Roadmap: How to Start with AI in Your Company (Stages, Impact Matrix and a 30-Day Plan)

Everyone talks about AI and nobody tells you where to start. This is the path we recommend after more than a decade implementing data, automation and artificial intelligence in companies across Latin America.

Key takeaways

• The gap between using AI and capturing value is organizational, not technological: what is missing is usage frameworks, security rules, a plan and indicators.
• AI creates impact in three areas: the business model, processes and the product. The greatest potential sits in core functions, not only in the back office.
• Your own data and processes are the one thing no tool can provide, and the place where most projects get stuck.
• To choose where to start, use a two-axis matrix, impact and cost, with a measurable hypothesis for each initiative.
• A first case measured within 30 days is worth more than a transformation plan with no evidence.

Contents

1. Why doesn't using AI translate into value for the company?
2. Where does AI create the most impact?
3. What does applied AI look like? Three cases
4. What needs to be in order before automating with AI?
5. What are the stages of AI adoption?
6. How do you choose where to start? The impact and cost matrix
7. What should an AI usage policy include?
8. What can you do in the first 30 days?
9. How do you ensure security and control?
10. Questions we received from the webinar audience

In many companies, someone on the team already uses AI every day: drafting proposals, summarizing meetings, organizing data. Almost always they do it with a personal account, without anyone else knowing, and without the board being able to see what changed in the results. That scene, which we see in most of the organizations we work with, explains an uncomfortable fact: AI use keeps growing while visible value does not.

This guide lays out how to get out of that situation. It brings together what we shared in our webinar "Everyone talks about AI and nobody tells you where to start" (delivered in Spanish) and adds the tools we use with clients: a prioritization matrix, a usage policy checklist and a 30-day plan.

Why doesn't using AI translate into value for the company?

Because the gap is organizational, not technological. People already use AI, but companies lack usage frameworks, security rules, room to experiment and an adoption plan that connects those experiments to business indicators. Without that, individual use does not scale and the board sees no measurable results.

There are four typical signs:

1. Personal, uncontrolled use. People work with their own ChatGPT, Claude or Gemini accounts, and few have a company account, with the security risks that implies. It is usually fixed with a plan change, but it requires a management decision.
2. Licenses nobody uses. The IT department often reports that everyone has a license, and employees do not even know they have one. Other times there is a mismatch between the tool the company chose and the one people prefer, especially between the Microsoft environment and the rest.
3. Isolated pilots. A proof of concept driven by one person's enthusiasm works, but it never reaches the rest of the organization, and it is hard to justify why it was chosen.
4. A board that sees no impact. Nobody can show a number that links AI to the business.

The studies back this up. In a recent McKinsey survey, 70% of respondents said they feel personally prepared to use AI, but only 27% of leaders believe their organization is ready for the shifts an agentic future requires. People are moving faster than the institutions they work in.

The BCG study of more than 1,250 executives shows the consequences: only 5% of companies generate AI value at scale, 35% are scaling, and 60% report minimal gains. The companies that do capture value grow revenue 1.7 times faster than their peers.

In Latin America the picture is moving. According to Microsoft's report for the second quarter of 2026, generative AI use among people aged 15 to 64 is 25.9% in Colombia, 24.0% in Chile, 23.4% in Argentina and 21.4% in Mexico, against a global average of 18.8%. The region is in the middle of adoption, which is why setting a direction now is not being late.

This is not only an issue for large companies either. In small and mid-sized businesses that apply these initiatives successfully, we see even stronger returns, mostly through process efficiency, because they have less bureaucracy to test, decide and scale.

Where does AI create the most impact?

In three areas: redefining the business model, optimizing processes and improving the product or service the customer sees. Most companies start by optimizing support functions, which is reasonable. The greatest potential, however, is in core functions, and it pays to get there sooner than seems natural.

1. Redefine the business model. AI makes it possible to serve demand that could not be served before for lack of capacity. An e-commerce company that only answers during business hours can add an assistant that responds about products and stock at any time and captures people who buy at night. A company that never bid on tenders because it had no one to analyze every call can now process them all and choose which ones to pursue. The question that guides this area: what demand are we not serving today because of lack of time, people or budget?

2. Optimize what you already do. Automate repetitive tasks, augment people with assistants and redesign entire workflows. This is where almost everyone starts, for good reason: you know your processes and your pain points. The back office and administrative tasks are strong candidates for early results. The question: which repetitive task consumes the most hours of valuable people?

3. Improve the product or service. Assistants, personalization, predictive services. Excel is not a new product, but an assistant that writes the formula you need has lowered the barrier to entry for a huge number of people: what matters today is what you need to achieve, not how a formula is written. Every customer touchpoint can be improved this way, including how long it takes to answer a request. The question: at what point in the customer experience is there friction AI could remove?

A note on the core. Support areas are comfortable places to start: there is less risk and savings show up quickly. The mistake is staying there. According to BCG, about 70% of AI's potential value is concentrated in core business functions. The functions that move the company and set it apart from competitors are the ones that pay off the most.

What does applied AI look like? Three cases

In the three cases below, the focus is on a specific process rather than on the tool. In each one, AI does the reading, classifying or analysis, and a person validates what matters. They are repeatable patterns, along with what we learned from each.

Case 1: invoice processing

• Context. A company processed thousands of invoices per month with an administrative team that spent a large part of its time on that task. According to the process manual, not every invoice needed the same level of control. In practice, everyone reviewed everything.
• Solution. An automated workflow built in n8n, an open source automation tool that can be hosted on your own server or in the cloud. The workflow reads each document, extracts the key data and assigns a confidence score to the reading, because a PDF generated by ARCA (Argentina's tax authority) arrives clean, while a blurry photo of a scan does not. Final validation stayed with a person, using a small platform to approve, reject or correct each extracted field. Integration with the management system was added later.
• Result. Turnaround times improved dramatically and manual work shrank to a fraction. The people who stopped keying in and checking invoices were not laid off: they were reassigned to key account management and other processes.
• Lesson. The confidence score avoids reviewing everything and also avoids trusting everything. The same pattern works for contracts, tenders and emails. A process like this also forces a reorganization of the team: new tasks appear and roles need to be redesigned.

Case 2: scenario simulator

• Context. Commercial proposals and operational decisions that were built from familiar prototypes and comparable cases.
• Solution. A simulator that makes it possible to customize each offer, project changes with more variables and score incoming proposals to decide which ones deserve attention.
• Result. Fewer hours spent on prospecting and forecasting, and better use of focus: human judgment concentrates on the cases that matter, and the rest can be delegated to another part of the team.
• Lesson. Solving one case is simple. The challenge is building a tool that works for many different cases. It is better to start with one specific case and generalize afterward.

Case 3: assistant over your own documents

• Context. During Greencode's ISO 9001 certification, documents, open items and audit comments had to be kept up to date as they constantly changed.
• Solution. An assistant connected to Slack that answers using the company's own documents and updates as new material is added. It uses an architecture called RAG (retrieval-augmented generation): the AI answers from the content of your own documents, not only from general knowledge.
• Result. The team could talk to its own information instead of waiting for the next meeting with the consultant to know what to do, and the assistant helped maintain the standard.
• Lesson. It is a simple pattern, with its own challenges, and useful in any area with living documentation: quality, legal, HR, technical support.

What needs to be in order before automating with AI?

Data and processes. Models are available to everyone; what differentiates a company is its data and its way of working, and AI does not generate either. Most projects stall because of dirty data, processes without an owner or a lack of indicators, not because of technology limits.

The most frequent blockers:

• Duplicated, incorrect or scattered data across systems that cannot be reconciled.
• Processes that each person runs their own way: you cannot automate "the process" when there are seven versions of it.
• Processes without an owner: nobody decides and nobody takes responsibility for the result.
• No indicators, and therefore no way to know whether efficiency improved.
• Instructions from leadership like "use AI" with no objective. More efficient for what, measured against what?

A very proactive team can compensate for all of this, but it is not wise to depend on goodwill. You need a plan, indicators, initiative and communication.

Five steps to get your data in order

1. Identify. What data does the business really need? Which of it already exists and which is missing?
2. Capture and organize. Where is it stored? Is it sensitive and does it require encryption? Does it update and sync on its own or by hand?
3. Analyze. Who sees which dashboard? Is there a single source of truth, or does each area have its own number?
4. Get insights to the people who decide. This is where AI shines. A dashboard gives a snapshot; AI helps interpret it and monitor variables that change often. But a dashboard only directors look at is useless if the salesperson deciding on a discount does not know whether stock is running out or piling up. Data has to reach operations. Agents work very well here, as long as the data is good.
5. Governance, from the first minute. Who owns each piece of data? Who has access, at what quality and for how long? It is not a final step: it runs through all the others, and it is a leadership topic, not only a technical one.

Signs a process is not yet ready to be automated

• Two people on the same team describe it differently.
• Nobody can be named as responsible for the outcome.
• There is no indicator showing whether it works well.
• The data it consumes lives in personal spreadsheets.

If any of these apply, the first project is to organize the process, not to automate it.

Work changes: from writing to verifying

When AI enters a process, people's work changes. Someone who drafts commercial proposals with AI moves from writing to verifying, and verifying well is a skill that has to be developed. If the search for new projects is automated, the bottleneck stops being finding them and becomes analyzing them. That is why handing out tools is not enough: redesigning the workflow is what captures value. Without that redesign, something predictable happens: you pay for the current team plus the automation.

What are the stages of AI adoption?

Enable, automate and reinvent. First you provide tools and training; then you automate processes and bring agents into decisions; finally you redesign the operating model. Each stage demands more solid data and processes than the last, and few organizations reach the third.

It is a reading similar to the "three horizons" McKinsey describes, and it helps you place yourself.

Stage 1: Enable

• What you do: provide tools and training; bring individual use into order.
• What you need: a usage policy and training by level.
• Typical risk: isolated pilots and unused licenses.
• Sign of progress: everyone speaks the same language.

Stage 2: Automate

• What you do: automate workflows and bring agents into decisions.
• What you need: clean data, processes with owners and indicators.
• Typical risk: automating a poorly defined process.
• Sign of progress: some initiatives show measurable value.

Stage 3: Reinvent

• What you do: new roles and workflows; a redesigned operating model.
• What you need: solid data and processes; aligned leadership.
• Typical risk: reinventing without a foundation and without evidence.
• Sign of progress: the new roles and workflows are already in place.

Enable is the most solid step. Directors, middle managers and operational teams should share a common language; otherwise they live in different worlds, where some say the tool is already too small for them and others say it still hallucinates. Middle managers and leadership need training more oriented to strategy than the operational team does. We lived this at Greencode: when we started training formally, we realized that without it, moving on to automation was very hard, because the culture was not ready to transform the teams.

Automate includes workflows and agents that start taking part in decisions. In our experience, this is where roughly a third of companies start to perceive value. Reinvent is where the greatest impact shows up: at that horizon, more than half of organizations report growth, although few get there.

When do you move from stage 1 to stage 2? A clear signal is when the AI or technology team receives more initiatives than it can handle. One company we spoke with told us things were getting out of hand: every team proposed projects and nobody could review them all. That is when a strategic plan stops being optional.

How do you choose where to start? The impact and cost matrix

By placing each initiative according to two questions: how much it costs to implement, in money, time and change, and what measurable impact is expected for the business. High impact, low cost initiatives are the quick wins to start with; those with high cost and unclear impact are reconsidered.

Step 1: estimate the cost

Each team estimates the cost of its initiative, and not only in money. Include:

• Implementation time.
• Effort from the team that will use it.
• Changes needed in processes and systems.
• The cost of moving from a prototype to something the whole organization uses.

That last point is key. Testing something on a laptop is fast, and it even makes the person doing it feel like a programmer. Offering it to the whole organization or to a customer is an enormous leap, and Finance will ask for that number. An initiative that cannot be priced stays a good intention.

Step 2: define the impact hypothesis

Every initiative needs a measurable hypothesis. A simple template:

"Today [process] takes [X] or costs [Y]. We expect that with AI it will go to [X'] or [Y'], and we will measure it with [indicator]."

If you cannot complete it, you are playing with the tool, and there is too much to do to spend time on that.

Step 3: place each initiative in its quadrant

• High impact, low cost (quick wins). For example, invoice or documentation assistants. This is where to start.
• High impact, high cost (strategic bets). They usually require board approval and take up next year's agenda. Plan them with their own budget.
• Low impact, low cost (incremental improvements). Report summaries, dashboards: they are a couple of prompts away. Do them when there is time.
• High cost, unclear impact. Initiatives whose return cannot be measured. Reconsider them; do not start here.

An illustrative example. A distribution company evaluates three initiatives: an assistant that reads supplier invoices, a customer service chatbot and an automatic summary of sales meetings. The invoice assistant has a contained cost and an impact measurable in hours per month: a quick win. The chatbot costs more and touches the customer, but it can open a new channel: a strategic bet. The meeting summary is fast and useful, but it does not move the business: an incremental improvement. The natural order is invoices first, the chatbot with its own plan and budget, and summaries whenever there is time.

A recommendation for the discovery process: leave room for open brainstorming. The initiatives that appear first tend to be about efficiency, because each area works on what it knows, and that is fine, but they are not necessarily the most interesting ones. Without that room, innovations get lost.

What should an AI usage policy include?

One page is enough if it answers seven questions: which tools are approved, what information is never uploaded, which uses need no permission, how output is verified, where ideas are proposed, who decides when in doubt and how often it is reviewed. Its purpose is for teams to know how far they can go.

There are two opposite mistakes: banning everything out of fear, or accepting anything out of carelessness. Both are costly. In addition, some teams have very good ideas and never propose them for fear of overstepping, and the company loses proofs of concept that could scale. The seven questions:

1. Approved tools, always on company accounts.
2. Information that is never uploaded to an external tool: customer data, contracts, sensitive financial information.
3. Permitted uses that need no permission, and uses that require approval.
4. Verification of what AI produces before it is used with a customer.
5. A channel to propose new ideas and a way to evaluate them.
6. Who decides when there are doubts.
7. Review frequency, because the context changes fast.

It is worth adding adoption indicators from the start. Measuring is what leaves evidence of what works and what does not, and there is still no golden manual: there are very good recommendations, but each organization has to test and adjust them. It is a continuous improvement process.

What can you do in the first 30 days?

Survey, prioritize, set the rules and launch pilot cases. The realistic goal is not to transform the company in a month, but to end it with two to four prioritized initiatives, owners and indicators defined, a published usage policy and a measured baseline for each case.

Week 1: survey.

• Each area lists its initiatives. We do it with two to three hour workshops per area, with asynchronous work beforehand so the workshop can be spent prioritizing. IT does not need to take part.
• Take stock of which tools and licenses exist, who uses them and who does not.

Week 2: prioritize.

• Estimate the cost of each initiative and complete the impact hypothesis.
• Build the matrix and choose between two and four initiatives.
• Assign an owner and an indicator to each one.

Week 3: set the rules.

• Write the first version of the AI usage policy.
• Define the training plan, with different content for leadership, middle managers and operational teams.

Week 4: launch and measure.

• Launch the pilot cases and record the baseline for each indicator.
• Define when progress is reviewed and who presents it.

Few initiatives, well measured. They build momentum and culture, and provide success stories to share with the rest of the company.

Common mistakes when starting

• Buying licenses before defining uses.
• Starting with what looks impressive instead of what is measurable.
• Automating a process nobody defined.
• Leaving everything in the hands of one enthusiastic person: if they leave, the pilot leaves with them.
• Not measuring the baseline, and then being unable to prove the improvement.
• Staying in the back office forever.
• Postponing the usage policy until an incident happens.

How do you ensure security and control?

With providers whose compliance can be verified and with corporate plans, with local AI when sensitivity requires it, and with agents that are traceable, have their own identity and have an off switch. These criteria let you move forward without slowing teams down and answer the first objections from IT, legal and compliance.

Data security deserves a full webinar of its own, but there are basic criteria:

• Providers with verifiable regulatory compliance, corporate plans and integration with Microsoft 365, Google Workspace or your user directory, so you can control who has which license.
• Local AI when the sensitivity of the data requires it: it can run on your own server or in the company's cloud, and behaves like one more piece of software inside your infrastructure.
• Traceable and observable agents, with an identity of their own that is distinct from the person who created them, so you can tell whether an action was taken by that person or by the agent they programmed. A third-party solution that leaves no record of what it did should not be used.
• A red button on every automated process, meaning a way to shut it down. It sounds obvious, but many processes today do not have one, and it is mandatory.

Questions we received from the webinar audience

How do we move from using AI as a lookup tool to really building it into the business, and how do we choose which processes to invest in?
With an initial survey: list processes and initiatives by area, estimate their cost, define the impact hypothesis and place them in the matrix. It can be done with your own teams or with a facilitated workshop. And leave room for creativity, in addition to efficiency.

How do we get the whole team working with AI, and how do we support middle managers?
First, train, so everyone speaks the same language. Then work with indicators and process owners. For example, in a sales area the useful question is not how people should sell, but what counts as a good sales process and how you know a meeting went well. With those indicators, middle managers can invite their teams to propose how to improve them with AI, and that brings objectivity.

How do we move from isolated initiatives to an integrated plan that connects with our systems and is secure?
With a strategic plan, a usage policy, adoption indicators and clear security criteria, like those in the sections above. Integration with systems comes afterward, on top of processes and data that are in order.

Does this apply to small and mid-sized companies too?
Yes, and in many cases with more of an advantage: less bureaucracy to test and scale, and an efficiency impact that is very visible from the first cases.

Three ideas to take away

1. The gap is organizational. People already use AI. What is missing is the company integrating and measuring it.
2. The value is in redesigning processes, not in handing out licenses, and in going after the core without staying in the comfortable place.
3. Few initiatives, well measured. They build momentum, create culture and provide success stories for the rest of the organization.

Next step

If you want to bring these ideas to your organization, at Greencode we support you from the initial survey of initiatives through implementation. Get in touch and we will set up a conversation, no commitment.

If you would rather start with the enable stage, you can learn about our AI Masterclass (4 hours, in person or virtual) and our hands-on workshops by area (2 hours, 30% theory and 70% guided practice), where each team finishes with a real project applied to its own work.

About the author. Ezequiel Lamónica is co-founder and director of Greencode Software and a professor at IAE Business School and Austral University, always in artificial intelligence and its application in business. Founded in 2014, Greencode has delivered more than 100 projects and operates in Argentina, Mexico, Chile, Uruguay, Colombia and the United States. Ezequiel on LinkedIn

Sources. McKinsey, From adoption to impact: Three horizons of AI transformation · BCG, Are You Generating Value from AI? The Widening Gap · Microsoft, AI Diffusion Report, Q2 2026

‍