Artificial intelligence in day-to-day management

Monday, eight in the morning. A project manager opens the summary that the AI tool prepared overnight: forty supplier contracts condensed into three pages of flawless prose. On page two, the structural works contract shows an annual price adjustment. The original says every six months. Nobody notices until the first progress invoice arrives with a figure that does not add up.

The scene is illustrative, but the mechanism is real. The summary was plausible, and plausible is not the same as true. The summary proves speed. Only review proves trust.

The subject has reached lawmakers. Regulation (EU) 2024/1689, signed on 13 June 2024, sets 2 August 2026 as the general date of application of the European AI rules. In Brazil, Bill PL 2338/2023, which "provides for the use of Artificial Intelligence", was approved by the Federal Senate on 10 December 2024. It went to the Chamber of Deputies on 17 March 2025. The question is no longer "does AI work?" but "who answers for what it produces?".

Our answer: AI applied to management pays off when it becomes a team of Consultants specialised by discipline, each with a defined scope, sources and limits. And when every relevant conclusion passes through a human specialist before it becomes a decision.

The gain comes from splitting the work by discipline

A generic assistant answers everything with the same confidence, informed or not. A Consultant dedicated to one discipline works differently. Document checking, schedule tracking and contract reading each get a bounded set of sources, their own instructions and a known exit criterion. The model underneath may be identical. What changes is the architecture around it.

An engineering guide published by Anthropic in December 2024 separates two designs. In workflows, models and tools follow predefined code paths. In autonomous systems, the model chooses its own path and tools. The guide advises starting simple and adding complexity only when it demonstrably improves results. Autonomy costs latency and money, and errors can compound from one step to the next. That is why the guide calls for extensive testing in sandboxed environments and appropriate guardrails.

Without a split by discipline, the same tool reads a land registry record, a construction contract and a cash flow sheet without telling what is critical in each. With the split, each Consultant knows what to look for, what to ignore and when to stop and ask for review. The gain is precision before it is speed.

AI can only organise what already arrives organised

Construction learned this lesson first. Brazil's Decree 10,306/2020 defines BIM as a "set of integrated technologies and processes" for creating and updating digital models of a building "collaboratively". Since 1 January 2021, on federal projects deemed highly relevant, the decree has required, among other uses, the detection of physical and functional clashes between the different disciplines. The value did not come from software, but from the duty to structure each discipline's information in a shared model.

AI applied to management follows the same path. It reconciles decisions only when it finds indexed contracts, versioned schedules and minutes with an owner and a deadline. Where information lives in inboxes and loose spreadsheets, it produces elegant summaries of an incomplete picture. Method first, tool second.

This work has a name: context engineering. It means giving each Consultant the sources authorised for the task, the company's vocabulary, the decisions already made and those that cannot be reopened. A company that does not document how it decides has nothing to give an AI beyond loose files. One that does gains twice: new staff learn faster and the Consultant makes fewer mistakes.

Controlling context also means restricting it. The schedule Consultant does not need buyers' personal data. Minimum scope reduces both error and exposure.

AI does not remove responsibility for the decision. It makes that responsibility more visible, because every conclusion now carries a trail, a source and a reviewer.

106

Review is a legal requirement, not a courtesy

Brazil's data protection law, the LGPD, lets data subjects request review of decisions based solely on automated processing that affect their interests (Article 20). The rule covers decisions that define their professional, consumer and credit profiles. On request, the controller must provide clear and adequate information on the criteria and procedures used. If a manager uses AI to screen buyers' credit or assess individual contractors, review and explanation belong in the design from day one.

The European rule points the same way for high-risk systems. It requires that natural persons can effectively oversee them and that the overseer stays aware of the tendency to over-rely on the output, known as automation bias. It also requires that the overseer can decide not to use, disregard, override or reverse the system's output. It does not directly reach a Brazilian operation with no link to the European Union, but it describes a universal risk.

The mechanism is easy to grasp and hard to sustain. When AI is right most of the time, the reviewer relaxes. Once the reviewer relaxes, the rare error slips through. Effective review demands deliberate sampling, written criteria for what to check and real authority to send the work back. A review that cannot refuse is a rubber stamp, and a rubber stamp protects neither the company nor the data subject.

AI governance already has a public framework

There is no need to wait for a Brazilian AI law to govern the technology. The NIST AI Risk Management Framework, released on 26 January 2023 for voluntary use, organises risk management into four functions: govern, map, measure and manage. On 26 July 2024 the institute published a profile specific to generative AI. Its official page states that the core framework is being revised.

In Brazil, the National Data Protection Authority (ANPD) runs a regulatory sandbox for artificial intelligence with three companies selected through a public call. According to the first partial report, from July 2026, the next cycles will go deeper into digital security, transparency, data governance, anonymisation and evidence production. That is the agenda the regulator is already watching.

Another European point matters outside Europe. Article 4 of the regulation asks providers and deployers to ensure a sufficient level of AI literacy among the people operating these systems. In short, the team must know what the tool does well and what it does badly. In InnovaSphere's reading, this is the highest-return, lowest-cost measure in any AI programme. A team that understands the tool reviews better and trusts it to the right degree.

The log turns AI into an accountability tool

The LGPD lists accountability among its principles. Adopting measures is not enough: a company must show they work. Article 37 supplies the instrument by requiring controllers and processors to keep records of the personal data processing they carry out. Applied to AI, that means keeping what went in, what came out, which version of the instructions was in use and who reviewed it.

The log has a second, less obvious use: it measures the AI itself. With a trail of inputs and outputs, you can compare the Consultant's conclusion with the specialist's final decision. Using the operation's own data, you learn in which discipline it is reliable and in which it needs more context or less autonomy. Without a trail, perceived quality becomes opinion. And opinion cannot support a decision to expand or pull back.

Numerical example: from the simulator's ceiling to the net gain

The simulator that accompanies this article takes two inputs, active Consultants and monthly hours replaced. It returns a ceiling: the cost of doing by hand what AI would take over, minus the cost of the Consultants. The example below adds what the ceiling leaves out. All figures are InnovaSphere hypotheses, not market data. Amounts are in Brazilian reais.

The ceiling suggests R$ 56,000 a month. Net of the hours AI does not absorb and of review time, the gain is R$ 23,000. That is 41% of the ceiling, with simple payback on implementation in three months, undiscounted because the period is short.

Now the stress test. If the absorbed share falls from 60% to 30%, all else equal, the net gain drops to R$ 9,500 a month and payback stretches beyond seven months. If review consumes half of the absorbed hours rather than a quarter, the gain falls to R$ 14,000 and payback moves to about five months.

Monthly break-even sits near 9% absorption: below that, the Consultants cost more than they save. The variable that weighs most is not the price of AI. It is how much of the mapped work AI actually takes on, with quality.

When it does not pay off

The same arithmetic that shows the gain also shows the limits. Three situations call for caution or delay.

• Low volume, or a task that changes every time. If repetitive hours are few, or each case demands fresh reasoning, mapping, configuring and reviewing cost more than the time saved. AI pays off on recurring work with stable criteria.
• Disorganised information. Without indexed documents, controlled versions and named owners, AI amplifies disorder while making it look orderly. Organise the information first.
• A decision that requires an explanation the company cannot give. When the output affects people, such as buyers' credit or contractor selection, and the company cannot disclose criteria or guarantee effective human review, automation creates more risk than savings.

Checklist for deploying AI Consultants in management

1. Map repetitive tasks by discipline, with real measured hours, together with each area's managers.
2. Classify the data each Consultant will read (personal, confidential or public) with the data protection officer.
3. Define each Consultant's scope, authorised sources and exit criterion with the discipline's specialist.
4. Establish who reviews, at what sampling rate and with authority to send work back, together with the area's leadership.
5. Record inputs, outputs, instruction version and reviewer for every task, with the technology team.
6. Measure accuracy every month, comparing the Consultant's conclusion with the final decision, with the operations lead.
7. Review contracts with AI vendors (data use and retention, confidentiality and security) with the legal team.

Three companies in the group work on adjacent ground. IncorpBuilding, the development engineering arm, operates with a native digital twin and digital observability of the works. InnovaPeople works on organisational design and context engineering. Valida Auditoria provides independent audit and technical validation under the motto "we validate reality, not the narrative".

Frequently asked questions

Does AI replace the management team?

It reorganises the work. Absorbed hours move from checking and compiling to review, decision and relationships. Human review is part of the cost, not a residue to eliminate.

Should we wait for Brazil's AI law before starting?

No. Bill PL 2338/2023 left the Senate in March 2025, and the Chamber of Deputies tracking record showed it "awaiting an opinion" in September 2026. The LGPD, however, already applies to any processing of personal data carried out with AI. The NIST framework lets you structure governance now.

Which Consultant should come first?

The one that combines high volume, low risk and clear criteria: document checking, minutes and open items, triaging technical correspondence. There, errors surface in review and savings show up fast.

How do we stop AI from making things up?

By restricting authorised sources, requiring every conclusion to state where it came from, separating what was verified from what was inferred, and reviewing by sampling. None of these measures removes the risk. Together, they make errors visible before they become decisions.

Is company data exposed to the AI vendor?

That depends on the contract and the configuration. Article 46 of the LGPD requires technical and administrative security measures able to protect personal data from unauthorised access. Minimum scope per Consultant, clear clauses on data use and retention, and records of processing are the starting point.

Legal basis

• Brazilian Law 13,709/2018 (LGPD): Art. 6, X (accountability); Art. 20, caput and para. 1 (review of automated decisions); Art. 37 (records); Art. 46 (security).
• Regulation (EU) 2024/1689: Art. 4 (AI literacy); Art. 14 (human oversight); Art. 113 (application).
• Brazilian Decree 10,306/2020: Art. 3, II (definition of BIM); Art. 4, I (first phase).
• Bill PL 2338/2023: approved by the Senate on 10 Dec 2024; pending in the Chamber.