Getting an AI agent to work in a company’s day-to-day operations is proving considerably harder than many anticipated. The demo shines, but when the agent lands on the real CRM, email, and ERP, the problems begin: duplicate data, permissions, processes that step on one another. That is precisely the bottleneck when it comes to training AI agents that are genuinely useful in a real enterprise environment, and not just on a presentation screen.
A recent piece of news illustrates this well. According to TechCrunch, the startup Arga Labs has raised a $10 million seed round to build "training environments" that reproduce enterprise software such as Salesforce, Workday, or the email client. Instead of a simple API access point, they create a complete replica of the program—a kind of digital twin—with its permissions and automations intact. The idea: to be able to rehearse thousands of times without breaking the real system.
Why training AI agents in the enterprise is so difficult
The article explains a very revealing contrast. AI tools for programming have advanced at full speed because the world of code already has mature instruments to deploy, roll back, and analyze changes. That makes it possible to set up test environments and repeat a scenario as many times as needed.
With management software, that is not the case. There is no "reset" button to return your CRM or your email to the exact starting state and relaunch the same test. And without that ability to repeat, reinforcement learning—rehearsing a task tens of thousands of times and keeping only the strategies that work—becomes unfeasible.
Arga’s CEO, Phillip Li, grounds it with an example any sales team will recognize: a potential customer enters as an opportunity in one system, while a colleague contacts that same company through a different tool. Is the agent able to realize it is the same company? To check that the email hasn’t already been sent twice? To decide who to write to between the two open opportunities? That ambiguity, so common in any office, remains slippery ground for agentic systems.
From a startup’s lab to an SME’s day-to-day
It’s easy to read this news as something distant, a matter for big tech companies with million-dollar rounds. But the underlying idea is very down-to-earth and affects any company considering using AI on its processes. As Yuri Sagalov of General Catalyst summarizes in the same article, much of the economic value of agents arises precisely from using business applications, and for that a repeatable test environment is needed.
At Tisa we see it with our clients: the interesting conversation is not "can AI draft an email?" but "can the agent reliably operate on my systems, with my data and my rules, without messing things up?". And the answer depends less on the trendy model and more on three utterly unglamorous factors: orderly data, clear processes, and a safe space to test before releasing the agent into production.
1. Orderly data: no foundation, no agent
The example of the duplicate customer is not an AI problem, it’s a data problem. If in your organization the same company appears in three different ways across three systems, no agent—however powerful—will make good decisions. That’s why we insist so much on data governance: quality, common criteria, and a single source of truth. It’s the preliminary work that makes any subsequent AI project make sense.
2. Clear processes and a place to rehearse
An agent learns (and is validated) by repeating. For that it needs an environment where it can make mistakes without consequences: test, measure, correct, and test again. Translated to an SME, this means having test environments of your ERP and your applications, with representative data, separated from real operations. You don’t need a lab-grade digital twin: you need method and a platform that enables it.
3. Start with a measurable use case
The temptation to "add AI" to everything at once is real, and it usually ends in frustration. We advocate exactly the opposite: choose one specific use case, with a clear and measurable benefit, and execute it from start to finish. An agent that reconciles duplicate opportunities, another that prepares responses to recurring incidents, another that reviews orders before invoicing. Start small, prove value, and scale on what works.
How Tisa grounds it
This is where the ecosystem we work in plays in our favor. Microsoft Dynamics 365 Business Central and Power Platform offer something that many standalone tools don’t have: separate test and production environments, permission control, traceability, and native automation with Power Automate. In other words, part of the "safe space to rehearse" that startups are rebuilding from scratch already exists when your management runs on a serious and well-integrated platform.
On that foundation, from Tisa’s custom development and AI applied to business lines, we accompany private companies on the realistic path: organize the data, define the process, set up the appropriate test environment, and develop the use case that truly moves the needle. With more than 30 years implementing management solutions in retail, distribution, construction, food industry, hospitality, sports centers, or automotive, we know the processes where an agent can add value—and also where caution is still advisable.
The practical conclusion
Arga’s bet confirms an underlying trend: the next wave of AI is not played out solely on smarter models, but on the ability to train AI agents on companies’ real software, with their data, their permissions, and their ambiguous cases. It’s a race currently led by the big players, but whose lessons are applicable to any SME: first order, then process, then AI.
If your company is mulling over intelligent automation or a first AI project, the best starting point is not choosing the model, but putting in order the foundation on which it will work. At Tisa we help you assess it with no obligation: write to us at info@grupotisa.com, call us at (+34) 971 305 885, or visit grupotisa.com. We’ll talk about your specific case and how to take a first step with sound judgment.
Reference source: TechCrunch, "Arga Labs is building a better way to train enterprise AI agents" (August 26, 2026).