What NOT to automate with AI (and what to automate): how an AI process automation company decides in Palma de Mallorca and the Balearic Islands
Published on
%3A%20c%C3%B3mo%20decide%20una%20empresa%20de%20automatizaci%C3%B3n%20de%20procesos%20con%20IA%20en%20Palma%20de%20Mallorca%20y%20Baleares.jpg)
Six weeks of development, around €8,000 and a three-month lifespan. That is the typical path of an AI assistant that replies on its own to guest complaints. We have seen it at holiday-rental management companies in Palma de Mallorca: the tool worked and wrote well in three languages… and they still ended up switching it off. A guest with a broken shower in the middle of August did not want a flawless reply; they wanted someone with a name and surname to take charge. Meanwhile, copying bookings from platform emails into the cleaning schedule —around twelve hours a week— was still being done by hand.
The figures are illustrative, but the script repeats itself across the Balearic Islands: it is not the artificial intelligence that fails, it is the choice of process. That is why, before a single line of code is written, a serious AI process automation company should help you decide three things: what to automate fully, what to keep in assisted mode and what not to touch. This article is only about that decision, with examples from Palma, Ibiza and Menorca.
Quick answers before you start
Can almost everything be automated with AI?
Technically, far more than is wise. The useful question is not “can it be done?” but “what happens when it gets it wrong, and who finds out?”. If the answer is “an angry customer” or “the tax office”, you probably need to assist, not automate.
What is the difference between automating and assisting?
Automating means the task runs without anyone reviewing it case by case. Assisting means the AI prepares the work (a draft, a classification, a summary) and a person approves it with one click. For many small and medium-sized businesses in Mallorca, assisting is the most profitable starting point.
How long does this classification take?
With a well-described list of tasks, a first classification takes days, not weeks. What takes longest is measuring what each task costs today.
What if my business is seasonal?
Then the decision changes with the month. A task that is not worth automating in March may be critical in July. In the Balearics it pays to assess every task against high-season volume, not against a Tuesday in November.
The three-question rule
Every task people bring us goes through the same filter. It is not an algorithm or a scoring sheet: it is three questions, in this order, that any manager can ask out loud.
First: does the task follow rules you could explain to a new hire in one afternoon? If doing it well requires “knowing this customer” or “having a feel for it”, AI can help, but it cannot decide. If it boils down to “when this arrives, do that”, it is a candidate for automation.
Second: what does a mistake cost, and who notices it first? A misfiled invoice that your accountant spots at the end of the month is a cheap mistake. A wrong message to a VIP guest in Ibiza, or a medical appointment confirmed in the wrong slot, is noticed by the customer before you. The more expensive and visible the mistake, the closer a person needs to be.
Third: does the task build or protect a relationship, or carry legal responsibility? Negotiating with a property owner, delivering bad news, accepting a quote or deciding on health data are not processes: they are decisions with a name attached. There, AI can prepare the ground, but never replace the person who is accountable.
If a task passes all three questions —clear rules, cheap mistakes and no relationship or responsibility at stake— it gets automated. If it only fails the second or the third, it gets assisted. If it fails the first and at least one other, it stays human, at least for now.
Automate, assist or keep human: what it means in practice
Automate
The system runs the task end to end and logs everything. A person reviews samples and exceptions, not every case. Examples: moving bookings from the platforms into the cleaning calendar, sending appointment reminders or reconciling payments with invoices.
Assist with AI
The AI proposes and a person decides: draft replies, incident classification, call summaries or extracting data from a PDF for validation. When a case moves from the machine to a person, a good handoff card saves them from starting from scratch.
Keep human
A person decides and carries it out, even if they use digital tools. At most, the AI prepares the context for them. Examples: serious complaints, negotiations, clinical decisions or special prices for key customers.
A task can move between groups over time. The sensible approach is to start in “assist”, measure for a few weeks and move to “automate” only what has proven reliable. The reverse costs money and, above all, trust.
The decision path, step by step
- Write the list without filtering. Between ten and twenty repetitive tasks, written as verbs: “copy”, “notify”, “check”, “reply”. If they are hard to spot, a bottleneck map helps you see them.
- Put a number on each one. Hours per week in high season and in low season. Approximate is fine; invented is not.
- Run them through the three questions. Rules, cost of a mistake, and relationship or responsibility. Write down the why.
- Pick just one task to automate and one to assist. Not ten. The first proves the savings; the second proves the team can trust the AI.
- Decide how to undo it. Before switching anything on, define what happens if it fails on a Saturday in August: who gets the alert and how you go back to the manual process within minutes.
- Review after 30 days. What worked moves up a group; what did not, moves down. It is the logic of our 30-60-90 plan, applied task by task.
One island, three decisions
These are typical examples of what we usually recommend on each island: not specific clients, but combinations that are common across the Balearic Islands.
Palma de Mallorca: private clinic with several consulting rooms
Automate: appointment reminders by WhatsApp or SMS, with confirmation and rescheduling through a link, connected to the real diary.
Assist: the email and web-form inbox. The AI classifies each message (new appointment, change, invoice, clinical query), extracts the data and proposes a reply; reception approves or corrects it.
Keep human: any reply with clinical content and the handling of complaints. Here, professional responsibility and the protection of health data come first, not efficiency.
Ibiza: restaurant with a terrace and private events
Automate: pulling bookings from the website and the platforms into a single system, with an automatic alert to the kitchen when a customer mentions allergies or intolerances.
Assist: event requests (birthdays, company dinners, small weddings). The AI summarises the request, checks availability and drafts a proposal with the right menu; the events manager adjusts price and tone.
Keep human: negotiating with large groups and replying to negative reviews. In the middle of the Ibiza season, a badly handled group costs more than many hours of manual work.
Menorca: boat rental company
Automate: sending the pre-departure documents (contract, departure checklist, harbour instructions) in the customer’s language as soon as payment is confirmed.
Assist: daily departure planning based on bookings and the weather forecast. The AI proposes the boat and skipper rota; the base manager decides.
Keep human: the decision whether to cancel a trip because of sea conditions, and the conversation with the customer when it is cancelled. It is safety and it is a relationship: two strong reasons not to delegate it.
If there is a CRM involved in your business, integrating AI agents into a CRM is usually where these decisions become concrete: which fields the AI writes, which ones it only proposes and which ones it must not touch.
Five things we would not automate (even if someone sells them to you)
- The first reply to a serious complaint. It can go out fast with the help of AI, but it should be signed by a person.
- Off-tariff prices and discounts. The AI can suggest; whoever knows the customer decides.
- Any action that cannot be undone. Payments, cancellations or bulk deletions without human approval are a gamble, not a process.
- Processes nobody currently does the same way twice. Automating chaos makes it faster, not better. First you organise; then you automate.
- Tasks whose rules change every season. If what worked in 2026 will not work in 2027, you may need a flexible tool rather than rigid automation.
Why a good automation company will tell you no
Be wary of anyone who agrees to automate everything you ask for. An AI process automation company that takes your business seriously will hand you, along with the quote, the list of what it would not do and the reason behind each “no”.
At Derek Solutions this is how we work: first we classify, then we build. When a task requires connecting systems that do not talk to each other, we solve it with custom software development in Palma; when the process lives on the phone of a field team, which is very common in rentals and services in Ibiza and Formentera, we turn to app development in Ibiza. And if what you are designing is an app with agents that act on their own, the reasoning is similar to our human, agent or hybrid task map, although there the focus is the product and here your internal processes.
Your list of 10 tasks, returned in three groups
Try it with your own business. Write down ten repetitive tasks your team does —one line per task, with the approximate hours per week— and send them to us. Within two working days we return the list sorted into three groups: automate, assist with AI and keep human, with the reason for each decision and the task we would start with. No commitment, whether you work in Palma de Mallorca, Ibiza, Menorca or anywhere else in the Balearics.
Send it to derek@dereksolutions.com · +34 680 283 973 and tell us your sector and your island. If any of those tasks should not be automated, we will tell you that too.