In Latin America the customer conversation does not start on a form. It starts in a chat. In Brazil, 8 out of 10 consumers prefer messaging to communicate with a business [1], in a country with 185 million internet users and 86.9% penetration [2]. That changes the starting point: while other markets debate the agentic era as a new channel to build, here the channel already exists, already holds the customer history, and already concentrates the commercial relationship.
The short answer: the agentic era does not arrive as a new channel, it arrives inside the conversation the customer already has. Redesigning service means trading the script for a decision layer: explicit objectives, context that travels, escalation rules written with the frontline, reversible actions, and measurement at the decision level.
Why does the agentic era land differently in Latin America?
Because you do not have to migrate the customer anywhere. They are already in the chat.
In markets where service was built on the phone-based contact center and the self-service portal, adopting agents means moving the customer to a new surface. In the region the order is reversed: the messaging thread is already the main channel, it already concentrates sales, support, and after-sales, and it already holds the relationship context. The practical consequence is that customer adoption cost is zero, so the bottleneck moves entirely inward: to the quality of the decisions your system makes inside that conversation.
That is where most teams fall short. Putting a menu bot on top of a fragmented process redesigns nothing: it automates the same maze, only faster.
What actually changes: from script to decision?
The structural shift is not about channel or model. It is about the unit of design.
For fifteen years service was designed as a journey: a predicted path for an assumed intent, with branches and decision trees written by hand. That design assumes the customer enters where you expected and wants what you anticipated. An agent breaks that assumption: it interprets context and picks the next step toward a goal instead of walking a script.
That moves human work upstream. Instead of writing every branch, the team defines what the agent optimizes, which trade-offs it accepts, where it stops, and who it hands the case to. The unit of design stops being the step and becomes the decision.
And the decision is exactly what needs governing, because it is where things break. When 1,642 execution traces of multi-agent systems were analyzed, most failures came not from model weakness but from incomplete specifications and misalignment between agents [3]. A bigger model does not fix a badly defined objective.
What does the customer expect from AI-powered service today?
Faster, without repeating themselves, and with a reason why.
Expectations rose and became concrete. FIG-1 summarizes what customers ask for today, and it reads better as a list of design requirements than as an opinion survey.
Data
| Customer expectation | Share of consumers |
|---|---|
| Want to know why AI makes its decisions | 95% |
| Expect faster responses than a year ago | 88% |
| Prefer one thread with text, image, and video | 76% |
| Expect service available 24/7 | 74% |
| Find it frustrating to repeat themselves at handoffs | 74% |
Notice the pattern: four of the five expectations are not about speed, they are about continuity. The customer assumes the system remembers who they are, what they bought, and what you promised. An agent with no memory and no access to history fails that test before it says a word.
Where is the gap almost nobody is closing?
In the explanation. Customers want to know why, and almost no company tells them.
This is the largest asymmetry of the moment, and the cheapest to fix. 95% of consumers want to know why AI decided what it decided, 80% of CX leaders acknowledge transparency will be non-negotiable, and only 37% of companies currently offer the reasoning behind an automated decision [4].
Data
| Value | |
|---|---|
| Customers who want to know why AI decided | 95% |
| Leaders who see transparency as non-negotiable | 80% |
| Companies that explain the reasoning today | 37% |
Closing that gap does not require a better model. It requires the system to record why it chose each action and to return it in plain language: "I am offering this date because your plan includes free rescheduling and Thursday has availability." That is traceability exposed to the customer, and today it is a differentiator precisely because almost nobody does it.
How do you redesign service for the agentic era?
With six decisions made on purpose, before writing the first prompt.
1. Define what the agent optimizes and which trade-offs it accepts. An agent optimizing response time behaves differently from one optimizing first-contact resolution or margin. If you do not choose, the agent chooses for you, usually whatever is easiest to measure. Write down the objective and also what you are willing to sacrifice.
2. Redesign the flow around decisions, not steps. Take the current process and mark it up: which points are real decisions and which are inherited administrative steps. The latter usually disappear. Automating a step that existed only because of a system limitation preserves the limitation.
3. Make the context travel. One thread, one customer identity, one truth about their history and their open case. This is the highest-return fix: 74% get frustrated repeating their story and 76% prefer a single thread where they can send text, image, and video without restarting [4]. If context is lost at the handoff, nothing else matters.
4. Write escalation rules with the frontline. The people on the floor know which case goes bad and on what signal. That experience is the best source for defining when the agent must stop and pass the case along. Escalation is not an agent failure: it is a design decision that executed correctly.
5. Keep actions reversible and human approval where it costs. Checking, quoting, scheduling, and explaining are safe, reversible actions. Charging, canceling, promising a commercial exception, or moving money are not. Those require approval. It is the same engineered agency principle we apply across our systems: autonomy is earned with evidence, not granted by default.
6. Measure at the decision level, not just the conversation. An average conversation satisfaction score hides where the system fails. You need to know which decision was made, on what context, whether it was right, and what happened next. Without that record, "the agent does not work" is an unfixable report.
If any of these terms are new, our agentic AI glossary covers the 20 concepts behind this design.
Where should you start?
With a narrow, high-volume flow whose actions are reversible.
The temptation is to start with the most visible case. The opposite is wiser: pick a flow where errors are cheap and repetition is high, because that is where you accumulate the evidence that later justifies widening autonomy.
What sets the order is not your industry, it is the shape of the flow. A clinic, a university, a retailer, an insurer, a bank, and a logistics operator run different processes but share the same six starting candidates, because all six combine high volume, clear rules, and actions that can be undone. Translate each one into your own vocabulary:
- Scheduling and rescheduling. Appointment, shift, service visit, or session: high frequency, clear rules, reversible action.
- Status of a request, an order, or a case. Pure lookup, near-zero risk, and usually the single largest consumer of volume.
- Qualifying and routing inbound requests. The agent gathers context and classifies; the substantive decision stays with a person.
- Reminders and due-date notices. High volume and sensitive: automate the notice, leave negotiation to a person.
- Tier-one support or after-sales. Frequent questions with real access to history, not an FAQ tree.
- Reactivating dormant contacts. Low risk, measurable, and it exercises the context layer.
Starting here is not a lack of ambition. It is how you reach the hard case with data in hand.
How do you know it actually works?
With a repeated test, not a demo that went well.
An agent that resolves one case proves nothing: systems that look strong on a single attempt collapse under repetition [5]. So we measure over repeated trials and at the decision level: resolution rate without handoff, classification accuracy, correct escalations versus late ones, and cost per resolved case. That is the dashboard that decides whether autonomy widens or narrows.
It is worth keeping the direction of travel in view. Gartner projects that by 2029 agentic AI will autonomously resolve 80% of common service issues, with a 30% reduction in operational costs [6]. Yet the same firm estimates that more than 40% of agentic AI projects will be canceled before the end of 2027, citing cost, unclear value, and weak controls [7]. Both are true at once: the destination is large and the road eliminates whoever improvises.
How MasterDragon builds this
We build AI-native, and in customer service that means starting from the real conversation, not the org chart.
We map the thread as it exists today on WhatsApp, mark the decision points, unify customer context so it travels between steps, and only then give the agent tools. On top we add the control layer: policies for what it may execute, reversible actions by default, human approval wherever money or a commercial commitment is involved, and traces of every decision so it can be explained and corrected. Autonomy widens when the measurement allows it, not when the demo impresses.
If your operation already lives in the chat and you want it to decide well, talk to our AI engineers. Start with our portfolio of shipped AI-native products.
References
- WhatsApp Business and Boston Consulting Group. (2024). Business messaging: a value driver for Brazilian businesses. https://whatsappbusiness.com/resources/resource-library/business-messaging-brazil-bcg/
- DataReportal. (2026). Digital 2026: Brazil. https://datareportal.com/reports/digital-2026-brazil
- Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Zou, J., Ramchandran, K., Sahai, A., Gonzalez, J. E., & Stoica, I. (2025). Why do multi-agent LLM systems fail? arXiv. https://doi.org/10.48550/arXiv.2503.13657
- Zendesk. (2026). Zendesk CX Trends 2026. https://cxtrends.zendesk.com/
- Yao, S., Chen, H., Yang, J., & Narasimhan, K. (2024). tau-bench: A benchmark for tool-agent-user interaction in real-world domains. arXiv. https://arxiv.org/abs/2406.12045
- Gartner. (2025, March 5). Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029. https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290
- Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027