From predictive AI to agentic AI: what changes when software starts to act

Agentic AI is where software stops suggesting and starts acting: it opens a ticket, updates a record, prepares a payment, talks to another system without anyone clicking "confirm." It is also where the question stops being purely technical. When a system acts on the company's behalf, someone has to have decided how far it is allowed to go.

We call a hybrid company one that designs that split on purpose: people set direction, limits, and exceptions; systems carry the volume and the complexity. Getting there has happened in layers, and each one changed something different about the work.

Five layers, by what they changed in operations

1. Classic machine learning. Models learn from history to predict churn, demand, credit risk, or fraud. Platforms such as Vertex AI, SageMaker, and Azure Machine Learning made this routine for data teams. What changed: recurring decisions gained a numeric estimate alongside them.
2. Deep neural networks. Software learned to handle images, audio, and unstructured text: reading invoices, visual inspection on production lines, transcribing support calls. What changed: data that used to need a person looking at it became processable at scale.
3. Generative AI. Language models such as GPT, Claude, Gemini, and Llama write, summarize, translate, and code. What changed: the cost of a first draft fell close to zero. Accountability for the draft still sits with whoever signs it.
4. AI agents. A language model with access to tools (APIs, databases, a browser) that decides its own next steps toward a goal. What changed: AI began executing multi-step tasks instead of only answering questions.
5. Agentic AI. One or more agents coordinating work end to end across a process, with little human intervention between steps. What changed: the whole process, not just a task, can be delegated. That is exactly why governance stops being optional here.

The layers don't replace one another. A mature operation uses all five at once, each where it fits.

Agent or workflow: the distinction that saves money

In December 2024 Anthropic published a guide that became a reference for teams building these systems (Anthropic, 2024). It separates workflows, where models and tools are orchestrated through predefined code paths, from agents, where the model itself dynamically decides which steps to take and which tools to use. The core advice is to start with the simplest solution and add autonomy only when the simple version demonstrably falls short.

In practice, many requests that reach us as "we want an agent" are better served by a workflow: classify an email, extract fields, query a system, draft a reply for review. Workflows are cheaper, more predictable, and easier to audit. Real autonomy pays off when the path can't be known in advance, such as tracing the cause of a mismatch across three systems or researching and cross-checking information from different sources.

What the market is showing

Enthusiasm is high and results are uneven. McKinsey's annual survey, published in November 2025, found that 62% of organizations are at least experimenting with AI agents, but only 23% say they are scaling agents in any business function (McKinsey, 2025).

Gartner was blunter. In June 2025 it predicted that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025). In the same release it named agent washing, vendors rebranding chatbots, assistants, and RPA as "agents," and estimated that only about 130 of the thousands of vendors claiming agentic AI deliver something that deserves the label.

Two infrastructure shifts help separate the serious from the rest. The Model Context Protocol (MCP), an open standard for connecting models to tools and data, was donated by Anthropic to the newly formed Agentic AI Foundation under the Linux Foundation in December 2025, with AWS, Google, Microsoft, and OpenAI among its members (Linux Foundation, 2025). The same month, OWASP published a Top 10 of risks specific to agentic applications, written for systems that plan, keep memory, call tools, and act with delegated authority (OWASP, 2025). A shared protocol and a published risk list are signs of a technology leaving the demo stage.

What weighs more for companies operating in Brazil

Personal data shows up early. Almost every useful agent in customer service, collections, or sales touches customer data. Under the LGPD, Brazil's data protection law, the company remains the data controller even when the one accessing the data is an agent running on a foreign vendor's API. That requires a legal basis, a vendor contract covering international transfer, and a log of what the agent looks up. Decisions made solely through automated processing that affect an individual can be subject to a review request under Article 20 of the LGPD.

The channel is usually WhatsApp. A large share of customer relationships in Brazil runs through it, which makes conversational agents the first use case at many companies. It is also the channel where a mistake is visible to the customer immediately.

Legacy systems are the bottleneck. An agent can only act where there is an interface to act on. An old ERP with no API, a spreadsheet as the source of truth, or a process that depends on someone operating a screen limits what can be automated safely. Often the first "agents" project is really an integration project.

An example autonomy ladder

Take a common finance process: resolving mismatches between a supplier's invoice and the purchase order. An illustrative example of how to climb the ladder:

1. Level 1, assisted: the system reads the invoice and the order, flags the mismatch, and suggests the likely cause. The analyst decides everything.
2. Level 2, with approval: the agent prepares the reply to the supplier and the adjusting entry. The analyst reviews and approves.
3. Level 3, autonomous within limits: mismatches below an amount set by finance and with a known cause, such as rounding or freight, are resolved without intervention. Everything else goes to a person.
4. Level 4, supervised by sampling: the agent resolves most cases and the team audits a weekly sample, adjusting limits based on the errors it finds.

Each step is unlocked only after the previous one shows, with data, an acceptable accuracy rate. And the level 3 threshold is the CFO's decision, not the technology team's. That is what we mean by a hybrid company at work: the machine carries the volume, the person decides how far it can go.

Before you start, check

• Does the process have enough volume to justify the effort? Ten cases a month rarely justify an agent.
• Is the path predictable? If so, start with an automated workflow, not an agent.
• Is there an API or integration for every action the agent would need to take?
• Is there a business owner accountable for the outcome who can set the autonomy limits?
• Can you measure accuracy today, without AI, so you can compare later?
• Does the personal data involved have a legal basis and a proper contract with the model vendor?
• Is there a clear path for a case to leave the agent and reach a person?

If most answers are yes, the next step is to validate the case with real data in a few weeks, before committing a full project budget. That is what the PoC Factory is for: a defined hypothesis, a three-to-six-week build, and a Decision Package recommending whether to scale, adjust, or drop it. If the case is already validated, see how we work with AI agents in production.