AI in Brazilian companies: results, costs and limitations

An employee finishes administrative tasks faster. A bank reduces the wait to deploy models. A hospital accelerates the handling of billing disputes. These advances already appear in public accounts from Brazilian companies.

The question for a manager is what happened afterwards. Was saved time directed towards useful work? Did an expense disappear? Was quality maintained? Was the investment recovered?

This analysis starts a documentary review of four cases. It does not represent interviews conducted by NEODY or an audit of the outcomes. Information comes from supplier publications and is identified as reported results. Our editorial contribution is to compare what each indicator demonstrates and what is still missing from the financial calculation.

The initial map: four cases, four economic questions

These examples were selected because they identify an organisation, an application and an operational outcome. They are not a representative sample of the Brazilian economy and cannot establish the average gain from adopting AI.

Company	Application	Reported outcome	What it does not yet prove
Localiza&Co	Copilot in work tasks	Average of 8.3 hours saved per employee monthly, according to Microsoft	Cash savings or ROI across all licences
Itaú	Infrastructure and deployment of machine-learning models	From up to six months to three to five days in some cases, according to AWS	An isolated AI effect or return from generative agents
Bradesco	Bridge platform and support assistants	83% digital resolution and 30% lower technology costs in the reported scope, according to Microsoft	A 30% reduction in the bank’s total costs
Marista Group	Billing-denial handling at Hospital São Marcelino Champagnat	From weeks to hours, according to Google Cloud	Recovered amount, full cost or audited ROI

An initial conclusion is that outcomes become clearer when companies define a process and an indicator. A technology presentation alone does not provide the same information.

In Cetic.br’s ICT Enterprises 2025 survey, AI adoption rose from 13% in 2024 to 17% in 2025 within the survey population. The survey interviewed 4,174 companies between February 2025 and January 2026. This describes use, rather than financial returns. It also does not justify applying that percentage to companies outside the study population.

Localiza: available time needs a destination

In a 12 May 2025 publication, Microsoft reported that Localiza&Co recorded an average reduction of 8.3 monthly hours per employee on repetitive tasks using Copilot. Among frequent users, the account mentions up to 19 hours. Applications include documents, presentations, research and meeting support.

These are reported productivity figures. In this account, the publication does not provide every element NEODY would need to calculate financial returns, including full costs, the measurement sample, review effort and economic use of freed capacity.

The business question is where that time starts producing value. A team preparing reports faster may bring decisions forward. Sales teams may spend the capacity analysing customers. These are value-capture hypotheses requiring additional measurement.

For a company evaluating a similar tool, multiplying hours saved by labour cost estimates capacity rather than an automatic cash inflow. Actual savings require checking expenses avoided. Growth requires checking additional output, demand and margin.

The distribution of usage matters too. A company-wide subscription can yield concentrated results in particular groups. Tracking active users, frequency and deliverables helps determine where to expand and where implementation needs adjustment.

Question for the next reporting stage: what additional outcome was produced with the recovered capacity, and how was it measured?

Itaú: removing the queue may matter more than speeding up a task

The AWS case study on Itaú describes changes to infrastructure used by machine-learning teams. Waiting time to deploy models, previously up to six months, fell to three to five days in some cases. The account mentions more than 3,200 solution users in April 2023.

This is a case of infrastructure, standardisation and machine learning. The improvement should not be attributed solely to a generative model or presented as proof that agents replaced teams. Technical scope matters when interpreting the result.

NEODY’s reading is that a queue can cost opportunities. When a model takes months to become operational, decisions depending on it may also be delayed. Shortening that interval can make projects economically useful earlier, even without reducing every employee’s working time proportionally.

Value must be investigated in the process receiving the improvement. How many models became operational? What changed in decisions supported by them? Were infrastructure costs reduced, margins increased or measurable quality improvements achieved?

A smaller company’s manager can use the criterion without copying the bank’s architecture: identify the dependency preventing regular use. Sometimes the constraint lies in data access, approval or records rather than model capability.

Question for the next stage: which business outcomes emerged because models became available sooner?

Bradesco: the denominator determines what a result means

The Microsoft story published on 18 December 2025 describes Bridge, a platform based on Azure services and connected to BIA assistants. Its summary reports 83% resolution in digital support, 80% in internal queries and a 30% reduction in technology costs in the context described.

These percentages refer to different indicators. Digital resolution is not equivalent to satisfaction, absence of error or lower spending. The reported reduction in technology costs should not be applied to the bank’s entire budget.

Understanding the denominator has practical consequences. A resolution metric may count closed conversations, answered requests or problems actually solved. Readers need to know the criterion, period and rate of repeat contact to interpret the gain.

An operation that closes more conversations but generates more return contacts may appear efficient without improving customer experience. Resolution rates should therefore be accompanied by handling time, reopened cases, complaints and answer quality.

Cost comparisons require identifying which components were included before and after. Architecture, infrastructure and process changes may occur together. Attributing the full effect to AI without decomposing it would go beyond the available evidence.

Question for the next stage: what cost base was compared, over which period, and which share of the improvement came from each change?

Marista Group: faster billing-denial handling does not reveal recovery value

In Google Cloud’s compilation of applications, Marista Group appears using Agentspace in internal operations, starting at Hospital São Marcelino Champagnat. The supplier reports that billing-denial handling fell from weeks to hours and mentions plans to expand to other areas.

Billing denials involve disputed or rejected charges requiring analysis and handling. This account concerns an administrative process. It does not demonstrate clinical outcomes and should not be used to claim improvements in medical treatment.

Reduced time is relevant to investigating the economics of the billing cycle. Earlier identification of a discrepancy may allow staff to submit documentation sooner. Still, disputed amounts, amounts actually recovered, lead times and review effort need to be measured.

The business interpretation is to look for processes where delay and rework have identifiable consequences. Documents, contracts, billing and records may provide clearer measurement units than a broad promise to “transform the company”.

A clear indicator does not remove quality requirements. Administrative workflows must ensure that information matches the consulted document and that exceptions reach the responsible person. Accelerating an incorrect charge may merely bring a problem forward.

Question for the next stage: how much of the lead-time gain became financial recovery after implementation, review and operating costs?

Productivity, savings and ROI need different evidence

A business publication must preserve these distinctions so readers know what they can bring into an investment meeting.

Evidence	What it supports	Documentation required
Shorter lead time	The process became faster in the measured scope	Before-and-after times, volume and quality
More capacity	Staff can deliver more or free up time	Output, net hours and use of capacity
Avoided expense	Spending disappeared or declined	Comparable financial baseline and additional costs
Additional revenue	Growth was associated with the project	Attribution criteria and an appropriate comparison
Financial return	Benefits exceed the costs considered	Investment, recurring costs, benefit and period

The FinOps Foundation distinguishes resource metrics from business-outcome metrics. In this series, that becomes questions about cost per resolved request, accepted document or task delivered to standard.

Additional revenue is also not additional profit. Discounts, shipping, taxes and service costs affect margin. An application producing more proposals can increase volume without improving financial outcomes if sales activity does not find profitable demand.

When costs and results are not disclosed, the appropriate conclusion remains limited. NEODY can record an application with a reported gain and identify missing data. This creates a useful review without treating missing information as proof of success or failure.

The cases do not guarantee that the same result will recur

These public accounts have an important characteristic: participants selected them for publication. Projects that failed or produced little value may receive less attention. The collection does not establish an adoption success rate.

Company size, infrastructure, staff experience and data quality also differ. A result achieved in an institution with structured systems does not automatically represent an operation relying on spreadsheets and inconsistent records.

Experimental evidence shows effects vary by task. The NBER study found gains in assisted customer support. The 2025 METR experiment found experienced developers completing tasks more slowly with the evaluated tools. Its 2026 update reports difficulties measuring more recent outcomes. These studies concern different populations and are not a direct product comparison.

The implication is to organise a pilot with a baseline and quality criteria. Choose a workflow bounded enough to measure cost and outcome, yet representative of daily work. Include exceptions, incomplete inputs and cases requiring staff intervention.

Before expansion, check whether improvement persisted when usage became routine. A useful pilot may need more support, integration or supervision at scale. Those components belong in the continuation decision.

What NEODY will investigate next

Each new case will be examined through its previous process, implemented solution, involved systems, investment, measurement period and reported outcome. Analysis will also record difficulties, remaining human work and limitations.

The initial review suggests three areas for deeper reporting: frequent administrative tasks, document workflows involving rework and infrastructure preventing regular use of data. This is an editorial interpretation of the reviewed cases rather than an ROI ranking of Brazilian sectors.

Suppliers can contribute documentation and access to customers. User companies can explain what worked, what cost more and what required correction. Participation does not imply sponsorship or guarantee publication. Any commercial support will be identified under the editorial policy.

If your company has a documented case, send information to NEODY (Portuguese). The most useful account combines process, investment, lead time and indicators, with permission to verify the information. Sensitive data may require a publishable scope that does not identify customers.

Chapter 01, How much does AI cost?, offers a method for organising the calculation. This chapter starts identifying where to seek evidence. Future chapters will investigate industry, retail, software and work, collected on the Super Series page.

Business outcomes emerge when improvements reach a process that can be measured. That is where this series begins its investigation.

Where AI is delivering value in Brazilian companies