APPLIED AI · Punta del Este · Maldonado
Chatbot or automation? Choose the work before the tool.
The question is not which tool looks more advanced. It is which part of your work needs an answer, a rule or a person.
A chat is not the whole system
A conversation can be the interface to a simple task or to connected tools. Evaluating products by appearance alone misses the work underneath. At GEMOS, we propose reviewing the input, the sources, the output and the person responsible for a decision.
A useful draft may solve one task. If nobody knows where to store it, who reviews it or which version gets sent, the process still needs work. Calling a chatbot an agent does not resolve unclear responsibility.
Separate three kinds of work
Repeatable steps can use rules: check whether an inquiry has dates or assign a category selected by the visitor. Language work involves interpretation, such as summarizing a request. Commercial decisions may need a person, such as accepting a discount or confirming a condition.
Not everything needs AI. A simple, verifiable rule is a reasonable first design choice when it solves the problem. Use a model where interpretation is genuinely useful and define how its output is reviewed.
Example: from inquiry to a stay proposal
Consider a request for two adults, four nights and a surf lesson. First, prepare a record of dates, guests, interests and missing details. Next, consult authorized sources. Then prepare separate pricing items and a reply for review.
Availability should come from a calendar or the responsible operator, not from persuasive language. Prices should come from agreed rates and conditions. A missing source needs to remain visible rather than turning an estimate into a booking.
The MansaBrava demonstration uses predefined data and fictional prices to illustrate this journey. It is not evidence of live integrations, available providers or completed bookings.
Choose a narrow first implementation
Start with a frequent, bounded task that can be reviewed. Organizing incoming requests and preparing an unsent draft is one example. Before adding integrations, check whether the team actually uses the record and whether the information helps.
Define the input, output, owner and expected failure cases. What happens when dates are missing, a request is ambiguous or a source cannot be reached? A useful implementation has a stopping point. The first goal is a repeatable improvement, not complete autonomy.
Ask for a deliverable, not a label
Agree the included tools, permissions, external costs and acceptance tests in writing. Establish who controls the accounts, what documentation is handed over and what happens when an integration fails. A polished demonstration cannot replace these decisions.
GEMOS proposes a supported implementation: review the process and feasibility first, then agree a deliverable and its tests. You do not need to buy a whole platform or arrive with a complete AI plan.
TAKEAWAYDo not buy autonomy in advance. Ask for a journey you can verify.