From AI pilot to daily use. Do you have the basics in order?

Many AI trajectories start with the visible side. A model, chatbot, or agent is chosen and a team starts a pilot. Only then does it become clear that the data is not findable, up-to-date, or consistent. A more powerful model does not automatically make those sources complete or reliable.

Veeam’s Data and AI Trust Gap Report 2026 shows the scale of this challenge. Only 7 percent of the surveyed organizations meet Veeam’s criteria for AI readiness. They have ambition, data visibility, and governance in order at the same time. Meanwhile, 95 percent say that data issues have slowed AI progress in the past year.

A pilot with carefully selected documents can work well. In daily use, different versions, access rights, and exceptions come into play. It is precisely then that it becomes clear whether the organization is ready to deploy the application more broadly.

Four roadblocks before the model

The first roadblock is that organizations do not fully know what data they have and where it is located. Information is scattered across SharePoint, Exchange, file shares, applications, and personal folders. Consider a service assistant that finds the general manual but misses the specific customer agreements in the CRM. The answer may be technically correct and yet unusable for that customer. Therefore, determine for each task which sources are needed together to provide a complete answer.

The second roadblock is quality. Duplicate documents, outdated versions, and contradictory registrations do not disappear with AI. A quote assistant using an old price list can repeat the same error in multiple quotes before someone intervenes. Data quality therefore requires ownership. Who determines which price list is leading? Who processes changes? And how do you ensure that expired information remains outside the sources used? These agreements partly determine how reliable the outcome will be.

The third roadblock is access. An employee does not automatically have the right to all information that is technically available. The same applies to an AI assistant or agent. If a folder with performance reviews is shared too broadly, an AI assistant can make that information more easily findable for employees who already had access. Inheriting existing rights therefore does not solve the problem of overly broad access. Check before linking whether those rights still fit someone’s role.

With an agent that is also allowed to modify data, an additional question arises. Which actions can it perform independently and when must an employee give approval?

The fourth roadblock is the lack of a measurable process goal. A chatbot that creates a summary within seconds does not yet prove that employees are saving time. How much time does it take to check and correct that summary? Therefore, choose a task with a clear owner and a baseline measurement. For example, measure the total handling time of a service desk ticket, including checks and remedial work. This makes it visible whether AI actually improves the process.

A practical start

First, map out one process and the associated data. Which sources are needed? Who keeps them up to date? Which information is sensitive? Which rights apply? And what result do you want to measure within the pilot?

Make that process manageable and test with real-world situations, including exceptions and different user rights. This is how you test whether the application is usable in daily work.

With the AI Readiness Scan from Comex, you get an initial picture of the data foundation, control, human application, and measurability of value within your organization in four minutes.