From a WhatsApp chat to a tracked operation

One change: every operation gets a record

Right now the website ends in WhatsApp and everything after that is done by hand. The new flow keeps WhatsApp as a help channel, but saves each operation in one place so people can check it, bill it and follow it.

How an operation works now

What gets in the way

• Nothing is saved by the website. Its only result is a WhatsApp message.
• Proof of payment is a screenshot checked by eye, with no record of who decided. That is a fraud risk.
• The same data is typed twice (Dashboard, then KeyFácil or Omni), which causes mistakes.
• There is no defined data structure. The board flags it in red: “dB Schema!!!”.
• Waiting. WhatsApp queues, and the receipt can take up to 3 days.

The business already lives on the phone

Next step: a web app built for the phone first and opened from a WhatsApp link; an Expo app later, once clients come back often.

What Kambista does that we can borrow

Kambista is a Peruvian exchange app. I walked through an exchange on a real phone: everything happens inside the app, in three steps. Full walkthrough.

• Ask for the same three things: bank, destination account and source of funds.
• Show where to pay, with copy buttons and a reminder to use the right currency account.
• Ask for the operation number and proof, then show a status timeline.
• Keep saved accounts so returning clients do not retype them.

Who does what, and in which system

The whiteboard flow, drawn with its actors. Blue boxes are what we build, purple boxes are the AI chat assistant, the grey box is the Dashboard that stays as it is, and the dashed purple piece is optional AI (use the switch to hide it).

Show the original whiteboard

The validator, the operator and the billing person may be the same people; the team should confirm (see Questions).

The client’s Web App and the team’s VIEW

Client: three new ways in, and the current human route

The first three save the same data in the same database, so the team always works from one VIEW. The chat is a core way in, with the AI assistant doing the talking. It is one assistant service behind both chat channels: the WhatsApp chat and the assistant inside the Web App run the same steps and save the same data, and only the channel differs. The assistant never approves anything: the validator does. The fourth is how it works today: a human seller attends the client and negotiates the rate before the operation is made. It keeps working as it is, and the assistant can hand a client over to it at any time.

1 · Web App form

2 · WhatsApp chat

3 · Assistant inside the app

4 · Human attention (current)

This flow already works in the business: a human seller attends the client and negotiates the exchange before it is made. It is how large or negotiated amounts get done today. Example: a client who wants to change 5,000 USD, a long queue and many messages.

Team: the VIEW

The VIEW replaces searching through chats. The validator checks the photo and the bank and sets the status; the operator marks and exports; billing results appear here too. Sample rows are made up.

Where the rates come from, and what “17:18” means

Two numbers: the Brasper rate is what the client pays and is set by the business; the market reference comes from a free API and is used to check it and as a fallback.

One table for orders, one for the Fact script

The team’s board (Operaciones > Contabilidad) is already one row per order. The MVP keeps only what is needed to validate, bill and send the PDF. Everything else (discounts, second proof, channel, rate history) waits.

Left out on purpose: issuer (derived from the send currency: PEN → INGENITECH S.A.C. in S/, BRL → Brasper 21 in R$), document type and environment (Fact settings), IGV rate (18% setting), discount % (already inside the commission), second proof and channel.

From imported orders to a PDF in the client’s inbox

A person imports a set of orders into Fact. For each order, Fact registers the sale with the tax authority, saves the official XML it gets back, builds the PDF from that XML and emails it. This replaces today’s manual billing steps: re-check, register, consolidate, email.

What Fact receives for each order

One orders row per order, exactly the fields of the first table, plus fixed settings (issuer by currency, 18% tax, document type). That is the whole input. Modeled on the boleta detail the team uses today (B001-00000003, provider APISUNAT):

{
  "order_id": "BR-0003",                       // unique, idempotency key
  "created_at": "2026-10-10T17:26:00Z",        // UTC; shown in Lima time
  "client": { "name": "MICHAEL CHECASACA MUÑOZ", "doc_type": "DNI",
              "doc_number": "48602704", "email": "client@example.com" },
  "commission": 11.98, "tax": 2.16,            // total 14.14 = 11.98 + 2.16
  "send_currency": "PEN"                       // → issuer INGENITECH S.A.C., S/
}

Fact checks total = commission + tax before sending, and writes back one invoices row. The screenshot names the provider APISUNAT while the diagram says KeyFácil: confirm which one Peru uses.

If something goes wrong

The chat assistant, and one optional extra

The chat assistant is part of the core flow. The rejection explainer is optional and experimental. I kept only AI that similar companies already use, and a person always decides: AI never approves a payment or changes tax data. Most sources are company write-ups, so their claims are unaudited.

Left out because I found no similar company using them: reading voucher photos automatically, a triage assistant for exceptions, asking the database questions in plain language, and a custom fraud model.

What we can ship in four days

The normal path only: an operation from quote to PDF receipt. Special cases and AI come later.

Suggested four-day path proposal

Needed before Day 1

• A sample import and export file from the Dashboard. Its format is the contract for the VIEW and for Fact, because the Dashboard cannot be changed.
• Test access or API keys for KeyFácil and Omni, ideally returning the XML. If they have none, Fact produces the files these tools already accept: one manual upload per batch, but no retyping.
• The accountant’s tax fields and what is billed.
• A WhatsApp business account, or a test number to start with.
• Who validates, who exports and imports.

How we will know it works

Questions that block or shape the MVP

Still unclear on the whiteboard: the “e” in the script box, the red circle at the end of its line, the first letters of the Dashboard label (“DAP”), and “<Expo>”. I read “IV” as Intención de Venta and “eje” as the central hub.