AI has reached 17% of Brazilian companies. What separates isolated tests from operational gains
The pressure to use AI arrived before clarity about where it can actually help.
In many companies, one team tests an assistant to write emails. Another uses it to summarize meetings. A third creates spreadsheets with AI-generated formulas. There is movement, curiosity, and sometimes individual productivity gains. Yet the process that consumes hours, creates rework, and delays decisions stays exactly the same.
That is what separates an isolated test from an operational gain. A tool can speed up a task, but only a well-chosen workflow changes how a company operates.
The TIC Empresas 2025 survey from Cetic.br showed that AI use rose from 13% to 17% of Brazilian companies between 2024 and 2025. The increase matters. It also raises the risk of investing in disconnected initiatives that are hard to measure and abandoned after the initial excitement.
For a mid-sized company, the useful question is not which tool to buy. It is where the operation loses time, margin, or decision-making capacity today.
AI adoption is different from operational gains
An isolated tool can create small local improvements. An employee drafts a proposal faster. A manager receives a meeting summary. An analyst organizes text with less effort.
Those uses can be helpful. The problem starts when the company treats an individual gain as operational transformation.
Operational gains happen when an entire workflow improves. There are fewer manual handoffs between teams, fewer repeated checks, less waiting for information, and more clarity about who does what.
Consider a supplier onboarding process. If data arrives by email, passes through three people, and depends on checks across different spreadsheets, an AI assistant that writes messages will not solve the bottleneck. The company still relies on email exchanges, manual validation, and scattered information.
The gain begins with more basic questions:
- Where does the process start and end?
- At which step does work stop or return for correction?
- How much time does each person spend on this workflow today?
- Which decision still depends on searching for data in several places?
- Who validates the output before it moves to the next step?
Without these answers, AI tends to become another layer on top of an already confusing operation.
What companies that capture value do differently
McKinsey's global AI research points in a consistent direction: organizations that capture value go beyond isolated use cases. They redesign workflows, establish governance practices, involve leadership, and track metrics.
For a Brazilian mid-sized operation, this means treating AI as part of a process that needs to work better, rather than as a tool demonstration.
There are four practical differences.
1. They start with a bottleneck that already has a cost
The best first process is rarely the most modern one. It is usually the one that has frustrated the operation for a long time.
Common signals include:
- a team copies the same information between systems;
- a report depends on consolidating multiple spreadsheets;
- approvals sit in email inboxes or WhatsApp conversations;
- customer service searches for client data in more than one place;
- leadership receives numbers too late to make decisions;
- the same task returns for correction repeatedly.
Rework comes before the calculation. Teams notice they are doing work twice before anyone calculates how many hours it consumes each month.
2. They measure the starting point
Without a baseline, every improvement becomes an impression.
Before implementing automation or applying AI to a workflow step, document the current state. This does not require a complex study. A simple measurement creates a useful reference:
| What to measure | Practical question |
|---|---|
| Time | How long does the process take from start to finish? |
| Volume | How many requests, registrations, or analyses go through it each week? |
| Rework | How many items return for correction? |
| Waiting | Where does work remain idle? |
| Cost | Which people spend recurring hours on this routine? |
| Quality | Which error affects customers, margin, or decisions? |
The metric needs to represent something concrete. “Reducing analysis time” is abstract. “An analysis that took two days is completed within the same shift” makes the change visible.
3. They assign someone to validate the output
AI can classify, summarize, suggest, and identify patterns. In critical processes, it does not remove the responsibility of the people who understand the operation.
Assigning a person to validate the output avoids two common extremes: releasing output without review or keeping a review so extensive that the gain disappears.
The question is simple: which decision can AI prepare, and which decision still needs human approval?
The boundary changes with risk. A commercial response draft requires a different level of validation than an analysis that affects credit, pricing, contracts, or sensitive information.
4. They adjust the workflow after implementation
Putting a solution live ends the build. It does not end the improvement work.
In the first weeks, the company needs to track where outputs still fail, which exceptions still require intervention, and whether the team has adopted the new routine. This is where the difference becomes clear between a tool that was purchased and a process that now works better.
The operation needs a clear record: who did what, when it happened, what was validated, and where the workflow stopped. That reduces dependence on memory, scattered conversations, and specific individuals.
A framework for choosing the first process
The first use case shapes how the company sees AI. If the project starts with a peripheral problem, the result feels optional. If it starts with a recurring bottleneck, it builds confidence for the next steps.
Use this framework before choosing any technology.
Step 1: List the processes that return for correction most often
Talk to operations, finance, sales, and customer service. Look for routines where someone needs to check, copy, chase, or redo work.
Do not start by asking where to use AI. Start by asking which work no one wants to repeat next week.
Step 2: Choose a workflow with clear boundaries
A good initial process has visible limits. For example: receive documents, extract information, validate fields, and send the item for approval.
Processes that are too broad make measurement difficult. “Improve sales” opens too many fronts. “Reduce the time between receiving an order and registering data for review” makes it possible to act and track results.
Step 3: Calculate the current cost without overcomplicating it
Estimate the monthly hours spent on the routine and multiply them by the cost of those hours. Add the impact of errors, delays, and lost opportunities when data is available.
The calculation does not need to be perfect at the start. It needs to be honest enough to support prioritization.
Step 4: Define a result metric
Choose one or two measures that leadership recognizes. Options include:
- total process time;
- number of items processed per person;
- percentage of items returned for correction;
- customer response time;
- weekly hours released for higher-value work;
- frequency of updates to an indicator used for decisions.
Avoid a long list. If no one can explain the result in one sentence, the metric is not clear yet.
Step 5: Map exceptions before automating
Every process has exceptions: unreadable documents, incomplete registrations, specific commercial rules, or conflicting data.
Mapping them beforehand reduces rework later. It also helps define where AI can act autonomously and where the team needs to validate.
Step 6: Choose the technology after designing the workflow
Once the workflow, metric, and exceptions are defined, the technical choice becomes more objective. Sometimes a simple integration solves the issue. In other cases, the company needs an internal system, an organized data base, or AI to interpret unstructured information.
Technology becomes the means to eliminate the identified problem. This sequence reduces the risk of buying a tool and looking for a use case afterward.
When it is still too early to automate
There are situations where waiting is the best decision.
If each person performs the same task differently, if there is no minimum approval rule, or if input data is highly incomplete, automation only speeds up inconsistency.
It also makes sense to pause when the process happens infrequently, has low volume, and does not create meaningful cost. Not every inconvenience deserves a project.
Being clear about “not yet” protects investment and increases the chance that the first move produces results when the company is ready.
AI starts with the process and ends with the result
The 17% figure from TIC Empresas 2025 shows that AI is no longer a distant topic for Brazilian companies. The next challenge is making it visible in the financial spreadsheet, response time, information quality, and time recovered by the team.
The path starts with a concrete problem: the workflow that returns for correction, information that arrives too late, approval that gets stuck between departments, or a routine that depends too heavily on one person.
Then comes the work of organizing the process, measuring the baseline, defining responsibilities, and implementing the technology that fits that reality.
If your company has manual processes, scattered data, or a team tied up in repetitive checks, we can review the situation in a free 30-minute diagnostic. The conversation helps identify whether there is a clear bottleneck to address now and what the most coherent first step would be.
Sources
- Cetic.br. Uso de Inteligência Artificial por empresas brasileiras avança e atinge 17%, aponta pesquisa do Cetic.br. TIC Empresas 2025, published June 12, 2026.
- McKinsey. The state of AI: How organizations are rewiring to capture value. Published March 12, 2025.



