No. 001—Conversation design for AI agents

Conversation designer for AI agents

I write the policies, flows and tests that decide how an LLM agent behaves with customers. Eight years designing chatbots, voice assistants and agents for banks, fintech and consumer brands across Latin America and the US.

See my work View CV
Misael Espinosa, conversation designer for AI agents, in Mexico City
Mexico City · EN / ES
CurrentlyAI Product Adoption Specialist, VaricentFebruary 2026 – present
AlsoFreelance consultant, nerds.aiHow the team designs, documents and tests agent products
PreviouslyAzumo · Stripe · PayPal · nerds.aiBanks, fintech and consumer brands, 2018 – 2026
002
About

I work in the gap between the spec and what real people type

Based in
Mexico City
Languages
English and Spanish
Since
2018, production chatbots for contact centers
Roles
Senior Conversation Designer · AI Agent Designer · Conversational AI Lead · AI Product Adoption

I started in 2018 at nerds.ai building production chatbots for contact centers in Latin America: intents, multi-turn context, error recovery, handoff to human agents, and the REST integrations behind them. I also ran customer success for those accounts, and that is where I learned to read what users do once a bot is live.

From there I designed support content and guardrails at PayPal, built agent workflows and integration documentation at Stripe, and at Azumo took conversational products for NLX and American Airlines from discovery to MVP.

Today I run the adoption program for Varicent's AI assistants. I read how enterprise customers use them and turn that into changes for the product and customer success teams. I also consult for nerds.ai on how their team designs agent products.

The part I enjoy is the gap between what a team wants the agent to do and what it does when a real person types something nobody planned for. I work in that gap: writing the spec, testing it against messy inputs, and fixing the policy until the behavior is predictable.

I'm based in Mexico City and work in English and Spanish.

003
How I work

From spec to usage, in that order

Four principles, taken from how I already work. Scroll to follow a product through them.

Stage 1/4 · Spec first
DeliverablePolicy document
UnitOne policy per job
Test assertsOutcome, not phrasing
After launchUsage reorders the roadmap
  1. 01

    Spec first.

    For agents built on LLMs, I write the policy as a document before touching any diagram. The diagram comes after, if it helps anyone.

  2. 02

    Modular policies.

    One policy per job (cancel, skip, return, track an order) instead of one long prompt that tries to do everything. Easier to test, easier to change.

    cancel_order skip_shipment return_item track_order
  3. 03

    Tests describe behavior, not wording.

    A test says what the customer should end up knowing or getting. It doesn't check exact phrasing or which tool fired.

    GivenA customer asks to cancel an order that already shipped ExpectThey end up knowing it can't be cancelled and how to return it
  4. 04

    Listen to usage.

    After launch, I read what people ask and let that reorder the roadmap.

004
Selected work

Four cases, with my part explained

Each one follows the same format: context, what I did, what stayed. Only numbers I can stand behind.

Context

At Azumo I designed conversational AI for two clients, NLX and American Airlines.

What I did

I took each product from discovery to MVP. I turned business goals into prompts, personas, tools and scoped knowledge bases, and worked with IT on roles and permissions for connector setup. I ran office hours and turned the questions that kept coming back into guides and FAQs.

What stayed

Specs the engineering team could build from, and guides and FAQs based on real questions from the client teams.

Context

Varicent, a sales performance management platform, launched AI assistants for research, incentives, plan design and sales planning. Enterprise customers had access, but many were not using them.

What I do

I run the activation program for assistant trials, from access setup to first use. I reclassified real customer prompts into 20 topics by 5 intents, so the product team can see what people ask for and where the assistants fall short. I review customer calls with account teams, triage AI product tickets, and wrote the playbook customer success managers use for first calls and low-adoption accounts. The playbook starts from the customer's goals instead of pushing usage.

What it produced

A taxonomy product managers use to read usage, a first-call playbook, and a shared page where customer feedback and requests on the assistants live.

Context

White-label chatbots for B2C brands in Latin America, from banking to food delivery and retail.

What I did

Designed conversations end to end for banking support, food ordering, a retail assistant on Messenger and a sales assistant. Standardized how the team specified bots before building them.

What stayed

A common spec format for the team's bot projects.

Context

A Latin American conversational AI company moving from button-based bots to LLM agents, with a design and integration process built for the older model.

What I did

Wrote a proposal for leadership on how the team should work on agent products: document agents as policies and SOPs instead of flow diagrams, write the spec before building, converge the designer and integrator roles into one profile, and use LLMs to draft specs and run automated QA tests.

What stayed

The team adopted the full proposal. Agent products are now documented as policies, designers and integrators work as one role, and specs and QA tests start from LLM-generated drafts.

05 Agentic Bot Diagram KitA free template for mapping how an AI agent works in four lanes: what the user sees, what the agent does, and the internal and external systems it calls. Figma Community ↗
005
Experience

Eight years of agents in production

Varicentnerds.aiAzumoStripePayPal
View CV
  1. February 2026 – present
    AI Product Adoption SpecialistVaricent

    Run the activation program for Varicent's AI assistants and turn customer prompts and calls into product and customer success decisions.

  2. 2026 – present
    Freelance consultantnerds.ai

    Advise on how the team designs, documents and tests agent products.

  3. March 2025 – February 2026
    Senior Conversation DesignerAzumo

    Designed conversational products for NLX and American Airlines from discovery to MVP.

  4. August 2024 – March 2025
    AI Operations Support SpecialistStripe

    Built internal automations and agent workflows for businesses integrating Stripe payments, and documented troubleshooting trees and API payload examples for technical onboarding.

  5. April 2022 – July 2024
    AI Content DesignerPayPal

    Designed content guardrails, structured knowledge and UX guidelines for support and in-product content across app and web.

  6. November 2018 – April 2022
    AI Conversation Design Managernerds.ai

    Built production chatbots for contact centers in Latin America and led customer success for those accounts.

006

Skills

Conversation design

Flows, policies, prompt writing, escalation to human agents, multichannel copy (chat, email, WhatsApp, voice).

Agents and LLMs

Agent policies, tool definitions, RAG content, test design for LLM behavior, prompt analysis and taxonomy.

Integration

REST APIs, webhooks, error handling and retries, data mapping, SQL queries over logs.

Tools

Figma, Notion, Jira, Confluence, n8n, Voiceflow, Dialogflow, Claude.

Languages

Spanish (native), English (professional).

Certifications

Design Thinking (Platzi), Generative AI (Coursera).

No. 007—Contact

Building an agent?

If your team is designing AI agents or conversational products and wants someone who writes the spec and stays for what happens after launch, I'd like to talk.

LinkedIn ↗ Email View CV ↗