Agentic AI: the real value is in redesigning processes, not in adding bots

One year after the commercial boom of agentic AI, systems that not only answer questions but execute sequential tasks, make intermediate decisions, and operate with minimal supervision, McKinsey published an uncomfortable conclusion for those who took the easier route: technology alone does not deliver value. What delivers value is the redesign of the work the technology will perform.

The consultancy analyzed implementations across multiple sectors and reached a direct diagnosis: the companies that achieved the highest returns were not those that layered AI agents onto existing processes. They were the ones that mapped their operations end-to-end, identified the real bottlenecks, and built monitoring into each step before deploying any automation. Those that did the opposite, and plugged a bot into a dysfunctional flow, accelerated the problem, not the solution.

What agentic AI actually is

Before getting into strategy, it is worth naming the object. Agentic AI is not a sophisticated chatbot. It is an architecture in which language models (such as GPT-4o, Claude 3.5 or Gemini 1.5) are combined with external tools, APIs, databases, CRM systems, ERPs, and given an objective, not just a question. The agent plans steps, executes actions, verifies results, and iterates until the task is complete.

Platforms such as LangChain, Microsofts AutoGen and Googles Vertex AI Agent Builder are today the main infrastructures for building these systems. Companies like Salesforce and ServiceNow have already embedded agents natively in their products, and Agentforce, launched by Salesforce in 2024, is a concrete example of how this arrives in the B2B market: an agent that handles support tickets, checks the customer history, opens cases, escalates to humans when necessary, and logs everything in the CRM, without manual intervention in most flows.

The most common mistake: automation without redesign

The trap McKinsey documents is familiar to any consultant working with operations: the company identifies a repetitive task, automates that specific task with an agent, and celebrates. But the process around it remains full of redundant approvals, mandatory fields that serve no purpose, and handoffs between departments that exist because of historical inertia.

The result is a fast agent operating inside a slow system. The efficiency gain is marginal and, worse, it creates a false sense of modernization.

What works, according to the analysis, is the opposite: first map the entire flow, then decide where the agent fits. This requires answering questions companies often avoid, such as why this approval exists, who actually uses this information, and what happens when this step fails.

Monitoring is not optional

Another critical point raised by McKinsey: agents in production need monitoring layers built from the start, not added later. This includes logs of every decision the agent makes, alerts for behavioral deviations, thresholds for human escalation, and periodic quality audits.

Without that, operational risk grows silently. An agent that approves purchase requests based on poorly calibrated rules can cause losses for weeks before anyone notices. Not because the technology failed, but because no one defined what "right" means in that context.

What this changes for SMEs in Brazil

For small and medium enterprises, which represent more than 99% of establishments in Brazil and account for about 70% of formal employment, according to Sebrae, McKinseys message has a liberating side and a challenging side.

The liberating side: the biggest opportunities are precisely where SMEs concentrate their most repetitive work. Customer support, triage of internal requests, credit approvals, supplier onboarding, HR help desks, these are all processes structured enough for an agent to operate effectively, provided the flow is clean.

The challenging side: redesigning processes takes time and requires someone inside the company to understand the business well enough to question the status quo. For many SMEs, that resource is scarce.

The practical solution Ive seen work with clients at FM Solutions is to start with the most complained about process, the one that generates the most tickets, rework, or internal dissatisfaction. Not the simplest. The most painful. Because that is where redesign has real urgency and where results appear quickly enough to sustain the project politically within the organization.

Human-agent collaboration, not replacement

A fact often ignored in discussions about agentic AI: the best-documented outcomes are not total automation, but structured collaboration between humans and agents. The agent handles the busywork, collects data, formats, verifies, and records. The human decides on edge cases and validates exceptions.

This model reduces operational workload without removing human judgment at moments when it truly matters. For SMEs with lean teams, it means a team of three can operate with the capacity of six, without hiring and without increasing fixed costs.

The right question to ask now

McKinsey documents what practice already showed, agentic AI is a lever, not a shortcut. Companies that skip steps and install agents without redesigning what the agents will execute will spend money automating chaos.

The question to ask today is not "which agent should I buy?" It is, "which of my processes deserve to exist the way they do?"

Those who answer that question honestly before adopting the technology are one step ahead, regardless of company size or budget.