Skip to main content

​
Reusable Agent Patterns for LAC Small Business

Created by Adrian Dunkley | maestrosai.com | ceo@maestrosai.com | Fair Use
These are seven proven patterns that cover the majority of useful small-business agents in Latin America and the Caribbean in 2026. Each pattern includes the design template from design.md, the recommended build path from build.md, and concrete regional examples. Copy a pattern. Swap the language, the currency, the tools. Ship in a week.

​
Pattern index

  1. Customer-Service Triage
  2. Invoice/Receipt Extraction
  3. WhatsApp Sales Assistant
  4. Bilingual EN/ES/PT Support
  5. Content Scheduling
  6. Supplier and Price Research
  7. Weekly Business Briefing

​
Customer-Service Triage

Goal: Sort incoming WhatsApp, email, and DM messages by intent and urgency, auto-reply the easy ones, escalate the rest. Decomposition:
  1. Read message + detect language.
  2. Classify intent (inquiry, complaint, booking, billing, other).
  3. Retrieve the right FAQ answer or escalate.
  4. Send reply, or tag and route to the right person.
  5. Log the interaction.
Tools: read_message, classify_intent, lookup_faq, send_reply, notify_human, log_interaction. Memory: Session only. Add long-term memory once you want to recognize returning customers. Build path: Low-code (n8n or Make) for most SMBs. Regional example: A Santo Domingo dental clinic handles 80 WhatsApp messages per day. The agent answers pricing and hours instantly, schedules simple appointments against the clinic calendar, and escalates anything that mentions pain, emergency, or insurance to the receptionist. Typical cost: 30to30 to 60/mo.

​
Invoice/Receipt Extraction

Goal: Turn PDFs, photos, and emailed bills into structured rows in Sheets or accounting software, with variance flags. Decomposition:
  1. Watch an email inbox or a Drive folder for new documents.
  2. OCR the document (Claude/GPT vision + a fallback if the image is poor).
  3. Extract vendor, date, line items, totals, tax.
  4. Cross-check against a purchase order if one exists.
  5. Write to Sheets or QuickBooks.
  6. Flag variances > 5% for review.
Tools: watch_inbox, extract_invoice, lookup_po, write_to_sheet, flag_variance. Memory: Long-term on vendors (name normalization, past totals). Build path: Low-code or code. For volumes over 500 invoices/month, consider a small-language-model backend (see SLM section) so documents never leave your country. Regional example: A Panamanian import-export firm processes 300 supplier invoices per month in 4 languages. Agent extracts each, matches to a PO, and pushes clean rows into QuickBooks. Owner reviews only the flagged 15%. Typical cost: 40to40 to 120/mo depending on volume.

​
WhatsApp Sales Assistant

Goal: Capture leads from Instagram/WhatsApp, qualify them, quote, send payment link, book. Decomposition:
  1. Greet and detect language.
  2. Ask 3 qualifying questions (what, when, how many).
  3. Check availability / inventory.
  4. Quote in the customer’s currency.
  5. Send a payment link (Stripe, PayPal, Mercado Pago, Pagadito, etc.).
  6. On payment confirmation, add to the calendar/order system.
  7. Send a human-readable receipt and next steps.
Tools: check_inventory, quote, fx_convert, send_payment_link, confirm_order, notify_owner. Memory: Long-term on repeat customers and their preferences. Build path: Low-code for first version (n8n + WhatsApp Business API). Graduate to code if you add inventory-system hooks. Regional example: A Cartagena boutique hotel runs a WhatsApp sales assistant that handles 60 bookings per week. Quotes in COP and USD, sends a Stripe link, auto-confirms on payment, and only wakes the owner for groups of 6+ or stays longer than 10 nights. Typical cost: 40to40 to 80/mo plus payment processing.

​
Bilingual EN/ES/PT Support

Goal: One agent that serves customers in English, Spanish, and Portuguese without code-switching errors. Decomposition:
  1. Detect language from the first message (not from metadata, which lies).
  2. Set the reply language for the whole session.
  3. Use a tone appropriate to the market (e.g., Brazilian Portuguese is more informal than European).
  4. If the customer switches language, follow them.
  5. Never mix languages within a single reply.
Tools: detect_language, reply_in(language, tone), translate_attachment. Memory: Session memory anchors the language choice. Long-term memory stores customer language preference. Build path: Any. This is about prompt engineering, not infrastructure. Model choice: Claude Sonnet 4.6 and Opus 4.7 have the strongest Spanish/Portuguese in the LAC Benchmark (see rankings/lac-benchmark.md). GPT-5.4 is a close second. For Kreyòl or Papiamento, pair a frontier model with a native-speaker review step. Regional example: A São Paulo-based SaaS company’s support agent handles English (US and UK), Portuguese (BR), and Spanish (MX, CO, AR, CL). Detects language in the first line, locks tone, and escalates anything involving billing disputes to a human CS rep. Typical cost: Model usage only, 5to5 to 30/mo at SMB volumes.

​
Content Scheduling

Goal: Take a topic calendar and turn it into drafted, scheduled social posts across Instagram, TikTok, and LinkedIn. Decomposition:
  1. Read this week’s content calendar from Notion or Sheets.
  2. For each item, draft copy for each platform (different length, tone).
  3. Generate or pick a visual (Canva API, DALL-E, Gemini Nano Banana 2).
  4. Queue into Buffer or Later.
  5. Owner reviews the week in one sitting; agent adjusts based on feedback.
  6. On Friday, agent asks for next week’s topics.
Tools: read_calendar, draft_copy(platform, topic), generate_image, queue_post, request_review. Memory: Long-term on voice, hashtags, best-performing post patterns. Build path: Low-code to code. Copy the output style from your best human-written 20 posts; feed them in as examples. Regional example: A Kingston fashion brand’s content agent produces and schedules 15 posts/week in Jamaican English, with local slang calibrated to the brand’s voice. Owner spends 30 minutes on Friday reviewing, not 10 hours/week writing. Typical cost: 20to20 to 60/mo plus image generation.

​
Supplier and Price Research

Goal: Given a category (e.g., “stainless steel fittings, 2-inch, food-grade”), find 5 qualified suppliers, request quotes, compile a comparison. Decomposition:
  1. Parse the part/category spec.
  2. Search regional B2B marketplaces (MercadoLibre B2B, Alibaba, local chambers-of-commerce directories).
  3. Filter to suppliers meeting volume and certification requirements.
  4. Draft and send RFQ emails in the supplier’s language.
  5. Track replies; follow up once at 72 hours.
  6. Compile a comparison table: price, MOQ, lead time, payment terms.
  7. Hand to procurement.
Tools: search_suppliers, send_rfq, track_replies, compile_table, ask_human. Memory: Long-term on suppliers: past performance, reliability score, language. Build path: Code (Python + LangGraph or Claude Agent SDK). Regional example: A Monterrey contract manufacturer runs a supplier-research agent for every new BoM. Agent finds Mexican, US border, and LatAm suppliers, runs the RFQ process, and fills a comparison sheet within a week. Procurement reviews and picks. Typical cost: 80to80 to 200/mo depending on volume.

​
Weekly Business Briefing

Goal: Every Monday morning, a one-page email with last week’s numbers, this week’s priorities, and anything that needs the owner’s attention. Decomposition:
  1. Pull sales from POS/CRM.
  2. Pull cash position from the accounting software.
  3. Pull social reach and engagement.
  4. Read the owner’s calendar for the week ahead.
  5. Compare to prior weeks; flag notable deltas.
  6. Draft a 1-page briefing (5 sections, 50 to 80 words each).
  7. Email to the owner by 6:30am local time.
Tools: fetch_sales, fetch_cash, fetch_social, read_calendar, draft_briefing, send_email. Memory: Long-term on baseline metrics for delta detection. Build path: Low-code (n8n scheduled workflow) or code. Regional example: A Lima boutique agency’s owner gets a Spanish-language briefing every Monday at 6:30am. Highlights: 3 projects trending over budget, 2 proposals that need a decision, 1 client flagged for churn risk. He makes decisions before coffee. Typical cost: 15to15 to 30/mo.

​
Combining patterns

Real businesses often stack 2 to 3 patterns:
  • Tour operator stack: Customer-Service Triage + WhatsApp Sales Assistant + Weekly Business Briefing.
  • Manufacturer stack: Invoice Extraction + Supplier Research + Weekly Business Briefing.
  • Agency stack: Content Scheduling + Weekly Business Briefing.
Build one at a time. Prove the ROI. Then add the next.

​
Where to go next

Created by Adrian Dunkley | MaestrosAI | maestrosai.com | ceo@maestrosai.com Fair Use, Educational Resource | April 2026 SEO: AI agent patterns | patrones de agentes IA | padrões de agentes IA | WhatsApp agent Caribbean | invoice extraction agent | supplier research agent LAC