How to implement artificial intelligence in your SMB without wasting money

Most artificial intelligence projects that do not work in an SMB do not fail because of the technology. They fail because they start with the tool — "we need a chatbot" — instead of starting with a concrete problem someone can measure. The result is usually a subscription nobody uses or a pilot that never reaches production.

This guide is practical: how to choose the first use case, what to expect from an AI agent, when to buy and when to build, what Colombian law requires about personal data and how to know whether the investment was worth it.

If you want to see the kinds of solutions we implement, check our page on AI for business.

Start with a measurable process, not the tool

The best first AI project is boring: a repetitive, high-volume process based on text or documents that today consumes hours of valuable people. Before touching any tool, measure how that process works today. Without that baseline you will not be able to tell whether it improved.

• It repeats many times a week and follows more or less clear rules
• It involves reading, classifying, summarizing or answering text
• It has an owner who suffers the problem today and will use the solution
• You can measure it: time per task, volume, errors or response time
• An AI mistake is detectable and correctable before it causes harm

First use cases that tend to work

FAQ assistant grounded in your documents

An assistant that answers customers or your own team using your manuals, policies, product sheets and FAQs. The technique is called RAG (retrieval-augmented generation): before answering, the system retrieves the relevant fragments of your documents and the model writes the answer based on them. This reduces the risk of it making things up, and you can show the source of each answer.

Data extraction from documents

Supplier invoices, purchase orders, delivery notes or forms that someone types into the system by hand. AI can read them, extract the fields and leave them ready for a person to review and approve. Human review still exists, but it shifts from typing to verifying.

Email and lead classification

Sorting the inbox by request type, urgency or responsible area, and scoring prospects based on what they write. It is a low-risk case with visible results within a few weeks.

Order follow-up

Answering "where is my order?" by checking the real status in your order system or with the carrier, over WhatsApp or chat, and escalating to a person when there is an issue. It requires integration with your systems, which is why it usually adds more value than a generic chatbot.

What an AI agent is (and what it is not)

An AI agent is a language model with a goal, access to tools and defined permissions: it can check your inventory, create a task, draft a reply or update a record, deciding in which order to do it. The difference from a chatbot is that it does not just converse, it also acts on your systems.

That is why an agent needs clear limits: what it can do on its own, what requires human approval and what gets logged. An agent that can issue credit notes or change prices without oversight is a risk, not an improvement. At Julsmind SAS we apply this approach in Opttia, our multi-agent platform for supply chain intelligence, where specialized agents reach recommendations by consensus, and in KAI, the AI assistant built into Katuq, which works on each company's order and inventory data.

If some of these terms are new to you, our glossary explains RAG, agents, LLMs and other concepts in plain language.

Build or buy

Buy a ready-made tool when the use case is generic (transcribing meetings, drafting text, summarizing documents) and the tool already integrates with what you use. Build when the solution needs your data, your business rules and connection to your systems: your ERP, your CRM, your inventory. That is where a generic tool falls short.

There is a middle path that is usually the most sensible: use models from established providers via API and build on top only the layer that is yours, that is, the integration with your systems, the rules and the controls. That way you do not reinvent the model and you are not tied to a tool that does not understand your operation. Whichever path you take, ask who owns the data, the prompts and the integrations if you switch vendors tomorrow.

Privacy and personal data: what the law requires

In Colombia, the processing of personal data is governed by Law 1581 of 2012 and its regulatory decrees. If your AI solution processes customer, employee or patient information, the same rules apply as for any other system: the data subject's authorization, an informed purpose, a published processing policy and security measures. Health data is sensitive data, with stricter requirements: explicit authorization and special care over who accesses it and why.

Many AI models run on servers outside the country, which may involve an international data transfer. Before connecting personal information to an AI service, check:

• In which country the data is processed and stored
• Whether the provider uses your data to train its models and how to opt out
• How long it retains information and how it is deleted
• Whether you can anonymize or minimize the data before sending it
• Whether your processing policy and authorizations cover this new use

How to measure whether it worked

Define the pilot with a start date, a cut-off date and a decision criterion agreed in advance: what result justifies scaling, what justifies adjusting and what justifies stopping. The metrics depend on the case, but they are usually along these lines:

• Time per task before and after
• Share of queries resolved without human intervention, and the quality of those answers
• Extraction accuracy compared with human review
• First response time to customers or prospects
• Cost per task, including API usage and supervision hours
• Real adoption: how many people use it each week

Common mistakes that make everything more expensive

• Buying licenses for the whole company before testing a concrete case
• Launching an assistant without up-to-date documents to learn from
• Giving an agent write permissions without human approval or logging
• Not measuring the baseline and then being unable to prove the result
• Ignoring data protection until a customer asks
• Treating AI as a one-off project instead of a process that gets tuned with use

AI in an SMB does not have to be expensive or complicated to add value. It has to be concrete: one process, one metric, one owner and a pilot with a cut-off date. From there it is much easier to decide the next step.