### Personal AI OS (Own project)

**The situation:** Everyone on the team uses AI differently, and knowledge and rules get lost.

Skills, routines, memory and rules in one place, versioned, set up the same way on every device.

### Knowledge map (Own project)

**The situation:** Nobody has an overview of which agents, rules and documents exist.

A 3D map shows everything and how it connects. Confidential items are redacted.

### Radar with workbench (Own project)

**The situation:** News on AI and ERP arrives daily, the time to read it doesn’t.

A daily run reviews videos, newsletters and changelogs and turns them into ideas. Nothing is implemented until I give the go-ahead.

### Own MCP server (Own project)

**The situation:** AI assistants guess what a business offers.

This website is a tool for Claude and ChatGPT: you can query me there directly and send an enquiry.

Connect: https://api-bi.dev/en/contact/

## We don’t really know what it costs

Invoices arrive by email, timesheets as spreadsheets, and at month end management asks.

### Service providers in view (From practice)

**The situation:** A provider’s invoices and timesheets are scattered across the inbox.

Attachments are filed automatically, line items land in a cost sheet, and hours are analysed.

### Cost and hours dashboard (From practice)

**The situation:** The numbers exist, the picture doesn’t.

A living page shows costs by month and topic against plan and updates with every run.

### Monthly report without manual work (From practice)

**The situation:** The report eats a day every month.

The month’s data is brought together into a report for management, as a page to read and comment on.

### Cost per unit of service (From practice)

**The situation:** Internal services are charged by gut feeling.

An allocation with traceable logic, from the data basis to the presentation for the decision.

## Our systems don’t talk to each other

ERP, shop, shipping and accounting: each fine on its own, together a patchwork.

### Data flow map (From practice)

**The situation:** Nobody knows exactly which data flows from where to where.

A map shows systems, interfaces and owners. Quick to create with AI and easy to keep current.

### Build your own system map (Own project)

**The situation:** You want to know where your landscape is stuck.

Click your systems together and get a first assessment, on this website or via the MCP server.

Try it: https://api-bi.dev/en/#karte

### Onboarding as a workflow (From practice)

**The situation:** New warehouse staff need all their access on day one.

A low-code workflow sets them up in the ERP, with permissions and notifications. Not everything needs an agent.

### Server runbooks (Own project)

**The situation:** Installation, migration and upgrades depend on one person.

Runbooks with parameters instead of fixed values, a version check and a log per run, using Odoo as the example.

## We lose time on routine

Inbox, meetings, back-and-forth. None of it is hard, all of it together eats the morning.

### Inbox with approval (From practice)

**The situation:** Writing replies takes longer than deciding.

The assistant drafts replies and sends them to the chat for approval. Nothing is sent before that.

### Meeting invitations (From practice)

**The situation:** Checking, moving, declining invitations.

If the calendar is free, it accepts. If there’s a conflict, a counter-proposal comes for approval.

### Morning briefing (From practice)

**The situation:** The day starts with ten open tabs.

Meetings, open items and news on one page, before the first coffee goes cold.

### Routine cockpit (From practice)

**The situation:** Who knows which agent runs when, and whether it worked yesterday?

An overview of all scheduled runs with their latest result and hints where something is stuck.

## A living page instead of a spreadsheet attachment

I like to deliver results as an artifact: a page an agent creates and keeps current, and that people read, comment on and approve.

- **Dashboard:** Numbers at a glance, kept current by a routine
- **Cockpit:** Monitor how agents are running
- **Report:** The monthly report without manual work
- **Map:** Understand data flows and connections
- **Presentation:** A decision paper, exportable as PowerPoint or PDF
- **Script and handout:** Talk and handout from one source
- **Calculator:** Play through scenarios yourself
- **Workbench:** A tool instead of a document, with tasks for the agent
- **Guide:** User documentation you can share
- **Checklist:** Make sure quality is right before handover

## Agent, workflow or interface?

With Claude, ChatGPT and MCP, a lot can be done today without a classic workflow tool. For processes that run thousands of times the same way, scheduled and subject to audit, a fixed workflow or an interface is often cheaper and safer. Which solution fits is the real architecture question. I’ll answer it with you.

## Lab

My own products, where I try out what AI makes possible in a short time today.

### A website for AI agents

This site: static, no tracking, with llms.txt, Markdown for every page, structured data and an MCP server.

Check your own site: https://api-bi.dev/en/agent-readiness/

### Ơi

A learning platform for Vietnamese with Hanoi pronunciation, with a handbook and its own design system.

### LinguaQuest and Wortreise

Language learning as a game world you walk through, with conversations, vocabulary and levels.

### Multi-tenant calendar

The permission model first, the interface second. Who sees what is defined in the data model.

### Offer radar

Rates listings by price, trust, photos and warning signs. The same pattern checks supplier offers.

## Do you recognise your problem?

Tell me where things are stuck. I’ll tell you which building block fits, and whether it needs AI at all. https://api-bi.dev/en/contact/