Where should a company start with artificial intelligence (AI)? With one specific process that takes up your team’s time today and follows clear rules — not by “putting AI into everything”. In practice, the companies that get results with AI start small, measure the gain and only then expand. Those that start with the technology and look for a place to fit it usually spend money without seeing a return.

In this article, I show where AI already works well in day-to-day business, the precautions you can’t skip, and a roadmap to take the first step safely.

What changed with generative AI

For a long time, automating a process required everything to be well structured: fixed fields, standardized forms, rules written in code. Anything outside the pattern — an email written in the customer’s own words, a PDF with a different layout, a voice message — still landed on someone’s desk.

Language models, like the ones behind ChatGPT, changed that. They understand free text (and, with automatic transcription, audio too), summarize, classify, extract information and answer questions in natural language. In practice, this means you can automate a good part of the work that used to depend on someone reading and interpreting.

But it’s important to understand what they are not: they are not a reliable source of facts on their own, they don’t know your company’s data unless you provide it, and they are wrong with the same confidence with which they are right. That’s why AI in business is less about “the smartest model” and more about how it is connected to your data and your processes.

Where AI already pays off

Customer service and triage

Classifying messages that arrive via WhatsApp (through the official API, so you don’t risk getting the number blocked), email or forms, understanding the subject, answering frequent questions and routing what needs human attention to the right person. It’s one of the most mature uses, and the gain shows up fast: the team stops answering the same question dozens of times a day.

Questions about company documents

Manuals, contracts, internal policies, technical catalogs. With a technique called RAG (Retrieval-Augmented Generation), the system first searches your documents for the passages that deal with the subject and hands those passages to the AI along with the question. That way, it answers based on your documents, instead of relying only on what it learned in training. The employee or customer asks in natural language and gets the answer taken from the right document — ideally, with a pointer to where it came from.

Extracting data from documents

Orders sent by email, receipts, reports in PDF or photo. The AI reads the document and fills in the fields in the system, instead of someone typing them. It works especially well with unstructured documents — electronic invoices, for example, usually come with a structured file that the system can import directly, without AI. And for fields with amounts, it’s worth keeping a check in place.

Summaries and records

Summarizing conversations, meetings, support tickets and long emails, and recording the essentials in the system. It may seem small, but it’s the kind of task nobody has time for — and that is sorely missed when someone needs to understand a customer’s history.

Analysis and decision support

Analyzing a request and suggesting what to do with it: whether an order fits a rule, whether a complaint is valid, what the priority is. Here, the safest approach is for the AI to suggest and a person to decide, at least at first.

The precautions you can’t skip

Data and privacy

Before sending any information to an AI service, you need to know what happens to it: where it’s processed, whether it’s stored and whether it can be used to train models. Personal data of customers and employees is protected by law — in Brazil, by the LGPD; in Europe, by the GDPR. The main providers offer business contracts and settings that address this — but someone needs to check. It’s also worth sending only what’s necessary (no ID numbers, for example, when they don’t change the answer) and remembering that many providers process data in other countries, which data protection laws also regulate. Watch out for free and personal versions of the tools: their data-use rules are usually different from business accounts.

Errors and “hallucinations”

Language models sometimes make up answers that look correct. The risk doesn’t go to zero, but you can reduce it a lot: have the AI answer only based on the company’s documents, instruct it to say “I don’t know” when the information isn’t there, and hand over to a person when the case is out of the ordinary. In decisions with financial or legal impact, the final word should belong to someone on the team.

Usage cost

Beyond development, AI services charge for the volume of text processed — what the AI reads and what it writes. The cost per operation is usually low, but it grows with volume. Also factor in the follow-up: reviewing answers and keeping documents up to date. It’s worth estimating the volume beforehand and monitoring it afterwards.

The user experience

An assistant that answers badly is worse than none at all. It needs to know when to hand the conversation over to a person, and that person needs to receive the context, without making the customer repeat everything.

A roadmap to start with artificial intelligence in business

Pick a process, not a tool. Look for repetitive, high-volume tasks that involve reading and interpreting text or documents. “Answering questions about delivery times” is a good start; “using AI in sales” is not.
Measure how it works today. How much time the team spends, how many requests come in, how long the customer waits. Without that number, you can’t tell whether the AI helped.
Set the AI’s limits. What it can do on its own, what it suggests for someone to approve, and when it hands over to a person.
Connect it to your data. AI is only useful when it accesses the right information: documents, the management system, the customer’s history. I explain how systems exchange information in Systems integration: how to make everything talk.
Start with a small group. One team, one type of request, one channel. Adjust based on real use.
Compare and decide. With the before and after in hand, it’s easy to decide whether to expand, adjust or stop.

It’s the same reasoning behind an MVP — I write more about it in MVP: como tirar sua ideia do papel sem gastar demais (in Portuguese).

Off-the-shelf tool or custom solution?

For individual tasks — writing a text, summarizing a document, reviewing an email — off-the-shelf AI tools already do a great job with a low investment — preferably on business accounts, without pasting customer data into personal accounts.

A custom solution starts to make sense when the AI needs to work inside your process: read data from your system, follow your rules, record the result in the right place and serve your customers under your brand. It’s the same logic as choosing between off-the-shelf and custom software, which I explain in Sistema pronto ou sob medida: como decidir (in Portuguese).

How WorkCodes works with AI

We have already put artificial intelligence to work in real products and projects: in SendZap, our WhatsApp platform, AI triages, answers and extracts information from conversations; in an ERP for a fuel distributor, it negotiates quotes over WhatsApp and reads reports from images; and, in a platform for an HR consultancy, it generates executive reports. The starting point is always the same: understand the process before choosing the technology.

If you still don’t know where to start, our technology consulting helps you find the right process and estimate the gain. When the answer is to build, we create the custom solution, starting with the essentials, with the first prototype within a few days and frequent deliveries. We work with companies remotely.

Get in touch and tell us which process takes up most of your team’s time. After an initial, no-strings conversation, you get a proposal with scope, timeline and price within 24 hours.

Artificial intelligence in business: where to start