What AI does well today
Today's AI, the kind built on language models, is very good at one type of task: working with text and messy information. Specifically:
- Reading and summarizing: contracts, long emails, documents, manuals.
- Classifying: sorting inquiries, organizing an inbox, tagging by topic or urgency.
- Drafting: emails, replies, reports or descriptions.
- Extracting data: pulling the fields you need out of a text or PDF, like dates, amounts or names.
- Conversing: answering in natural language over a base of information you provided.
They all share something: there's a text in and a text out, and a person can review the result before it has consequences. When the task looks like that, AI usually works well from day one.
The rule: if the result gets reviewed and can be corrected, AI already works. If the result runs on its own and a mistake is expensive, it has to be designed with far more care.
Use cases that work today
These aren't theoretical ideas: they're things already working or in advanced development in real projects.
An assistant that answers with official information. In TurisRed, the municipality loads its tourism information and visitors ask questions in natural language. The assistant answers only from that information, with guardrails: pinned answers for sensitive topics, blocked subjects and conversations the municipality can review. The AI doesn't make things up: it answers with what it was given and stays within bounds.
An agent that sorts messy messages. In CertiCarga, drivers report statuses, locations and expenses by chat, each in their own way. An agent interprets those messages and turns them into recorded data, without anyone typing them in. It's the kind of task where AI shines: turning free text into structured information.
Drafts that save hours of writing. In PsicoAssist, the professional enters the patient's data and the AI produces a report draft that the person then reviews and signs. The AI doesn't diagnose or decide: it assists. It's an example of using AI in a sensitive area without handing over control.
Practice with personalized feedback. In Work English, the AI builds real-work situations and corrects grammar, vocabulary and tone on every message. What used to need an available teacher now happens on the spot, as many times as you want.
In all four cases the pattern is the same: the AI does the heavy, mechanical work (reading, drafting, sorting, answering) and a person keeps control of what matters. That's where AI pays off.
Where it doesn't make sense yet
For a project to go well, you have to be honest about what AI still doesn't do well:
- Decisions involving money or health without human review. AI can propose, but responsibility has to stay with a person.
- Exact calculations and critical numbers. AI can get a sum or a total wrong; that's what deterministic software is for.
- Tasks with no data. If you haven't loaded the information, AI makes it up. It isn't a truth engine: it answers with what it has, and fills the gap when it doesn't.
- Processes that keep changing. Get the process stable first; otherwise AI just automates the mess.
- Replacing a professional's judgment. It assists, it doesn't replace.
It's also worth being wary of "let's do everything with AI". Most of the time the right answer is AI in one specific step of an existing process, not an end-to-end AI product. And often, for a task, a simple rule works better than a model.
There's one case to avoid: using AI as a search engine for data you already have in order. If the answer lives in a database, a plain query is faster, cheaper and more reliable than a model. AI adds value with messy text, not with a table.
How to start with a small pilot
The most expensive mistake is starting with the biggest thing. I recommend the opposite: a small pilot, with a concrete task and a way to measure whether it worked.
- Pick a repetitive, narrow task. Not "all of support", but "classify the inquiry emails".
- Define what a good result is before you start, not after.
- Load the information the AI needs: documents, data, examples of correct answers.
- Put a human in the loop at first, especially if customers are involved.
- Measure for two to four weeks and decide: scale, adjust or drop it.
A small pilot takes weeks, costs little and settles the question without risking the whole operation. And if it works, you already have the proof to convince whoever needs convincing.
What to have before you start
An AI project doesn't start with the model: it starts with the information and the process. Before writing a line, it helps to have four things:
- The information the AI will use, organized and up to date. If it's scattered or stale, get it in order first: AI doesn't fix bad data.
- Someone responsible for reviewing and deciding when the AI doesn't know or gets it wrong, especially at first.
- The task clearly defined: what goes in and what has to come out, with a couple of examples of correct results.
- A way to measure, even a simple one: hours saved, errors, response time.
When those four things are in place, AI is an implementation detail. When one is missing, the project stalls right there, not in the technology.
How to measure whether it works
"It feels better" isn't enough to decide. Some concrete ways to measure:
- Time saved: how many hours a week stop going into that task.
- Errors: whether they went down or up compared with doing it by hand.
- Response speed: how long it takes to resolve an inquiry or an order.
- Quality: a sample reviewed by a person, with a clear sense of what "good" means.
- Real usage: how many people use it and how often, not just whether it's available.
With those numbers the pilot stops being an opinion. And if it improves nothing, better to know early and stop investing.
Privacy and data
When you use AI with your company's or your customers' information, you have to decide carefully what data leaves and where it goes. A few rules I apply:
- Send the minimum. If the task doesn't need the name or the ID number, it doesn't go.
- Encrypt what's sensitive and define what's stored and for how long.
- Choose where it's processed. A model you host isn't the same as an external vendor; each has its conditions.
- Keep a log of what was processed, so you can answer if someone asks.
- With health, minors' or financial data, check the legal framework with a specialist.
Privacy isn't an afterthought: if you neglect it, the project fails a legal review or puts you at risk. It's better to solve it from the start.
How I approach it
I start by understanding what task you want to solve and what information you have, because everything depends on that. Then I build a pilot on part of the operation, with human review at first and an agreed way to measure. If it works, I scale it; if not, we drop it without having overinvested. AI fits into what I do in AI integration, always as one piece of a process, not an end in itself.