AI in your company: What you need to know before getting started

Artificial intelligence (AI) promises to automate tasks and improve efficiency, but before considering models or licenses, there is a preliminary step that often determines whether the project will succeed or fail: compiling the data and setting up the internal structure.

Carles Abarca, Vice President of Digital Transformation at Tecnológico de Monterrey, said that although AI sometimes produces surprising results, it is nothing more than an algorithm.

“The accuracy and quality of the results depend entirely on the data we provide it with,” he adds.

Along the same lines, Tec professor Adolfo Ernesto Arroyo points out that data compilation is “fundamental” for a model to work: “It’s the real bottleneck we have in models.” 

CONECTA presents the key points concerning how to use artificial intelligence (AI) in companies and businesses.
 

A clear, measurable goal

The first step is not to choose ChatGPT, an agent, or “trendy” software, but to define what problem you want to solve: customer service, predictive maintenance, sales, financial risk, inventory, or analytics.

“The big mistake is to jump on the bandwagon because it’s the fashionable thing to do, not because it’s necessary. Abarca explains that the first thing you have to identify is the problem you want to solve.

Quick checklist:

• What area do you want to improve and why?
• How will you measure success? (time, cost, errors, satisfaction, revenue)
• What indicator will show that it did, in fact, work?

Defined processes: AI might replicate disorder

According to Abarca, many organizations assume that AI fills in information gaps in the same way a person does, but that’s not how it works. If the process is incomplete or poorly documented, AI will only replicate that disorder.

“Artificial intelligence will not be creative. If we don’t have good process definition and data quality, the project is doomed to failure,” says Abarca.

Before working with AI, define:

• Actual process flow (from start to finish)
• Inputs and outputs for each step
• Exceptions (“What if…?”)
• Managers assigned to each stage

Ready data, not just “lots of data”

For Arroyo, the change lies in shifting from volume to quality with useful data for use cases. One of the common mistakes is to believe that it is enough to “have a lot.” 

“Entering data today is very easy; what matters is quality and that it is well regulated for the use case,” he explains.

Signs that your data are NOT ready:

• Duplicated entries (the same client appears more than once)
• Inconsistencies in names/addresses
• Incomplete or non-standard data
• Lack of metadata (source, date, responsible party)

Cleanliness, order, and minimal data traceability

According to the professor, preparing data means that the company can answer basic questions: Where does this data come from? Who generated it? When was it updated? What system does it reside in?

Arroyo summed up: “Without a well-structured, auditable, well-defined dataset, AI will perform like a poorly calibrated instrument,” he assured us.

Recommended action:

Create a minimal data classification structure.

• Data catalog (what exists and where)
• Common definitions (same meaning throughout the company)
• Required fields by system
• Quality indicators

Close technological gaps before scaling AI

If the company still relies on manual or paper-based processes, or its digital infrastructure is fragile, AI will not be scalable.

Abarca explains that this is one of three typical mistakes: attempting AI when there is “a huge technical debt” and processes have not been digitized or standardized.

What to review:

• Which processes are still done manually?
• Which systems are not interconnected?
• How many tasks depend on manual data entry?

Detect and correct data outside the system

One of the most common risks is that “real” information lives outside official systems: spreadsheets, macros, personal databases, or files on laptops.

For Abarca, this is the most common problem: “Since I can’t find an answer, I download everything to Excel... but that Excel file is already outside the scope of management and will no longer be taken into account when designing the AI.”

Key actions:

• Identify “critical” reports made with Excel.
• Find out who updates them and which sources are used.
• Gradually migrate to institutional sources.

Data governance: Roles, permissions, and traceability

Governance is what prevents data from becoming a “no man’s land.” For Arroyo, it is not a technical, but an organizational issue: responsibilities, processes, and limits are defined.

“Data governance is not a technical issue, it is an organizational institution, formal roles must be defined,” he explains.

Roles he says are necessary:

• Data Owner (responsible for quality/use/security)
• Data Steward (daily quality and access policies)
• MLOps (model lifecycle management)
• Legal and privacy issues

Documented architecture: Knowing what systems exist and how they interconnect

Abarca explains that good architecture must consider “what information is on which servers,” what the primary sources are, and how applications relate to each other, in addition to formal connections via APIs.

“Organizing architecture” begins with something very specific: documenting what already exists.

Minimum requirement:

• Systems and data flow maps
• “Golden” sources (the official truth)
• Stable integrations (no file exchange)
   

Security and compliance: Encryption, anonymization, and access control

When AI connects to internal or sensitive data, the risk increases: leaks, fines, or misuse. Arroyo warns that failing to comply with legal frameworks may have serious consequences.

“Legal frameworks help regulate this, but internally we must have technological practices in place: encryption, access control, and anonymization,” he points out.

Minimum requirements for getting started:

• Role-based access (not “everyone sees everything”)
• Classified sensitive data
• Change audit (who moved what)
• Clear retention and deletion policy

Executive sponsorship and change management

Although the focus is on data and structure, the human factor defines its adoption. Abarca explains that an AI project requires involvement from key areas and support at the highest levels.

“Sponsorship is needed; it has to be a decision made at the highest executive level; there’s a very important human and change-management component,” he says.

What internal preparations to make:

• Project team (business + data + IT)
• Basic training by role
• Clear communication: what will change and what will not

Events for building digital foundations

Beyond specific tools, implementing AI has become a matter of competitiveness and transformation for organizations. 

In this context, Tecnológico de Monterrey has promoted spaces for dialogue and learning where experiences, challenges, and case studies are shared to bring technology to real company applications.

An example of this is incMTY 2026, Tec’s entrepreneurship festival, which will focus on artificial intelligence and the use of technology to spark innovation in businesses and startups.

According to experts, these types of events open up the conversation not only about “what AI can do”, but also about which areas many companies need to improve: reliable data, clear processes, and an internal structure capable of sustaining long-term projects.

READ MORE: