North Marketing
ESEN
Book your diagnosis

Front 2

Autonomous agents

An autonomous agent is software that converses, decides and acts. It receives the message, checks your business information, takes actions in your systems and replies, with clear limits and a person available when needed.

North team · Updated October 7, 2026 · 6 min read

24/7service with no shifts
< 5 sfirst response
es · endefault languages
100%conversations logged

What an agent can do

Each capability is a specific function or rule. You turn on the ones the business needs.

AnswerReplies with your informationquestion → RAG → answerRates, menu, policies and hours. If the data isn't in the knowledge base, it hands off.
QuoteCalculates the exact pricecotizar(dates, guests, extras)Season, nights, guests and extras come from your rate table. The function does the math, not the model.
BookSchedules in your systemconsultar_disponibilidad → crear_reservaHolds the spot for a few minutes while the person pays, so nobody else takes it.
CollectSends the payment linkgenerar_enlace_pago → payment webhookStripe or Mercado Pago. The booking confirms itself when the payment notice arrives.
QualifySeparates browsers from buyerskey questions → CRM stageDate, budget and group size. The salesperson gets the ready buyers first.
Follow upRe-engages people who didn't buyno payment after 22 h → messageA message with the context of the inquiry. It stops if the person buys, replies or asks not to get more.
RemindConfirms and reminds appointmentsstart − 24 h → utility templateWith buttons to confirm, change or cancel. The reply updates the schedule.
LanguagesServes people in the language they write indetect language → same knowledge baseThe information is loaded once. The agent replies in Spanish, English, Portuguese or French depending on who writes.
Audio and imagesUnderstands voice notes and photosaudio → transcription → agentTranscribes the voice note and replies. Reads the photo of a bank transfer receipt or of a product.
Hand offPasses to a person with a summaryrule → derivar_a_personaLarge groups, complaints and special requests. Whoever takes over sees the full history and a summary.
ReviewsAsks for the review after the visitcheckout + 3 h → link to the profileThe same request for everyone. Asking only satisfied customers for reviews goes against Google's policies.
ReportSums up the weekMonday 9 a.m. → reportConversations, bookings, handoff reasons and the questions the agent couldn't answer.

Methods

How to build an agent that doesn't make things up, acts safely and improves with use.

RAG: answering with your information, not with what it imagines

The business documents (rates, policies, menu, frequently asked questions) are split into fragments and converted into numeric vectors (embeddings). For each question, the system retrieves the most similar fragments and the model answers based on them.

Retrieval for one questionlog
question: "Do you allow pets?"

similarity  fragment
0.91        politicas.md § Pets: "We accept dogs up to 10 kg, with a $350 per night fee."
0.74        habitaciones.md § Ground floor: "The rooms with garden access…"
0.52        faq.md § Check-in: "Check-in is from 3 p.m.…"

answer: "Yes, we accept dogs up to 10 kg, with a $350 per night fee."

Recommendation. If the answer isn't in the retrieved fragments, the agent hands off to a person instead of filling in with guesses.

Function calling: actions with validation

The model doesn't write directly to your systems. It asks to run a function defined in advance, with typed parameters; the server validates and executes.

A function definitionJSON
{
  "name": "consultar_disponibilidad",
  "description": "Available rooms for a date and number of guests",
  "parameters": {
    "type": "object",
    "properties": {
      "fecha_entrada": { "type": "string", "format": "date" },
      "noches": { "type": "integer", "minimum": 1 },
      "personas": { "type": "integer", "minimum": 1 }
    },
    "required": ["fecha_entrada", "noches", "personas"]
  }
}

Recommendation. Every action with consequences (booking, charging, canceling) goes through a function with validation and is logged.

Evaluation before launch

Before serving real customers, the agent answers a bank of real anonymized conversations and every answer is measured.

Result of an evaluationlog
test bank                 100 real anonymized conversations
correct answers           96 / 100
correct handoffs          18 / 19
made-up information        0
decision                  launch with 20% of traffic and review daily

Recommendation. Launch gradually and expand traffic only when the metrics hold for a week.

Handoff rules

The cases that go to a person are defined in writing: large groups, complaints, special requests, sensitive topics. The person receives the conversation with a summary.

Recommendation. Review the handed-off conversations every week: they are the best source for improving the agent.

What we recommend in each case

If this is happeningWe recommendWhy
Lots of repeated questionsAn agent with RAG over your FAQIt handles the volume and frees up the team
Sales that require paymentAn agent with booking functions and a payment linkIt closes the sale in the same conversation
Inquiries after hours24/7 agent with handoff when you openThe first reply arrives in seconds, at any hour
International customersA bilingual agent that detects the languageIt replies in the language of whoever writes
Sensitive topics (health, legal, complaints)An agent only for sorting and handing offA person makes the decision

Strongest case studies

Agents and automations with measured results.

6xappointments per month (from 2 to 12)Lomas de Angelópolis · AI agents89%of 21,526 conversations resolved by the AIGHT Hotels · Automated inquiry follow-up96%response rateVitaldent · Automation9xmore bookings through InstagramGold's Gym · AI agent for Instagram60%of 400 inquiries answered by the agent in his last campaignGuitarras David Alcántara · AI agent for WhatsApp

Anatomy of an agent

Five layers, from the outside in. Each one can be reviewed and measured separately.

Channels
WhatsApp Business PlatformInstagram Messaging APIWebsite chat
Orchestration
WebhooksMessage queueConversation memory
Intelligence
Language modelInstructions and limitsRAG over your information
Tools
AvailabilityBookingsPayment linksCRM
Control
Handoff to a personLog of every stepWeekly review

A conversation, step by step

What happens inside when someone writes on a Sunday night. Every line is logged.

Execution log · conversation 7f3alog
23:40:02.118  input     WhatsApp · "Do you have a room for Friday and Saturday? There are two of us."
23:40:02.204  context   new customer · language: es · no history
23:40:02.611  rag       3 chunks: rates, rooms, check-in policies
23:40:03.020  function  consultar_disponibilidad(2026-11-13, 2 nights, 2 guests) → 2 options
23:40:03.940  output    "Yes. Would you prefer an ocean view or a garden view?"
23:41:15.502  input     "Ocean. We'll arrive late, around 10 pm."
23:41:16.330  function  crear_reserva(suite_mar, 2026-11-13, 2 nights, arrival 22:00) → RES-48213 pending
23:41:16.902  function  generar_enlace_pago(RES-48213) → link sent
23:44:58.077  event     pago_confirmado RES-48213 → CRM updated

AI agent, rule-based bot or person

CapabilityRule-based botAI agentPerson
Understands questions phrased in their own words
Replies in seconds, at any hour
Uses your up-to-date information
Takes actions: book, charge, record
Handles exceptional cases
Handles a hundred conversations at once

● yes · ◐ partly · ○ no. The combination that works best: the agent handles the volume and the person handles the cases that need one.

The agent dashboard

With each weekly review the agent learns from the conversations handed off and resolves more on its own.

Booking agent · 8 weeksIllustrative example
Resolved without a person71% ▲ 17 ptsweek 8
Conversion to booking14% ▲ 3 ptsof conversations
First response4 s ▼ 33%median
Satisfaction4.7 / 5 ▲ 0.4end-of-chat survey

Conversations resolved without a person (%)

W1W8

The agent's instructions

An excerpt from the configuration: what it can do, what it cannot, and when it hands the conversation to a person.

Booking agent configurationYAML
rol: booking assistant for Casa Aurora
idioma: reply in the customer's language
tono: warm, brief, no emojis
herramientas:
  - consultar_disponibilidad
  - crear_reserva
  - generar_enlace_pago
  - derivar_a_persona
límites:
  - do not confirm prices outside the rate table
  - do not promise room changes or discounts
  - groups of more than 10 people → derivar_a_persona
  - complaints → derivar_a_persona with a summary

Security and privacy

Which channel fits each job

The same agent can handle several channels. Each one has its own rules, set by Meta or by the medium.

ChannelStrong atLimit to keep in mind
WhatsAppClosing, collecting payment, reminding and following upAfter 24 hours from the customer's last message, only approved templates, which are paid per message
InstagramTurning the interest from posts, stories and comments into a conversationA single private reply per comment, within 7 days; outside the 24 hours you can't send automated messages
Website chatServing people who are already looking at pricesIf they don't leave a phone or email, the contact is lost when they close the tab
EmailDetailed confirmations, invoices and long follow-upsSlower replies; works as a backup for WhatsApp

Services in this front

Autonomous agentsAI agentsThey reply, understand and sell around the clock.Autonomous agentsAutomationRepetitive tasks run on their own and your team gets hours back.

Frequently asked questions

Does WhatsApp allow AI agents?

Yes, if they are limited to your business. Since January 15, 2026, the WhatsApp Business terms prohibit distributing general-purpose AI assistants through the platform, such as a ChatGPT inside WhatsApp (source). An agent that answers about your products, books and collects payment is allowed.

Which language model do you use?

Whichever handles each case best. We test two or three models on the same set of conversations and choose by accuracy, speed and cost per conversation. Since models change often, that evaluation lets us switch models without starting over.

What happens if someone tries to get the agent to say or do something else?

What the customer writes doesn't change the limits. Prices, discounts and bookings go through functions that validate on the server: even if the message asks for a discount, the function doesn't apply it.

Let’s start with a free diagnosis

30 minutes to see where your business is losing customers and what to do first. If you don’t see results in 30 days, we give your money back.

Book your diagnosis

We’ll message you on Instagram with times.

Keep reading

Front 1Strategy and operationsA plan with numbers, organized data, returning customers, email and partnerships.Front 3Digital positioningContent that sells, ads measured in sales, and presence on Google and in AI.