Topic
Applied AI, responsibly
Where AI belongs in a serious process, where it does not, and how that boundary is enforced in practice.
Prompt injection: what an outside text can make your company's AI do
Any text an AI reads, whether an email, a WhatsApp message, a PDF or a transcribed voice note, can carry an instruction, and that is the first risk on the OWASP list for applications built on language models (LLM01:2025). No prompt can shield the model, because the rule and the attacker's text arrive through the same channel. The defense is to decide what the AI does on its own, what needs approval and what it never does, with the lock in code: at Balcão, the WhatsApp agent reads everything, but it has no way to start a conversation or do anything beyond replying, recording and handing off to a person.
17 min read
17 min read
AI agent on WhatsApp Business: what you pay for, what is free and what changes between the official and an unofficial API
Choosing between the official WhatsApp API and an unofficial library such as Baileys is not only about price: on the official API, Meta's rules shape the service, with a 24-hour window, approved message templates, one integration per number and per-message billing; on an unofficial one those rules disappear and what remains is the risk of the number being banned, because the use breaks WhatsApp's terms. The official API is not just the expensive option, since every number gets 1,000 free replies a month, and the unofficial one is not free. On the WhatsApp customer service platform we run, we operate both and choose client by client.
18 min read
18 min read
Code review with AI: what it catches and what slips through
AI code review is good at logic errors, side effects in other parts of the system and obvious security flaws. It does not catch business rules nobody wrote down, architecture decisions or what only the team knows about production. Across 26 pull requests in our projects, between August and September 2026, it flagged 51 problems before merge, and the largest group was wrong output on screens and reports. It only works when it reviews against the team's written standard, and the merge is always approved by a person.
15 min read
15 min read
How much it costs to put AI to work inside a company system, and what makes the bill go up or down
When AI runs inside a company system, you pay per token, and a customer service conversation costs cents: on Balcão, our WhatsApp customer service platform, triage comes to about R$ 0.15 (roughly US$ 0.03) per conversation. What weighs most is the text the AI rereads on every reply, and prompt caching bills that part at 10% of the price or less; model size comes next, with up to a 100x gap between the largest and smallest model from the same company. The bill also changes when you measure the result: in the simulation of a sales agent, the model that cost four times more brought more leads to the call, and the math by result came out in its favor.
17 min read
17 min read
When an app built on Lovable needs to leave Lovable, and when it doesn't
Leaving Lovable can mean moving the hosting, moving the database out of Lovable Cloud or no longer building with the tool, and the first two work without the third: at Sistema Forja, infrastructure went from about R$ 2,000 to about US$ 20 a month and the client kept editing the app in Lovable. If you are still validating the idea, stay where you are. Rebuilding only makes sense for one part, when search is the product or when money flows through the app.
15 min read
15 min read
How to organize company knowledge so AI can work with it
For AI to work the way your company works, you do not need to train a model. You need to organize what the company knows in a place the AI can read while it works. We use Tiago Forte's second brain method: capture everything that comes in, organize it into four folders (projects, areas, resources and archives), write each process down only once and split the work with a simple rule. AI takes the volume, which is capturing and organizing, and people keep the judgment, which is deciding and producing.
17 min read
17 min read
AI chatbot, assistant or agent: how to pick the right level for each task in your company
Chat, assistant and agent are three levels of delegation to AI. In a chat, you ask and it forgets. An assistant has learned one of your processes and helps when you call it. An agent decides on its own when to act and where to leave the result. What picks the level is how much you already trust what it delivers and what a mistake costs, not how sophisticated the task is: the most important process in our company, the project quote, runs entirely at the assistant level, and the contract that used to take up to 40 minutes to put together now takes 5.
18 min read
18 min read
How to make AI follow your company's method
Constraining a model to your company's methodology is not writing a better prompt or training your own model — it is taking out of its hands whatever cannot go wrong. Four layers: the method written outside the prompt and read at run time, the arithmetic in code, the banned list along with the vocabulary that replaces it, and a brake that stops the action instead of asking for good behaviour. In the Sistema Cuidar report that means AI writes 5 of the 15 sections and not a single number in the document came from a language model.
19 min read
19 min read
What can be salvaged from AI-generated code
What breaks in an AI-built product isn't ugly code: it's the boundary. In the app we audited, all eight server endpoints accepted unauthenticated requests — payment endpoints included — and any user's private data was readable by everyone else. None of it required a rewrite: the client kept shipping with the same AI after the foundation was fixed.
12 min read
12 min read