Gartner projects that companies will spend US$ 206.5 billion on AI agent software by 2026, a 139% increase from US$ 86.4 billion in 2025. For comparison, the overall AI market is expected to grow 47% during the same period. This means that AI agents are growing almost three times faster than the rest of AI. When money moves at this speed, it's telling a story.
This post translates this story for those who decide on budgets: what changed between the era of beautiful demos and the era of agents that actually work, why a large chunk of these billions will be wasted, and how to be on the right side of this trend as a medium-sized Brazilian company, not a big tech.
The number, and the catch within it
First, the vocabulary. An AI agent is not a chat that answers questions. It's a program that takes on an end-to-end task: reads the request, consults the company's systems, decides what's routine and what needs human intervention, executes, and delivers the result. The difference between a chat and an agent is the difference between a customer support agent explaining how to fill out a form and an employee filling out the form for you.
Now, the catch. Within the same set of projections, Gartner points out that only 17% of organizations have AI agents up and running, and estimates that more than 40% of AI agent projects may be canceled by the end of 2027. Combine these three data points: spending more than doubles in a year, a minority have put agents into production, and nearly half of projects are expected to be canceled.
This isn't a contradiction. It's a snapshot of a market that has proven the value of technology but still struggles with execution. The difference between those who reap results and those who cancel projects isn't the size of the check. It's what comes in the next sections.
What changed from 2024 to 2026
In 2024, every vendor had a demo of an agent. A robot chatting nicely in a sales meeting, answering three rehearsed questions, impressing everyone. Many companies bought this demo and discovered in the following months that impressing in a meeting and withstanding real-world operations are different sports. We've written about this in detail: why agents fly in demos but die in production.
In 2026, the money is going elsewhere. Companies no longer pay for demos; they pay for agents that take on an entire workflow: first-line support, document verification, order triage. With two conditions that have become market standards: human review on critical points and proven savings in reports, not just promises on slides.
This change explains the 139% growth. Buyers have matured. They're no longer paying for impressive technology; they're paying for delivered work. And delivered work can be measured: hours saved, errors avoided, deadlines shortened, customers responded to at 2 AM.
The hard part isn't intelligence; it's the rest
This is why 40% of projects will fail, and it's the most important point in this text: the AI model is almost never the problem. Today's models are good enough for the vast majority of routine tasks in a company. What determines success is everything around the model.
Think about hiring a brilliant professional, fresh out of school, on their first day of work. Their intelligence alone doesn't solve anything. Without training, they don't know the company's rules. Without performance evaluation, you don't know when they make mistakes. Without defined authority, they sign off on things they shouldn't. It's the same with AI agents, and there are exactly three key components:
1. The right context at the right time. The agent needs access to what your company knows: current exchange policies, the customer's history, updated price tables. An agent responding with outdated or generic information isn't an agent; it's a risk with a monthly subscription. Most of the work in building a good agent is organizing and delivering this information to them when they need to make decisions.
2. Error measurement before customers do. All agents make mistakes. The question is whether you catch the error in internal testing or when a customer complains. Serious operations run continuous tests on the agent, like quality control on a production line: daily samples, purposefully difficult cases, minimum scores to stay online. Those who don't measure don't know they're making mistakes, and those who don't know they're making mistakes cancel the project six months later without understanding what happened.
3. Limits that prevent agents from making things up. Clear rules on what they can never do alone: promise unconfirmed deadlines, offer discounts outside their authority, respond to topics they don't master. And an escalation path: when in doubt, pass it to a human, don't improvise. We've also explored this in more depth: what truly protects an AI operation.
None of this shows up in a demo. All of this shows up on the third-month invoice.
How it plays out in practice
The pattern of successful cases is always the same: a specific workflow, with human review and measured results. In the Brazilian market, there are already mature examples, such as WhatsApp assistants that respond to customers 24/7 and escalate to human teams when necessary.
At Steply, the case we use as a benchmark is in finance: an agent that took on payment verification and reconciliation, a task that consumed hours of qualified personnel comparing lines. The agent handles routine tasks, humans review exceptions, and savings appear in counted hours, not just impressions. We detailed this in how we unlocked the bottleneck in reconciliation.
Notice what these cases aren't: they're not the entire company turning into AI, not 15 agents launched at once, not two-year projects. They're a tedious, repetitive, and expensive workflow being handed over to an agent with supervision. This is how the US$ 206 billion turns into returns instead of becoming 40% canceled projects.
Buying tools or building operations
A divide is forming among companies, and it will become more visible throughout 2026. On one side are those who only sign up for generic AI tools: they have exactly what their competitors have, at the same price, with the same results. Generic tools aren't competitive advantages; they're new expenses.
On the other side are those who build agents on top of their own processes: customer support with in-house rules, finance with in-house exceptions, triage with in-house criteria. These companies use the US$ 206 billion to their advantage because the entire market is making base technology cheaper and more mature, and they apply this technology to bottlenecks that only they understand. We've shown how this logic is shaking up even the software subscription model: money is shifting from paying for access to paying for results.
Where to start without spending like a big tech
If your company doesn't have any agents up and running yet, the recommendation is almost anticlimactic: start with a small agent for a real and annoying problem. Repetitive, with clear rules, with measurable pain in hours or in real dollars. Include human review from day one. Measure before and after. Only then consider the second agent.
The question of 2026 is no longer "will we use AI?". The market has already answered that with US$ 206 billion. The question that separates those who reap results from those who cancel is another: which workflow do we hand over to an agent first, and how will we know, with numbers, if it's working well? Those who answer this before signing any contract are already ahead of the majority.
