Cases No 01 2026

Autoflow A debt-collection platform we built.

Built for a consumer-credit company in Brazil. The collectors had been doing the job with spreadsheets, WhatsApp, and a CRM none of them liked. We replaced all of it with one screen, and used AI for the parts of the day they wanted to do least.

portfolio today 142 aging 38 at risk 12 0–7d 8–30d 31–60d 60d+ 01 Dashboard portfolio · prioritised contract · 0428‑3 issued due +7d +30 AI · contract read next · settlement offer 02 Contract live data · AI read inbox · triage PAY NEG PAY N/R 03 Triage intent · reply · next
Fig. 01 The three screens a collector actually looks at. The dashboard tells them where to start the day, the contract view holds everything about a specific debtor, the triage view sorts inbound messages by what each one needs.

Numbers What changed

  • i ~50% of a collector's day used to be operational work. They get most of it back.
  • ii 2 days minutes Invoice issuance used to take the team two days each cycle. It runs by itself now.
  • iii ~3x Target cases per collector with the AI doing its share. We're tracking toward it.

No 01

The problem

When we started, the team's day was roughly half negotiating and half operational. The operational half was issuing boletos, copying contract numbers between systems, reading every inbound message to figure out what needed a reply, and updating statuses afterwards. Most collection teams we've worked with look like this.

The brief was to give them the operational half back.

No 02

What the AI does

There's no chatbot. The AI shows up in four places, doing four different jobs.

  1. i

    Triage

    Inbound messages on WhatsApp and SMS get a tag: paying, negotiating, ignoring, asking a question, complaining. The triage view shows them sorted by tag, and a collector can clear fifty without context-switching every other one.

  2. ii

    Reply

    When the collector opens a conversation, a suggested response is already drafted using the contract that conversation belongs to. The draft knows the instalment numbers because we pull them at draft time.

  3. iii

    Contract read

    Each contract gets a short summary: what's happened, what's overdue, what the AI thinks should happen next. This was the feature the team managers asked for first.

  4. iv

    Flows

    A no-code builder for SMS, WhatsApp, and voice campaigns. Triggered on time, on calendar, or on contract events (invoice issued, payment received, X days overdue). Managers run it without an engineer in the loop.

inbound WhatsApp · SMS AI · classify intent → lane ground live contract AI · reply suggested · sent instalments · settlement · history · status
Fig. 02 What happens when a message comes in. We tag it, pull the contract it's about, and draft a reply with that context. The collector reads the draft, edits if they want, and sends.

No 03

Where it's going

We planned the product in three steps. It ships on step one today. The other two are work for the next year and change.

  1. 01 Shipping

    Copilot

    Where we are today. The AI drafts the reply; the collector reads it, edits if they want, sends. A human is in every loop.

  2. 02 In progress

    Assisted

    The AI handles the patterns it has seen enough times (reminders, simple promise-to-pay confirmations) and asks for sign-off when there's a negotiation or a discount on the table.

  3. 03 Where it points

    Autopilot

    Most cases close without a human reading them. The exceptions queue is short enough that a person can actually work through it.

No 04

The stack

Surfaces
Dashboard, contract view, conversation triage, flow builder, team monitor.
AI
Intent classifier, reply drafter, contract summariser, decision branches inside flows, AI voice for outbound calls.
Channels
WhatsApp, SMS, and AI voice, all running through the same flow engine.
Data
The platform talks to the company's credit system in real time. Nothing the AI sees is more than a few seconds stale.
Live today
Two AI agents in production. Four automated flows running.