AI Agents in Finance: how to apply artificial intelligence with control, traceability and practical results

The conversation about AI in finance has matured.

The doubt is no longer whether the technology works. In many cases, it already works. The point now is different: how to apply artificial intelligence in real financial processes, with control, traceability and adherence to daily operations.

This matters because most of the financial work still happens outside the ERP, spread across emails, PDFs, portals, spreadsheets and systems that don't talk to each other. It's in this environment that rework, manual error, low predictability and difficulty scaling without increasing the team arise.

This was exactly the focus of the AI Agents in Finance class, presented by Catarina Pinheiro, CFO of Abstra and PhD in Artificial Intelligence: showing how to combine automation and AI to reduce operational effort in critical processes, without giving up governance.

The real problem of finance is not just automation

When talking about automation, many people still think of linear, predictable and structured flows. This type of automation remains extremely valuable. The problem is that it only covers part of the operation.

In practice, finance deals with:

• high volume of operational and repetitive tasks
• exceptions and lack of standardization
• documents in different formats
• external portals and systems without API
• processes that require control, audit and security all the time

Therefore, the challenge is not just to automate. The challenge is to automate with context.

This is exactly where AI Agents start to make sense.

What is an AI Agent, in practice

An AI Agent is not just a model answering questions.

In practice, it is a system that receives an objective, interprets the context, breaks the problem into steps, uses tools and makes decisions throughout execution until reaching a result.

Think of an objective like this:

"Process all energy invoices for the month."

To execute this, the agent can:

1. access a portal
2. download files
3. extract data from documents
4. validate inconsistencies
5. forward the result to ERP
6. generate alert in case of divergence

The central difference is here: instead of following only a fixed path, the agent can act based on objective, context and available tools.

Traditional automation and AI Agents don't compete with each other

One of the most important points from Catarina's class is that this is not a substitution discussion.

It's not automation OR agent.

It's automation PLUS agent.

Traditional automation remains ideal for the structured part of the process:

• direct integrations with ERP
• deterministic rules
• predictable flows
• steps without variation

Agents come in where there is variation:

• PDFs and emails
• exceptions
• external portals
• decisions that depend on context
• systems without API

This combination is what allows automating the real process, end-to-end.

Where AI Agents make the most difference in finance

Not every process needs AI. But some hardly scale well without it.

1. Tasks in portals and systems without API

This is one of the most common bottlenecks in finance. Many processes still depend on legacy systems, government portals and interfaces that require manual navigation.

In this scenario, the use of browser automation with agents allows the system to act like a human user: accesses pages, fills fields, clicks buttons, extracts data and follows the necessary operational flow.

The absence of API ceases to be an absolute technical blocker. The process remains automatable.

2. Reading and interpreting documents

Another critical point is the volume of documents that reaches finance:

• invoices
• payment slips
• receipts
• contracts
• utility bills

The gain is not just in "reading PDF". It's in classifying the document, extracting the right fields, validating rules, associating this document to the correct process and ensuring the result is structured for audit and operation.

3. Accounting and financial classification

Classifying transactions consistently remains an operational and accounting challenge.

With AI, it's possible to suggest accounting account, cost center and category based on description, supplier, history and process context. After that, the flow can automatically accept, forward for human review or escalate exceptions, according to confidence level.

This type of application reduces manual typing, increases standardization and improves the quality of the base used by accounting, controllership and FP&A.

A classic case: NFSe extraction in government portal

A didactic example presented in the class was the extraction of service invoices in portals like NFSe.gov.br.

Manually, this process usually consumes hours from the team every month:

• login with digital certificate
• navigation to the correct area
• period filtering
• opening notes
• PDF downloads
• information collection for later processing

With an agent, the flow can be structured to:

• access the portal
• navigate the interface
• extract relevant data from each note
• locate the correct access key
• download documents
• return output in structured format

The most interesting point here is not just the time savings. It's the fact that the process stops depending on repetitive manual execution and becomes a more consistent, auditable and predictable flow.

Another high-impact case: utility bills

Energy, water, telecom and rent usually seem simple, but generate relevant operational volume when distributed across multiple units, suppliers and due dates.

In this type of process, an agent can:

• access the portal or billing link
• download the bill
• extract value, due date, consumption and unit
• create the entry in ERP
• associate cost center
• compare with history
• signal deviations or charges outside the standard

Besides reducing operational effort, this improves due date control, visibility for FP&A and ability to detect anomalies before payment.

OCR in finance is not just document reading

A common mistake is treating OCR as the end of the process. In finance, it's just a part.

Real value emerges when document reading comes accompanied by:

• file type classification
• structured data extraction
• mandatory field validation
• checking against internal policy
• association with orders, contracts or expenses
• database organization for audit

This is what transforms OCR into truly useful automation.

Contracts also fit this logic

Another very relevant use is reading contracts with AI.

Instead of depending on manual reading of long PDFs to register clients, validity, readjustment, fine, payment method and other conditions, the process can extract this data in a structured way to feed systems and reduce operational delay.

Fields can include, for example:

• company and CNPJ
• start and end dates
• value and frequency
• payment method
• renewal clauses
• termination fine
• readjustment index
• contracted plan and usage limits

This type of extraction reduces operational cost, decreases transcription error and accelerates processes like registration, billing and contractual governance.

Intelligent classification improves operation and accounting quality

In the classification part, the most robust architecture usually doesn't depend only on pure AI.

The most reliable pattern usually combines:

• a quick heuristic, when available
• fallback to AI in more ambiguous cases
• validation by confidence level
• final decision: accept, ignore or send for human review

This approach is important because the objective is not to create "magical" automation. It's to create reliable automation.

When well implemented, intelligent classification improves:

• consistency in chart of accounts
• closing speed
• report quality
• audit capacity
• detection of anomalies and necessary reclassifications

The real challenge is not technology. It's orchestration

This might be the most important point of the article.

Technology already exists. The problem is no longer proving that OCR works, that agents can navigate portals or that AI can classify transactions.

The real challenge is in orchestrating all this within the financial environment, where:

• there are exceptions
• rules change
• systems don't talk
• control is non-negotiable

Isolated automation doesn't solve the entire process.

Isolated AI doesn't either.

ERP alone, even less.

What solves is the combination of:

• deterministic automation
• agents for context and variation
• business rules
• human approval points
• audit trail
• governance from architecture

How to start the right way

The recommendation presented in the class is simple: don't start trying to automate everything.

Start with a process that has:

• high manual volume
• clear operational impact
• relatively understandable rules
• measurable result

Good entry points are usually:

• invoices
• recurring bills
• accounting classification
• pre-payment validations
• flows with document reading

From there, the most solid path is:

1. Map the priority process

Understand where there is greater manual effort, greater error risk and greater operational cost.

2. Run a pilot with real case

Validate accuracy, saved time, exceptions and failure points before scaling.

3. Build governance from the beginning

Audit, fallback, human approval and traceability should not come later. They need to be born with the flow.

4. Evolve to orchestration

Connect automated processes and move from isolated tasks to integrated financial flows.

In summary

AI Agents make sense in finance because they solve a part of the problem that traditional automation, alone, cannot cover well: the part of variation, context and exceptions.

They don't replace rules, integrations or governance.

They expand what can be automated.

When used correctly, they allow finance to advance from isolated automations to truly smarter, traceable and scalable processes.

FAQ (Frequently Asked Questions)

What differentiates an AI Agent from traditional automation?

Traditional automation follows fixed rules and predictable flows. AI Agents can interpret context, handle exceptions and make decisions during execution, adapting to process variations.

In which financial processes do AI Agents make the most difference?

Especially in processes with unstructured documents, portals without API, classification that depends on context, and flows with many exceptions - like NFSe processing, utility bills and accounting classification.

Do AI Agents replace ERP?

No. They complement ERP, automating processes that happen before, after or outside the main system - like portal extraction, document processing and complex validations.

How to ensure control and audit with AI Agents?

Through architecture that includes audit trail, human approval points, validation by confidence levels and fallback to manual review when necessary.

What's the first step to implement AI Agents in finance?

Choose a process with high manual volume and clear impact (like invoices or classification), run a controlled pilot and build governance from the beginning before scaling.

Can AI Agents work with legacy systems?

Yes, through browser automation they can navigate portals and systems without API as if they were human users, overcoming technical integration limitations.

---

Want to understand how to apply AI Agents in finance in practice?

👉 Access the complete AI applied to finance class with Catarina Pinheiro, CFO and PhD in AI.