Superintelligence for Outsourced HR: Screening, Payroll and Onboarding

Superintelligence, in the sense the term takes on in marketing copy, is not what an outsourced HR team needs to buy. What it can prepare with AI agents are narrow tasks: organizing resume screening against defined criteria, assembling the onboarding document package, and checking inconsistencies before the payroll department closes the month. All of this should run with restricted access to personal data, a record of every action, and human review for any decision that affects a person.

Key takeaways

• The agent helps screen and prepare information, but it does not decide hiring, dismissal, pay adjustments or disciplinary measures.
• Onboarding is a checklist process: the agent follows up on missing items, confirms receipt and logs what is still pending, without keeping more data than necessary.
• In payroll, the useful role is flagging discrepancies and preparing questions for whoever closes the period, not recalculating charges on its own.
• Health data, identity documents and bank information stay outside the agent's reach by default.
• Start with a single workflow, with logs, a metric agreed in advance, and clear criteria for expanding, adjusting or stopping.

What an agent does well in an outsourced HR firm

An HR consultancy handles many clients, many open positions and many short deadlines. Much of the work is repetitive and depends on checking information against a list: the job posting has its stated requirements, the document was sent, the payroll period includes every expected event. These are exactly the tasks where an agent with clear instructions and limited tools performs well.

Before thinking about technology, it helps to separate three types of work. The first is objective screening, where a resume is compared with explicit requirements such as required education, verifiable years of experience and declared availability. The second is administrative organization, such as tracking documents and reminding people of pending items. The third is judgment, such as assessing cultural fit, interpreting an interview or deciding whether a discrepancy is an error or a legitimate situation. The agent can help with the first two. The third belongs to people who are accountable for the decision.

A practical rule keeps this boundary clear: the agent recommends, the human decides, and the recommendation must be traceable. If an output cannot be explained from the criteria in the job posting, it should not exist.

Resume screening without inferring what is not written

In screening, the main risk is not speed but inference. A language model can suggest age, marital status, religion, sexual orientation or health from indirect signals in a resume, such as dates, institution names or gaps in employment history. None of this is a legitimate selection criterion, and treating sensitive data as an input to a ranking creates a legal and ethical problem that no efficiency gain justifies.

For that reason, the instructions to the agent must be restrictive. It evaluates only the criteria listed in the job posting, responds with the evidence found in the text (quoting the relevant passage), states when a requirement does not appear, and never produces personality scores or personal profiles. Fields such as photo, full home address and date of birth should be kept out of the input, because they are not needed for the assessment.

The people making the decision also need to know what the tool does not do. A sorted list of candidates is only a reading aid. A recruiter should review the rejected candidates, not only the approved ones, because wrongly excluding a strong candidate is the error that usually goes unnoticed. For the topic of automated decisions and the right to review, see LGPD automated decision-making and the right to review (Art. 20).

Onboarding: checklist, deadlines and a secure channel

Onboarding is one of the best places for an agent, precisely because it is a checklist process with a lot of repetitive communication. The agent can keep the status of each candidate, remind people of pending items and notify the team when a document has arrived or is unreadable. The central question is where those documents travel.

A common mistake is asking candidates to send personal documents directly in an open WhatsApp conversation, mixing personal data with routine messages. A safer approach is for the agent to explain which documents are required and send an upload link with controlled access, from which it only receives confirmation of receipt and the status, without storing the file contents. Before designing this, check the usage rules of the messaging platform, such as the WhatsApp API 24 hour window, and how data is transferred and stored, especially if any service processes information outside Brazil, a topic covered in international data transfer under LGPD when using AI providers.

A simple work plan for onboarding can follow this order:

• Define the document list by employment type and by role, validated by the client's payroll department.
• Separate what the agent can see (status, date, type of pending item) from what only the responsible person can access (document contents).
• Create message templates for reminders, confirmations and pending items, without improvised text about legal deadlines.
• Log every event: who sent what, what the agent replied, and when a human took over.
• Route any question, mismatched document or request for an exception to a named person.

Payroll support: flag discrepancies, do not recalculate everything

In payroll, the temptation is to let the agent do the calculations. This is risky. Social charges, deductions, termination pay and events submitted to eSocial depend on specific rules, collective agreements and individual contracts, and a language model should not be the source of any payroll figure. What it does well is comparison: a line item that appeared with no history, an employee with two entries in the same month, a pay adjustment applied without a record of the agreement.

In this model, the calculation stays in a deterministic system validated by the team, and the agent works on the output. It receives only the fields needed for the check, usually an internal identifier, the line item, the amount and the period, and returns a list of alerts with the reason and a reference. The person closing payroll decides what to correct. This reduces exposure of salary data, which is usually sensitive within a company, and makes clear who is responsible for each number.

This design also avoids treating the model's answer as accounting truth. An alert that turns out to be wrong costs a few minutes of review. A wrong figure sent to a government system costs far more.

Personal data: what stays out of the agent

The LGPD treats health data, for example, as sensitive personal data, which requires additional care. Medical certificates, admission exams, leave records and disability information should not enter prompts, knowledge bases or conversation logs. If the process requires that data, it should be handled by an authorized person, in a system with access control, and the agent should only know that the item exists and what its status is.

Some practical principles for operations:

• Minimize: the agent receives only what the task requires, not one column more.
• Segregate: separate profiles and permissions for recruiter, payroll and agent.
• Record: log every query and every action, with the identity of whoever authorized it.
• Review: check instructions and connected tools periodically, especially after workflow changes.
• Restrict tools: the agent should not have write access to payroll systems without human approval.

An internal AI use policy helps align the team. A practical model is described in company policy for employee ChatGPT use, and the same logic applies to agent tools. The Brazil AI regulation framework is worth following, but it does not replace a case-by-case LGPD analysis. For official guidance on data protection, the Brazilian National Data Protection Authority (ANPD) publishes material on its website.

How to start: map, limit, pilot and hand over

The method we apply in AI projects has four steps: map the current process and its baseline, limit data and actions, pilot a real workflow with logs and an agreed metric, then hand over and decide whether to expand, adjust or stop. In outsourced HR, this path prevents the common error of starting with the most visible tool and discovering too late which data it touched.

In the mapping phase, the question is how much time the team spends on each step and where errors usually appear. Without that baseline, no improvement can be verified. In the limiting phase, you define which data goes in, which actions the agent can take, and what requires approval. In the pilot, you choose a single workflow, such as chasing onboarding documents for one group of clients, and one simple metric, such as the time until a pending item is resolved, compared with the baseline. At handover, the internal team takes over the operation, with written instructions, an exception procedure and periodic review.

Each stage can be tested in controlled environments with fictitious data before any real data is used. This practice is described in how to test AI agents with mocks and a sandbox.

Illustrative scenario: onboarding at a firm with several clients

Illustrative scenario: an HR firm serves mid-sized companies and, during intense hiring periods, handles many onboardings at the same time. The team notices that most of its time goes to chasing documents and answering the same questions, while each client's payroll department waits for complete packages before registering the employee.

The chosen pilot covers only pending-item follow-up for onboarding at two clients. The agent receives the document list for each role, sends an upload link, confirms receipt and records the status. It does not open the files, does not evaluate their content, and routes any exception to the responsible analyst. Every action generates a record. The agreed metric is the time between opening the process and having a complete package, compared with the same indicator from earlier months.

After a few weeks, the team reviews the log and finds two situations that were not anticipated: candidates answering in conversations with photos of documents, and one role with a requirement that differed from the standard list. Both led to changes in the instructions and the channel. The decision to expand to the remaining clients comes only after that review, which is the point of the method: the pilot exists to discover what does not work before the operation grows.

Frequently asked questions

Do I need superintelligence to automate screening or onboarding?

No. For these tasks, what matters is an agent with clear instructions, objective criteria and limited tools. Today, the term superintelligence describes a subject of research and debate, not a product an HR team buys. A simple, well-defined solution with human review is usually more useful than a generic promise.

Can the agent calculate payroll on its own?

It is not advisable. Social charges, deductions and events depend on specific rules and agreements, so the calculation should stay in a deterministic system validated by the team. The agent can check results, flag discrepancies and prepare questions for whoever closes the period. Any correction should go through recorded human approval.

Is it safe to receive onboarding documents through WhatsApp?

It depends on how the process is designed. Sending personal documents in an open conversation mixes sensitive data with routine messages and makes access control harder. An upload link with restricted permissions, where the agent receives only confirmation and status, is usually a more appropriate alternative. Review the legal basis and your internal policies before deciding.

How do I know if the pilot worked?

Define a simple metric with the client before starting, compared with a baseline, such as the time until an onboarding package is complete or the number of unanswered pending items. Review the log with the team, check errors and exception cases, and then decide whether to expand, adjust or stop. A pilot that does not generate enough records to support that decision is not yet ready to be evaluated.

Sources

• NIST AI Risk Management Framework
• OWASP Top 10 for LLM Applications
• Anthropic: Building effective agents
• ANPD: Autoridade Nacional de Proteção de Dados