Loading...
Skip to Content

PortaLonja Case Study · Custom Software and AI Integration

The system runs the work. The agent prepares it. A person decides.

How a commodity broker left WhatsApp, email and spreadsheets for its own platform, with a personal AI agent connected to the system.

A commodity broker selling South American grains and seeds worldwide ran every deal across WhatsApp, email, spreadsheets and PDFs. We built PortaLonja, a platform where each deal lives in one place, and connected every person on the team to their own AI agent, which reads, organises and prepares the work inside that system.

It is our example of both at once: custom software for the part no standard ERP covered, and AI built in where the work happens. PortaLonja is now available to other exporters: portalonja.com.

Updated: 27 October 2026 · By Andreas Walther, founder of Contrast Concepts

Before and now

Six everyday jobs.

Where a deal stands

Before: spread across chats, emails and the memory of whoever ran it. To know how it was going, you had to ask.

Now: every deal has one record with buyer, supplier, products, payment terms, shipments, documents, tasks and everything that was said, in order.

The documents

Before: invoice, bill of lading, packing list and certificates were compared by hand, field by field.

Now: the system reads every document, extracts the data and cross-checks it. If a weight, a name or a number does not match, it is flagged before the container leaves.

WhatsApp and email

Before: what was agreed on WhatsApp stayed on one person's phone, and attachments got lost in the inbox.

Now: messages, transcribed voice notes and emails with their attachments are recorded on the right deal. Whatever nobody answered shows up.

The shipments

Before: a tracking spreadsheet someone updated when they remembered.

Now: every shipment moves through stages, from booking to delivery. A stage only closes when the documents it requires are there, and the system says what is missing.

The freight

Before: quote requests emailed to each forwarder and answers compared in a spreadsheet.

Now: the request goes out prepared from the deal, quotes are recorded and compared side by side, and the chosen freight is added to the sale price.

Loading the container

Before: instructions by phone and loose photos on WhatsApp.

Now: a load order per container that the crew opens on their phone, with the steps, photos and signature recorded on the deal.

How it works inside

One central record and rules the system enforces.

One database. Companies, contacts, products, deals, shipments, documents and communications are all connected. A piece of data is entered once.

Roles and permissions in the database. Sales, logistics and finance see what is theirs. The rule lives in the database, not on the screen, so it holds the same for a person as for an agent.
Background processes. Reading emails and attachments, transcribing voice notes, extracting document data and cross-checking it all run on their own, without anyone triggering them.

A record of everything. Every change is logged with who made it and when, person or agent. So any step can be reviewed later.

The personal agent, connected to the system

It can drive almost everything. It only sends what a person approves.

Each person on the team uses their usual AI assistant, on the computer or the phone. The system connects to that assistant over MCP, an open standard, and offers it concrete tools: search, open a deal, log a note, prepare a reply.

The agent acts as that person and sees exactly what they see. There is no "superuser" account for the AI.
So instead of opening five screens, the person writes "what is missing for this week's shipment to leave?" or "tell the buyer the cargo leaves on Thursday", and the agent looks it up, gathers it and prepares it.

What it learns on the job, such as which documents each destination asks for, it keeps in a company playbook that the team can also read and correct.

Three levels of permission

  • Reading, freely. Deals, shipments, documents, emails, chats, prices, receivables: everything the person's role can see.
  • Internal changes, announced. Creating a task, logging a note, filing a message on its deal, advancing a stage. The agent says what it is about to do and the person confirms.
  • Anything that leaves or moves money, only prepared. Replies to customers, new emails, freight requests, payments and price changes become a proposal in a review tray. A person edits and sends it, or approves it.

A morning with the agent

The daily brief

The agent goes through the agenda, the unanswered messages and the shipments under way, and leaves a summary on the person's home screen.

Sorting what came in

A supplier's WhatsApp chat with no deal assigned: the agent suggests which deal it belongs to and files it once the person confirms.

Checking documents

The draft bill of lading has arrived. The agent reads it alongside the system's check and points out the field that does not match the invoice.

Preparing the reply

The agent drafts the email to the forwarder asking for the correction, in the same thread. It waits in the review tray.

A person decides

The person reads the draft, changes a sentence and sends it. The system records who prepared it and who sent it.

Questions about this case

What is MCP?
MCP (Model Context Protocol) is an open standard for connecting AI assistants to systems. The system publishes tools with clear rules, for example look up a deal or prepare a draft, and the assistant uses them on behalf of a person, with that person's permissions.
Can the agent send anything on its own?
No. Anything that leaves the company or moves money, such as an email to a customer, a payment or a price change, the agent only prepares. It waits in a review tray in the system, where a person edits it and sends it, or approves it.
What data does the agent see?
Exactly what the person using it sees, no more and no less. If that person's role cannot see margins or payments, neither can the agent. The system enforces the permissions, not the agent.
Does it only work with one particular AI assistant?
No. Because MCP is an open standard, any assistant that supports it can connect. The system does not depend on one AI provider: if another model makes more sense tomorrow, you change the assistant and the tools stay the same.
Can you build something like this for my company?
Yes, the same way of working fits other industries: one central record per deal or job, documents read and checked by the system, and an agent connected with clear rules. In Paraguay we work at USD 70 an hour or USD 560 a day, or at a fixed price if we agree one beforehand.

Where does each deal in your company live today?

Book a free 60-minute call. We’ll tell you which part of your work a system can carry and which part an agent can. No sales sequence.

Pick a topic and we will call you

We reply within one business day with a time that suits you. Your details are used to answer you and nothing else. No list, no newsletter, no third party.

Or call us directly

Office Paraguay

Call Us
+595 98 65 21 088
Nicolas Krisvoschein 384
1767 Asuncion

Office Switzerland

Call Us
+41 44 586 90 96
Bahnhofstrasse 10
CH-5630 Muri

Office Germany

Call Us
+49 36 25 96 30 70
Hinter den Höfen 6
99880 Waltershausen