# AI-agent collections case · $2.5M recovered | Shimli > AI agents across a consumer portfolio: outreach at the optimal moment, negotiation within policy, promises recorded. $2.5M recovered in 2025. Collections ## Automated collections with AI agents: from cold receivables to confirmed payment. Autonomous collections is a process where an AI agent reaches the debtor, negotiates within the institution's policy and records the payment commitment without human intervention. Here it runs across the whole portfolio in parallel, reaching out at the optimal moment. Consumer banking and lenders · Central America The starting point The team ran the numbers before calling us. It was never a question of effort: it was arithmetic. **420** accounts assigned to each officer × **6 min** per full action: dial, wait, explain, record = **42 h** to work through the list a single time vs **40 h** in an entire working week And that list was early arrears only. Whatever went uncontacted in the first days moved to late-stage arrears, where recovery costs several times more. How it runs ### From portfolio upload to the validated payment. - 01 Portfolio intake Daily API load: days late, amount and balance. - 02 Per-customer analysis History, effective channel and response hour. - 03 Outreach and negotiation WhatsApp or call at the optimal moment. - 04 Promise recording Amount, date and reason to the CRM, with a reminder. - 05 Validation and closing OCR validates payment. Off-policy cases escalate. What it uses from Shimli ### The pieces behind automated collections. #### AI agents They negotiate within authorized extensions and escalate anything outside policy. - How it was used here - The extension table the committee approves was loaded: up to two payments, no principal forgiveness. - The agent only offers a plan if balance and days past due fall inside that table. - Outside it, it doesn’t negotiate: it hands the case to an officer with the full conversation. **CrediAmigo** Agent · actions live buscarUsuario Database Look up the customer by ID clasificarUsuario Logic Classify the credit level requisitos Data Validate approval requirements aprobacion Logic Conditions to approve credit cuotaMensual Compute Compute the monthly payment Code Interpreter Web Search #### Flow Builder The whole journey: trigger by days past due, branches by balance and action by channel. - How it was used here - The trigger is day 3 past due, not month-end. - Branches by balance: low amounts go to WhatsApp, high amounts to a call. - With no reply in 48 hours, the flow retries through the other channel. Flow canvas Start Branch Contact · 5068… Contact · 5066… Else AI Agent CrediAmigo Digital assistant · live #### AI voice A natural-voice call when the amount or the profile calls for it. - How it was used here - It only kicks in on high balances and customers who never read the WhatsApp. - It uses the same script and the same extensions as the text agent. - It’s recorded and transcribed inside the same CRM case. AI call · in progress 01:24 **Customer · Collections** +504 ···· Voice Agent A reminder of your $1,250 payment. Want to defer it? Customer Yes, in two installments please. #### Mass campaigns Reminders by arrears bracket, with delivery and reads measured in real time. - How it was used here - One send per bracket: 1 to 15 days, 16 to 30, 31 to 60. - Approved utility template, at each bracket’s peak read time. - Whoever doesn’t read it enters the one-to-one outreach flow. **2,608** Enviados **2,506** Entregados **1,782** Leídos **395** Rechazados Sent 100% Delivered 96% Read 71% Measured in production · 2025 **$ 2.5 M+** recovered in overdue receivables with AI agents · 2025 **7 ×** proven return recovered per dollar invested **$ 31 M+** in payment promises recorded and validated · 2025 **80 %** of queries resolved without human intervention The outcome ### What changes in collections operations. 01 **Portfolio coverage** The whole portfolio managed in parallel, not just the fraction the team could reach before cases went cold. 02 **Cost per action** Volume grew without adding headcount: cost stopped rising with every new customer. 03 **Traceability** Every agent decision is audited: what it analyzed, what it decided, under which policy and when. The institution's name is withheld under a confidentiality agreement. We can share a direct reference during the commercial process. The other three processes ### The agent runs the same way on every one of them. [Credit From the incomplete file to a decision with everything in view. See the case](https://www.shimliapp.com/en/casos/credito) [Support From the message queue to the reply that resolves. See the case](https://www.shimliapp.com/en/casos/atencion) [Operations From the WhatsApp order to the invoice issued. See the case](https://www.shimliapp.com/en/casos/operaciones) ### What people ask before getting started. Straight answers on what Shimli is, what it costs, what it integrates with and who is behind it. What is Shimli? – Shimli is a platform to build and run AI agents inside a company's operation. An AI agent is a program that understands the context of a conversation, queries the institution’s systems, decides within the policies defined for it and carries the task through to the end. It is built by Sheilim LLC, headquartered in Miami. What is the difference between a chatbot and an AI agent? + A chatbot replies from a script: it recognizes an intent and returns text written in advance. An AI agent understands context, analyzes information, executes tasks in the institution’s systems and escalates to a person when needed. A chatbot replies; an agent resolves. How much does Shimli cost? + One plan at $139 a month with every feature on the platform: there are no editions or tiers. It includes 3 human seats and 1 admin, 1 Flow Builder, WhatsApp, Facebook, Instagram and web connection, and unlimited WhatsApp templates. Added separately: an extra instance at $59 a month, an extra Flow Builder at $19, an extra human seat at $9, and AI usage at $49 per 2,450 queries. You can create the account and get into the platform yourself. ### Let's build this same case on your operation. We review your **collections** process, identify where the agent fits and define with you the indicators we're going to move. No commitment, nothing to install. Design my collections case