What to do when your team already uses its own AI agents
- AI
- Agentic AI
According to Work:InProgress 2026, a study by IDC and Provokers for Google Cloud, 44% of professionals surveyed in five Latin American countries regularly use personal AI agents at work, and only 28% can point to a clear usage policy. Our reading is that this behavior fills a gap the company left open, which is why a ban won’t close it. What closes it is an official path that’s easier to use than the personal one.
If you lead technology, there’s one finding worth looking at first. Among professionals who use a personal AI agent for work, 36% say they do it because they feel they have more privacy. The privacy the employee gains is visibility the company loses: what data leaves, which service it goes to, and under what terms. And in that same group, 54% see little or no risk.
This isn’t a fringe behavior either: 44% of respondents regularly use agents their company didn’t give them.
What the study says, and what it doesn’t
IDC and Provokers ran Work:InProgress 2026 with Google Cloud. It’s based on 3,200 online interviews conducted in July and August 2026 in Argentina, Brazil, Chile, Colombia, and Mexico, with professionals at companies of 500 or more employees. Keep three things in mind as you read it:
It measures self-reported perceptions, not observed usage.
The sample covers mid-size and large companies, 16% of them in the public sector. It doesn’t describe a 30-person startup.
It’s published by a vendor that closes the report by recommending its own platform.
None of that invalidates the numbers, but it does change how you read them.
There’s also a wording detail that matters. The question asks about agents “for personal use, not provided by the company” (our translation); it doesn’t ask whether they’re prohibited. Some of that use may be allowed, or simply something nobody has thought about yet. Calling all 44% “shadow AI” works in a headline, but it leads to the wrong response: treating as a violation something that, in many cases, is a gap on the company’s side.
The data on what companies provide points the same way. When it comes to AI agents, 65% of respondents don’t report corporate access, 63% don’t report regular training, 69% say no one has shown them use cases, and 72% can’t identify a clear policy. At the same time, 85% say their company is already showing interest in agents or running initiatives with them, while only 12% describe mature adoption, with governance and integration across processes.
In most cases, the interest is there, but the tool, the rules, or the training is missing. In that gap, people make do with whatever is at hand.
Why people use their own agents
When asked why they use a personal agent, people give reasons that have more to do with convenience and confidence than with rebellion: 38% say it’s more accessible and easier to use, another 38% feel safer using it, and 36% mention privacy. Gaps on the company side also show up among those reasons: 21% say the corporate tool doesn’t integrate with the other solutions they need, 19% say their company doesn’t offer AI resources, and 11% say they weren’t properly trained to use the corporate tool.
The mechanism is the same as with any poorly designed internal tool. If the official path doesn’t exist, or it exists but forces people to copy and paste across three systems, they take the shortcut. No bad intent is required; a Friday deadline is enough.
The same study shows the other side: 61% believe they get more value from agents built into the tools they already use every day, either on their own or combined with specialized apps. For most people, then, the value lies in bringing AI to where they already work rather than adding yet another tool.
A case of ours with the same pattern
We ran into something similar in a different area. One of our deployment teams didn’t review the usage licenses for some libraries and fonts. The software went to production, and later we received a claim about the fonts. Fortunately, they weren’t in use.
Nobody decided to use something they shouldn’t have. What was missing was a step in the process that required checking third-party material before it went into the product, so that’s what we put in place: a license verification process.
The same thing happens with personal agents, only faster and with data instead of fonts. Every time someone pastes a contract, a customer list, or a piece of code into a tool the company hasn’t evaluated, material goes into a third-party system without anyone reviewing it. That’s why reminding people to be careful isn’t enough. What fixes the problem is putting the control at the point where the data flows.
Before you offer “agents,” define what an agent is
When the study asked professionals to describe in their own words what an AI agent is, only 7% mentioned autonomy and decision-making. One in five associated it with automation and task execution.
That has a practical consequence for whoever builds the corporate alternative. Much of what people call “my agent” is probably an assistant they ask questions, or a sequence of steps that repeats the same way every week. A fixed sequence doesn’t need an agent. It needs a workflow: a path defined in code where every step is planned in advance.
At Activeone, we have a rule for this: we don’t build an agent if a workflow solves the problem. A workflow is more predictable, can be audited step by step, and its cost per run can be calculated up front. An agent, which decides which steps to take and which tools to use, makes sense when the task genuinely changes with context. Even then, if you can’t explain what each run costs, it isn’t ready for production yet.
Seven steps to get started without buying anything
If your company is among the 85% showing interest but not yet in the 12% with mature adoption, this sequence lets you start with what you already have. Each step answers one question and comes with a warning sign that tells you to stop and review before moving on:
Step | Question it answers | Warning sign |
|---|---|---|
1. No-penalty inventory | Which personal agents does each team use today, and for what tasks? | If nobody reports anything, people probably don’t trust that the inventory comes without penalties. |
2. Classification by data | What information goes into each tool: customer, financial, code, personal data? | Sensitive data in tools nobody has evaluated. |
3. Comparison with the official option | Is there a corporate alternative for that task, and does it integrate with where people already work? | If it doesn’t exist or doesn’t integrate, a policy is unlikely to change behavior. |
4. Type of system | Does the task follow fixed steps (workflow), or does it require deciding what to do based on context (agent)? | Asking for an agent to do something that’s always done the same way. |
5. One-page policy | What’s allowed, with what data, what isn’t, and whom to ask. | A policy nobody can sum up in one sentence. |
6. Cost per run | How much does each task completed with the official alternative cost? | Scaling without knowing that number. |
7. Measurement | Has the use of personal tools for tasks involving sensitive data gone down? | Measuring only how many licenses were activated. |
The order matters. The policy comes in step 5, not step 1, because a policy written before the inventory regulates usage you don’t know about yet.
Where this doesn’t apply
Not every personal use is a risk. Someone who uses their own assistant to organize ideas, with no company data involved, doesn’t need a governance project, and treating that case the same as someone uploading a customer database undermines the policy’s credibility.
Offering a corporate alternative also has a cost: licenses, integration with existing systems, and someone to maintain it. For a small team, a list of evaluated tools and a clear rule about which data stays inside may be enough, and building more would be overengineering.
And it bears repeating: these numbers come from large companies and from what people report. Your organization may be doing better or worse, and the inventory in step 1 will tell you more precisely than any regional survey.
What does seem hard to dispute is that a policy that takes more effort than a personal agent will lose to it. The 44% have already chosen the easiest path; the company’s job is to make the official path that one.
Frequently asked questions
What percentage of professionals in Latin America use personal AI agents at work?
According to Work:InProgress 2026 (IDC and Provokers with Google Cloud), 44% of 3,200 professionals at companies with 500 or more employees in five countries use them regularly. These are agents their company didn’t provide.
Is a personal agent the same as shadow AI?
Not necessarily. The study asks about agents not provided by the company, not about prohibited tools. Some of that use may be allowed or simply unregulated, and it’s worth telling the two apart before penalizing it.
Why do employees prefer their own AI agents?
Among those who use one, 38% say it’s more accessible and easier, another 38% feel safer with it, and 36% mention privacy. In addition, 21% say the corporate tool doesn’t integrate with what they need.
Should you ban personal AI agents?
A ban without a more convenient official alternative tends to hide the usage, not eliminate it. It’s more effective to start with a no-penalty inventory, classify uses by the data involved, and offer an integrated alternative.
When should you use a workflow instead of an AI agent?
When the task always follows the same steps. A workflow is more predictable, auditable, and has a known cost; an agent makes sense when the task requires deciding what to do based on context.
Sources
Google Cloud, IDC, and Provokers. Work:InProgress 2026: Descubrimientos sobre cómo la IA está transformando el trabajo [Discoveries on how AI is transforming work], Spanish-speaking Latin America edition. In Spanish. Accessed November 19, 2026.