AI adoption is already widespread. Now the first data is emerging from the next stage: AI agents actually performing the work. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, driven by spiraling costs, unclear business value and weak risk controls. McKinsey, in its 2026 State of AI, adds a wider signal about AI overall, not only agents: 80% of the executives surveyed say it improved their individual productivity, while only 37% see a real contribution to operating results (EBIT), a share that has not moved in a year.
Those figures describe a gap between adopting AI and turning it into results. In our view, closing that gap requires sound process architecture and governance. The technology for building agents already works; the challenge is closing the gap between a pilot that impresses in a demo and a process that runs every day, with rules, permissions and owners. Closing that gap is, above all, a design problem.
Why pilots stall
The barrier usually lies in integration with the real process: generic tools don't learn the workflow or adapt to it. They resolve an isolated query and stop where the real operation begins, with its systems, its data and its exceptions. The pilot looks good in a demo and breaks down when it has to integrate with the real business process.
Gartner points to a second factor: agent washing. Of the thousands of vendors that market themselves as agentic, its analysis identifies around 130 with genuine capabilities. This creates the risk of inflated expectations, underestimated integration challenges and no clear plan for control.
How do you automate a sales process with a team of agents?
Take a sales operation. A new lead comes in, and it has to be identified, researched, validated, checked against criteria, recorded, moved to the next action and, in some cases, handed to a person. These are distinct tasks, with distinct information and distinct levels of risk.
A single generalist agent could handle the entire workflow. A team of specialized agents breaks the process into distinct roles, each handled by a designated agent, and coordinates the work, with control points wherever the risk calls for one. How that work is divided and governed is an architecture decision, and it weighs heavily on whether the project reaches production. In the lead example: one agent gathers and normalizes the information, another checks it against the business rules, another updates the records and proposes the next action, and the account owner signs off before any outreach.
Autonomy is governed by risk level
An agent that only looks up information carries limited risk. One that edits records, sends communications or executes transactions needs precise rules before it operates. Before switching it on, you define what information it can access, what systems it can use, which actions it is authorized to perform, what it must escalate, when it needs human authorization, and how everything is logged. These are business decisions, not only IT ones.
This is the role of agentic AI infrastructure: providing the permissions, rules, orchestration and traceability that govern how agents operate. If an agent made a decision, you should be able to reconstruct what data it used, what rule it applied and what happened next. Without that layer, autonomy is precisely the risk Gartner points to.
What happens when something fails
A system is ready to operate when it knows what to do when something goes wrong, not only when it works in the ideal case. Before it goes live, you test it with real-world cases, including exceptions, set minimum quality thresholds and keep everything monitored. If an agent gets something wrong or a connected system goes down, the process has to be able to stop, retry or hand off to a person, with a record of what happened. That ability to halt and escalate is part of the design, alongside permissions and approvals.
How do you take a team of agents to production?
A team of agents that reaches production follows an ordered path. At GoCode, we structure this work around five stages:
Process assessment
Data and permissions
Team design
Governance and human oversight
Production and ROI
The order matters. Starting with the technology, before defining the process, the controls and the measurement, can undermine the implementation. The path that reaches production starts with the process.
Working with a specialist or building in-house
A 2025 MIT study measured this directly. Solutions bought from a specialized vendor or built through a partnership succeeded about 67% of the time, against a third of that rate for in-house builds. It is a correlation within a reported sample, not a causal law. And the biggest return did not show up where most of the budget goes: more than half is spent on sales and marketing tools, while the highest return sits in automating internal processes.
McKinsey 2026 also shows the gap by size: large companies that have already scaled agents in at least one function went from 27% to 40% in a year, while smaller ones stayed flat, near 22%. For a mid-sized company in Chile or Latin America, and for whoever runs the operation, that does not dictate where to start. It makes sense to weigh the commercial and the operational opportunities against the same criteria: expected benefit, available data, complexity, risk and the capacity to sustain the solution over time. A specialist brings experience and can reduce the learning curve; the decision depends on the process you choose and your in-house capabilities.
When does a team of agents make sense, and when is conventional automation enough?
Not every process justifies a team of agents. A team of agents makes sense when a process has several stages, several systems, high volume and distinct roles. A few clear use cases:
Legal: reviewing contracts against the company standard, with a summary of deviations for the lawyer.
Finance: reconciliations and close preparation, with alerts on thresholds.
Insurance: intake, classification and verification of claims before review by a claims adjuster.
Operations and procurement: preparing orders and comparing quotes, with approval from the person responsible.
Across all of them, the agents prepare, verify and document, and a person approves what is sensitive. Where the process is a single step or low volume, a narrow task is handled by a single agent or by conventional automation, depending on its complexity. Recommending the architecture that fits the process, rather than forcing agents into everything, is part of the design.
| Criterion | Conventional automation | Team of AI agents |
|---|---|---|
| Process | One or a few stable steps | Several stages that vary from case to case |
| Rules | Fixed and predictable | Require judgment and involve frequent exceptions |
| Systems | One or two | Several; information has to be cross-checked |
| Decision | Deterministic | Interprets, compares and prioritizes information |
| When to choose it | High volume, low variation | High volume with variation and distinct roles |
The underlying question has shifted to the operational: which part of the operation would work differently if some of its roles were run by a team of agents, and how we want that team to work with the systems and with people. The next leap comes from designing that team and its governance better, more than from waiting for a smarter model.
To explore how this could work in your business, learn about GoCode's AI agent team services for businesses in Chile and Latin America.
Frequently asked questions
When does a team of AI agents make sense instead of conventional automation?
When the process has several stages, crosses several systems, handles exceptions and runs at high volume. If the rules are stable and the steps predictable, conventional automation is usually enough. You choose the architecture to fit the process, not the other way around.
Do you have to replace your current systems (CRM, ERP) to use AI agents?
No. A team of agents is designed to operate on the infrastructure you already have: it looks up information, cross-checks data, updates records and executes actions with minimum permissions. It is a new layer over your existing systems.
What happens if an agent gets something wrong or a connected system fails?
The process has to be able to stop, retry or hand off to a person, with a record of what happened. That ability to halt and escalate is defined in the design, before anything goes live, together with the human approval points.
What prevents AI pilots from reaching production?
Because the distance between a demo and real operation lies in integration, governance and measurement. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, driven by costs, unclear value and weak controls.
How do you start automating a process with a team of agents?
With a narrow process that has volume, clear rules and a measurable result. You map the current process, define data and permissions, divide the agent roles and the human control points, and measure against a baseline before scaling.
How much autonomy should you give an AI agent?
As much as the risk allows. An agent that only looks up information carries limited risk; one that edits records or executes transactions needs precise rules, approvals and full traceability. Autonomy is governed by risk level, decision by decision.