By Gabriel Simonet | CMO at GeneXus, a Globant Company

At this year’s Partners Meeting, five different talks arrived at the same conclusion, each from a different angle: The technology already works. What determines the success of an AI initiative is the organizational structure built around the models.

As we compared the experiences shared, these nine challenges kept coming up.

1. The bottleneck moves; it doesn’t disappear

When a development team adopts AI and becomes ten times more productive, the bottleneck shifts to another link in the chain: marketing, legal, compliance or training. These functions are organized around a pace that has already changed. I spoke about this because I’ve seen it firsthand in many companies, and it connects with something Andrew Ng has been warning about for some time: friction now sits between teams rather than within them.

2. The stakeholders are changing

Business teams are taking a greater role in the conversation alongside IT. Gisela Bertelli shared the case of an airport, where the digital transformation sponsor works in legal, something that would have been unthinkable three years ago. Business functions feel the urgency, understand the impact and lead adoption. IT becomes the partner that orchestrates the work.

3. There is a real gap between prototype and production

Virginia Paternostro talked about how she built a working sales bot with Claude in fifty minutes. The final stretch was deceptively difficult, though: getting it into production, integrating it with HubSpot, ensuring stability and meeting audit and data security requirements took weeks and a technical expert. Building the agent is the easy part. Building the plumbing that lets it operate in production is where the real work happens.

4. Designing autonomy and human oversight

Gaston Milano described the architect’s role as deciding what agents can do on their own, what requires approval and when a person should intervene. As organizations delegate more work to agents, they need to make those boundaries explicit: who is responsible, which decisions can be automated and how results are checked. The level of autonomy becomes a design decision for each part of the process.

5. Turning individual work into shared knowledge

As people work with agents, they generate solutions, decisions and lessons that can stay buried in private conversations and personal workspaces. Organizations need ways to make that work visible and reusable across teams. Otherwise, everyone gets faster individually while teams keep solving the same problems in isolation.

6. Connecting agents to business context

An agent is only as good as the context it can access. Eugenio García highlighted data integration as a central challenge: connecting documents, APIs, databases and scattered data lakes so agents can work with real business information. Gastón Milano added another dimension with company as code: documenting and versioning the policies, processes, roles and decisions that guide the business. Together, these require reliable connections, clear access permissions and a shared source of truth that people can maintain and agents can use.

7. Avoiding platform lock-in

It’s easy to build on the first platform that solves the problem. But anything built in a specific environment brings its own harness: the integrations, patches and configurations needed to make it work there. With the “best model” changing every few weeks, moving an agent to different infrastructure becomes difficult if portability wasn’t considered from the start.

8. Planning in the fog

The pace of change is so high that even the team building the platform can only project its roadmap a few releases ahead. The context, model capabilities and rules of the game change before you reach the finish line. The ability to adapt quickly is what wins.

9. Measuring business impact

Adopting AI is a means to a business outcome: moving from individual productivity to orchestrating complex use cases, and translating that capability into measurable impact, such as lower costs, faster time to market or greater customer satisfaction. Becoming faster as an organization is a deliberate decision.

And also, the human dimension

These nine challenges also have a human dimension: loneliness in the age of AI. In studying how Anthropic and OpenAI’s Codex team work, we saw both the risk of losing interaction with colleagues and the value of creating spaces where people can meet and solve problems together.

When a question we once asked a colleague gets answered by an agent, we may gain speed and lose an opportunity to learn from each other, discuss how to make a judgment or discover something nobody had thought of.

If each person can solve more on their own, we need to create reasons to come together: review a prototype, share what we’ve learned or ask for another perspective on a decision. Part of leading this transformation is nurturing those relationships and turning individual capability into shared learning and a culture of collaboration.

We need to create spaces where working with AI helps us work more closely together.

You may also be interested in reading:

GeneXus and the Power of Knowledge

Does GeneXus for Agents Work? Here’s What Its Users Think

GeneXus for Agents: what it is, how it works, and how to get started