Start with a process that happens every day, follows clear rules, has high volume and can be measured: sorting incoming email, moving invoice data into your software, or sending an alert before something expires. Measure what it costs you today, run it in draft mode with a person reviewing, and compare after a month. If rules are enough, you do not need AI.
Automating with rules and automating with AI are not the same
There are two ways to automate, and it pays to keep them apart from the start.
Rule-based automation follows fixed instructions: "if an email arrives from this supplier, save the attachment in this folder and tell accounts". It always does the same thing. If it fails, it always fails the same way, and you fix it by changing the rule.
AI, specifically the language models offered by Anthropic, OpenAI, Google or Microsoft, is for what rules cannot do: read free text, understand what an email is about, pull data out of documents that do not always share the same format, or write a draft. The price is that it makes mistakes. Anthropic admits in its documentation that even the most advanced models can generate text that is incorrect or inconsistent with the context, and that its techniques to prevent this reduce the problem but do not eliminate it (Anthropic, "Reduce hallucinations"). Microsoft warns that responses from its Copilot are not 100% guaranteed and should be checked before you send them to anyone (Microsoft Learn).
| With rules | With AI | |
|---|---|---|
| What it is for | Data in a fixed format: forms, spreadsheets, dates | Free text, varied documents, drafts |
| How it fails | Always the same way, easy to spot | Unpredictably, and sometimes with apparent confidence |
| Running cost | Low | You pay for model usage |
| Human review | By sampling | Of everything that goes outside |
The safest combination is this: AI reads and suggests, rules decide what happens with that suggestion, and a person approves anything that goes to a customer.
How to choose your first process: four filters
Make a list of the tasks that repeat in your business. Then run each one through these four filters and score them from 0 to 2.
- Repetitive. It happens many times and almost always the same way. Answering "can you send me March's invoice?" is. Preparing a proposal for a new client is not.
- Rule-based. You could explain how it is done on one page to someone who has just joined. If the answer is "it depends" and you do not know on what, it is not ready yet.
- High volume. The more times it happens a week, the sooner you recover what it costs to set up. A task that happens twice a month will take a long time to pay off.
- Measurable. You know how often it happens and how long it takes. If you do not, count it for a week before deciding.
Add two more questions. What happens if the system gets it wrong? Where is the data? To start, choose a task where a mistake is spotted and corrected without harm, and whose data already sits in something that can be connected: email, a spreadsheet or your management software.
| Task | Repetitive (0-2) | Rule-based (0-2) | Volume (0-2) | Measurable (0-2) | Total |
|---|---|---|---|---|---|
| Your task 1 | |||||
| Your task 2 | |||||
| Your task 3 |
Start with the highest scorer among those that do the least harm if they fail.
How much time you save: do the sums with your own numbers
An honest calculation includes two things: the time you will spend reviewing what the system does, and the cases you will end up doing by hand anyway. You need four figures.
| Figure | What it is | How you get it | Your figure |
|---|---|---|---|
| A | Cases per week | Count them for a week | ____ |
| B | Minutes per case today | Time ten cases and take the average | ____ |
| C | Share of cases the system gets right, from 0 to 1 | Measure it in the trial, do not guess it | ____ |
| D | Review minutes per case with the system | Measure them in the trial | ____ |
With them you do three calculations:
- Time today = A × B.
- Time with the system = (A × D) + (A × (1 − C) × B). The first part is reviewing every case. The second is redoing by hand the ones that go wrong.
- Weekly saving in hours = (time today − time with the system) ÷ 60.
To see how it works, with numbers chosen only for the example: if A = 100 cases, B = 5 minutes, C = 0.8 and D = 1 minute, today you spend 500 minutes. With the system it would be 100 minutes of review plus 100 minutes of cases by hand (100 × 0.2 × 5), 200 in total. The saving is 300 minutes: 5 hours a week. Now change two figures. If C were 0.5 and D were 2 minutes, the time with the system rises to 200 + 250 = 450 minutes and the saving falls to 50 minutes. That is why C and D are measured in a real trial, not estimated.
To turn it into money:
- Cost per hour of the person doing the task: what they cost the business per year, divided by their working hours per year.
- Monthly saving = weekly saving in hours × working weeks per month × cost per hour.
- Months to recover the investment = project cost ÷ (monthly saving − monthly running cost).
The monthly running cost includes AI model usage, hosting and maintenance. And one warning: hours saved only turn into money if you spend them on other useful work.
Examples by sector
| Sector | Process | What AI does | What a person checks |
|---|---|---|---|
| Property management | Turning residents' emails into incidents for each building | Reads the email and suggests building, type and urgency | The building and the urgency, before calling the contractor |
| Estate agency | Answering enquiries from property portals and proposing a viewing | Reads the enquiry and drafts the reply | Price, availability and any detail not in the listing |
| Advisory firm | Moving received invoices into the accounts | Extracts the data from each invoice | The ones that do not add up or where the system is unsure |
| Shops or services | Alerting before a contract, insurance policy or certificate expires | Nothing: dates and a rule are enough | That the dates entered are correct |
| Any office | Spotting customer emails with no reply | Nothing: rules are enough | Who answers each one |
| Any office | Finding a scanned document by its content | Reads the text of the scans (OCR) | Nothing, it is a search |
For the advisory firm (in Spain, an asesoría or gestoría, the firms that handle a business's accounts, tax and payroll), the law gives you the list of fields to extract. Every invoice must show its number and series, date, the issuer's name and NIF (Spanish tax ID), address, a description of the transaction, the tax rate and the tax amount, among other details (Royal Decree 1619/2012, art. 6). If the system cannot find one of them, the invoice goes to a person.
We go deeper into the first two examples in AI for property managers in Spain and AI for estate agents in Spain. You will find more cases by sector under property managers, real estate and advisory firms.
We have built the last three as products: Renewals sends email alerts every morning before something expires, Unanswered-mail alerts sends a daily summary of customer emails with no reply in each Microsoft 365 mailbox, and Archive finds documents by their content, scans included, with OCR on your own computer and no cloud.
The controls you cannot skip
- A person reviews everything that goes outside. Emails to customers, documents, posts. The AEPD (Spain's data protection authority), in the internal rules it has set itself for using generative AI, requires results to be reviewed and validated by hand before they are used or published (AEPD, summary of obligations, January 2026). It is the agency's internal policy, but it is a useful benchmark.
- Start in draft mode. The system suggests and the person sends. Only once you have measured how often it gets things wrong should you think about fully automating the lowest-risk part.
- Minimum access, a log and a way to stop it. INCIBE (Spain's national cybersecurity institute) recommends that an AI connected to your systems only reaches what is essential, keeps a record of what it does, is tested before use and can be stopped quickly. Also that it never pays, deletes, grants permissions or sends sensitive data on its own (INCIBE, September 2026).
- Watch out for hidden instructions. The AEPD explains that a PDF, an email or a web page can carry instructions that are invisible to a person but that the model reads as orders. Its example is an email with white text on a white background that an agent obeys while summarising it (AEPD, agentic AI, February 2026). That is why it proposes its "rule of 2": an agent that reads content from anyone, has access to sensitive information and acts on its own is a setup that should not be allowed. If it reads outside emails, it should not act without human supervision.
- A business account and a contract. The AEPD's Lab reminds you that a free account is not the same as a business account with contractual guarantees (AEPD, August 2026). On their business plans, the big providers say they do not train their models on your data by default: OpenAI on its API, Anthropic on its commercial products, Microsoft in Copilot and Google in Workspace. Check where the data is processed too. Microsoft keeps its EU users' traffic inside its EU Data Boundary, although today it excludes from it the Anthropic models it offers inside Copilot. OpenAI has a European region on its API, with prerequisites. The provider processes data on your behalf, so you need a data processing agreement (GDPR, art. 28).
- Personal data: only what is needed. The GDPR requires data to be "adequate, relevant and limited to what is necessary" (art. 5.1.c). If one paragraph is enough, do not upload the whole file. And note: swapping a name for an "X" does not always anonymise, warns the AEPD's Lab. Avoid decisions based solely on automated processing that produce legal effects on a person or similarly affect them: GDPR art. 22 gives people the right not to be subject to them, with some exceptions. And if the processing is likely to result in a high risk to people, you need an impact assessment before you start (art. 35).
- Measure the errors. During the trial, note every case where the system got it wrong and what the right answer was. That is your figure C, and it also tells you which rules are missing.
On human review, a 2026 warning from the AEPD to a company preparing an AI to score job applications sums it up well. Human involvement has to be effective: whoever decides must be able to weigh the system's score with their own judgement, not simply confirm it (AEPD, warning in case EXP202600427).
What the EU AI Act asks of you
If your business uses an AI tool in its work, the regulation treats you as a "deployer": anyone using an AI system under their own authority, unless the use is personal and non-professional (Regulation (EU) 2024/1689, art. 3.4). The regulation was amended by the so-called digital omnibus, Regulation (EU) 2026/1744, published on 24 July 2026 and in force since 27 July, according to the European Commission's timeline.
| Obligation | From | What it means for a small business |
|---|---|---|
| AI literacy (art. 4) | 2 February 2025 | Take measures so your staff know how to use the AI they use. After the reform, these are measures to support that training, with no set level required of each person |
| Prohibited practices (art. 5) | 2 February 2025, plus two new ones from 2 December 2026 | They do not affect normal office uses. The new ones ban systems that generate realistic intimate or sexual content of people without their consent, and child sexual abuse material |
| Transparency (art. 50) | 2 August 2026 | Anyone talking to an AI assistant must know it, unless it is obvious. If you publish AI-generated or manipulated images, audio or video that could pass as genuine, you must say so |
| High risk, Annex III | 2 December 2027 | If you use AI to recruit staff, for decisions affecting working conditions or to assess people's creditworthiness, you will have extra obligations (Annex III) |
Three practical details. First, the duty to tell people they are talking to an AI falls on the provider (art. 50.1), but the regulation treats as a provider anyone who has a system developed and puts it into service under their own name or trademark (art. 3.3). If someone builds a custom assistant for you and you use it under your brand, expect the duty to be yours. Second, if you publish AI-generated text to inform the public on matters of public interest, you must say so, unless there is human review and someone takes editorial responsibility (art. 50.4). Third, for high-risk uses, Article 26 will require you to use the system according to its instructions, assign human oversight to people with the necessary competence, training and authority, and inform workers' representatives and the affected workers before using it in the workplace.
AESIA, Spain's AI supervisory agency, publishes practical guides to the regulation and warns on its website that they will be updated after the digital omnibus. Always check dates against the text on EUR-Lex.
A 30-day plan to get started
| Days | What you do | What you have at the end |
|---|---|---|
| 1 to 5 | Choose the process with the four filters. Count A and time B | A chosen task and what it costs today |
| 6 to 10 | Write down the rules, the exceptions and which cases always go to a person. Choose a provider with a business account and a data processing agreement | One page with the process and its limits |
| 11 to 20 | Set it up in draft mode and run it alongside normal work | A log of every suggestion: right or wrong |
| 21 to 30 | Calculate C and D, redo the sums and decide | Carry on, adjust or stop, based on data |
In parallel, write a one-page AI use policy. The AEPD's Lab suggests starting with a simple inventory: which tools are used, by whom, for what and with what data. Then decide which tools are allowed and which data never goes in (AEPD, August 2026). INCIBE adds that banning all AI tends to push people to use it in secret, and that it is more practical to approve specific tools and require human review of their output (INCIBE).
Mistakes worth avoiding
- Starting with the hardest process. The first one is for learning. Pick a small one.
- Counting only the time gained. Review and the cases done by hand count too.
- Letting the AI send on its own from day one. Draft first, then data, then decision.
- Using free personal accounts with customer data. Without a contract you do not know what happens to that data.
- Putting nobody in charge. Someone has to look at the error log every week.
If you want us to build it
At Quantum Nova, based in Alicante and working with businesses across Spain, we design and build automations and AI agents connected to the tools you already use. First we run a free 2 to 5 day assessment and give you a fixed price. Our published ranges, ex VAT: one automation workflow, €1,500 to €6,000; several workflows, €5,000 to €20,000; a first AI product (MVP), €4,000 to €12,000. A typical automation is up and running in 2 to 4 weeks, with the data hosted in the EU and no subcontracting. You will find the details under process automation and AI solutions.
Sources
- Regulation (EU) 2024/1689, the Artificial Intelligence Act, arts. 3, 4, 5, 26 and 50.
- Regulation (EU) 2026/1744 (digital omnibus on AI), which amends arts. 4, 5 and 113 of the above.
- European Commission, AI Act application timeline.
- European Commission, AI Act Service Desk, Annex III.
- AESIA, guides to the AI Act (in Spanish).
- Regulation (EU) 2016/679 (GDPR), arts. 5, 22, 28 and 35.
- AEPD, Agentic artificial intelligence from a data protection perspective, v1.2, February 2026 (in Spanish).
- AEPD, Basic summary of obligations and recommendations for managing generative AI at the AEPD, January 2026 (in Spanish).
- AEPD Lab, "Prompts, personal data and confidentiality", 25 August 2026 (in Spanish).
- AEPD, warning in case EXP202600427 (in Spanish).
- INCIBE, "The human factor in cybersecurity for SMEs and the self-employed", 3 September 2026 (in Spanish).
- Royal Decree 1619/2012, Invoicing Regulation (in Spanish), art. 6.
- Anthropic, "Reduce hallucinations".
- Anthropic Privacy Center, is my data used for model training?.
- OpenAI, "Data controls in the OpenAI platform".
- Microsoft Learn, "Data, Privacy, and Security for Microsoft Copilot".
- Google Workspace, Generative AI Privacy Hub.