What should you not automate in your business? Anything that calls for judgment, empathy or negotiation; choices with serious consequences; any process that is still changing or rarely happens; sensitive data without someone checking it; and the core of your work. These are best kept by a person. Repetitive, high-volume jobs with clear rules, on the other hand, are good candidates for automation. For the opposite question, see our guide to which tasks to automate first.
For many small businesses, automation can save time and improve efficiency, but the technology only applies the rules it was given. Before you start, it pays to work out which business processes gain from it and which lose something important.
This guide is written for freelancers and SMEs in Spain, so its legal points refer to the GDPR and Spanish law. It also draws on public guidance from the UK, published by the National Cyber Security Centre (NCSC), the Information Commissioner's Office (ICO) and the government, and on reports from the CIPD, the professional body for HR in the UK. For an organisation based in Spain, those texts set out principles and good practice rather than requirements.
What tasks should not be automated?
A task should stay with a person when a machine would remove what gives it its value: a genuine conversation, a decision you can't undo or the skill you are paid for. In practice, that means:
- Sensitive customer moments: the first proper reply to an enquiry, a complaint, a tense conversation about money.
- Judgment and negotiation: anything where the right response varies with the context.
- High-impact choices: pricing, hiring, legal and tax matters.
- Unstable processes: ones you can't yet explain step by step, or that keep changing.
- Rare jobs: things you do once or twice a year, where building the automation would cost more than it saves.
- Sensitive data without oversight: medical and bank details, customer records.
- The core of your work: the skill customers come to you for.
Customer relationships that need a personal touch
An automatic message reads the same for everyone, which is exactly what an upset or worried customer doesn't want. In these moments, the customer decides whether to keep trusting your business. Customer communication and support are tempting candidates for automation, but turning these conversations into templates can cost you the customer.
The first reply to a new enquiry
An automatic acknowledgement that a message has arrived is helpful and bothers nobody. The reply to what they actually asked works better coming from you: two lines of your own take about the same time to write as setting up a template, and the difference shows.
Complaints and apologies
A customer who writes in anger and gets a prefabricated reply feels that nobody has looked at their message. In complaint management, use an automatic notice to confirm that the complaint has arrived and give the time frame for your response; the apology, the explanation and any exception to your returns policy need your own voice. If the complaint arrives as a review, see how to respond to negative reviews on Google.
Difficult conversations about money
Reminders for overdue invoices fit well into a workflow, and chasing unpaid invoices is often one of those jobs you never get round to. Negotiating a payment plan with a regular customer going through a bad patch is another matter: let the system send the reminders and keep any related decision for yourself.
Judgment calls that should stay yours
An automation follows a rule: if this happens, do that. It works when the rule is unambiguous, and becomes a risk when the right decision depends on the context. Offering a goodwill gesture, spotting a job that looks like a scam or choosing which project to accept: each of these requires experience and critical thinking. An app can gather the information and flag what looks unusual, but the final word is yours. Bigger organisations have committees, teams and approval processes for these choices; in a one-person business, you make the call.
Pricing shows this well. A program can compare your rates with the market, but raising your prices or making an exception for a key customer depends on your strategy, your margins and the relationship, and none of that fits in a formula.
Hiring, sick leave and dismissals
Software can sort CVs or keep the daily record of working hours (the registro de jornada) for your employees. Anything related to hiring, an employee's sick leave or a dismissal affects people: an automatic filter can discard exactly the profile you were looking for, and a dismissal should never arrive by automatic email.
Legislation points the same way. Where a firm has workers' representatives, Spain's Workers' Statute (Estatuto de los Trabajadores) gives them the right to be informed of the rules behind the algorithms that affect decision-making on conditions of employment, hiring or keeping a job, including profiling. In the UK, the CIPD's employment law notes stress that no single piece of legislation governs their use in employment: employers have to deal with several areas of law at once, from discrimination to the protection of personal details.
Legal and tax matters
Accounting software takes a good part of the paperwork of a sole trader off your hands. But deciding whether an expense is deductible, how to respond to a formal request from the Tax Agency (Hacienda) or which clause to accept in a contract calls for professional knowledge and judgment: yours or your adviser's. Software can help you meet deadlines, but it can't decide which obligations apply to you.
Processes that are poorly defined or keep changing
Automation copies a process; it doesn't fix it. If the process is unclear or still moving, a machine repeats the confusion, only faster. If nobody can explain who does what, in which order and with which information, there is no solid base to hand over yet. Bad assumptions built into a workflow are hard to spot, because they are applied without complaint, and incorrect data then spreads across your connected systems. Map the process first: define it, tidy it up, write down the steps and run it by hand until it runs smoothly. Our guide on how automating a process works, step by step shows what to write down before you start.
The same applies to anything still evolving. If you are adjusting your rates, testing new services or reorganising your team every couple of months, building it now means freezing a version you are about to abandon. Each change forces you to undo and rebuild, and that back-and-forth soon outweighs the time saved. Let it settle first: something that keeps changing is not a bad candidate forever, just a bad candidate for now.
When automation isn't worth the cost
Building and maintaining an automation takes about the same effort whether you use it a lot or a little, but the time it actually saves depends on how often the task happens. An annoying task is not necessarily a frequent one, and low volume rules out many automation projects.
A task you do every day or every week deserves a close look; one you do three times a year almost never does. For those rare jobs, use a simple checklist instead: it is free and never breaks down on a Saturday. To put an annual figure on a specific task, use the formula for calculating the cost of an automation.
An automation isn't finished once it is built either. An app changes its interface, a form gains a new field, a supplier changes the layout of its invoices, and someone has to fix and maintain the workflow. Otherwise, one day the automation stops without warning. The cost of maintenance is often underestimated: include it before deciding whether a task is worth the effort.
Personal and sensitive data
If your business handles files on customers, patients or staff, or other people's personal details, automation is still possible, but with care: using an app doesn't change who is responsible for them. Under the GDPR and the LOPDGDD, the Spanish law that complements it, you remain the data controller, responsible for respecting the rights of the people concerned, and every service you share them with is one more point where they can leak. Take particular care with:
- health data and the other special categories of personal data, which get stronger protection under the GDPR;
- bank details and payments;
- automatic decisions that have a significant effect on a person, such as rejecting a job candidate;
- external tools that keep copies of your documents;
- any new processing that could seriously affect individuals' rights, where the GDPR may oblige you to carry out an impact assessment beforehand.
What the GDPR says when a machine decides alone
The GDPR gives everyone the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects for them or similarly significantly affects them. There are exceptions, such as a decision that is necessary to enter into or perform a contract, or one based on the person's explicit consent. Even then, the controller must protect the person's rights and freedoms with suitable measures, including at least the right to obtain human intervention, to express their point of view and to contest the decision.
The UK has taken a different route. The ICO explains that recent UK legislation opens up a wider range of lawful bases for decisions with serious effects made solely by a machine, provided organisations keep appropriate safeguards, while special categories stay more protected. That reform changes UK law; an organisation established in Spain remains bound by the GDPR and the LOPDGDD, and the Spanish Data Protection Agency (AEPD) publishes practical guidance on applying them.
A review that really counts
A final sign-off is not enough if the person in that role simply approves whatever the machine proposes. For the review to mean something, ensure the reviewer has the relevant information and evidence to understand the case, the time to look at it and the authority to change the outcome. If they can't say no, they are only rubber-stamping. And keep an open channel for the person affected to contest the outcome.
None of this should stop you: check how each service protects privacy and security, limit access to what is strictly necessary and let a person review the exceptions.
The core of your work and your brand
There is a difference between what surrounds your work and the craft itself. A photographer in Valencia can let an app handle quotes, reminders and file management, but not the way they compose an image; a physiotherapist in Alicante can let patients book online, but not hand over the session. A machine can help you prepare and organise around your craft, but what makes customers choose your business is the last thing to hand over.
Your brand works the same way. Scheduling posts is admin; the writing, the content and the personal contact with your customers are your signature. In sales, a machine can handle the follow-up, not the relationship.
Regulated professions: your code of conduct still applies
If you are one of the regulated professionals in Spain, such as lawyers, doctors, psychologists or physiotherapists, your colegio profesional and its code of conduct keep applying when a machine is involved. Confidentiality of client or patient details, your personal responsibility for advice or a diagnosis and, in many codes, the duty to stay up to date in your field remain yours, and so do the professional standards your colegio sets. Before handing over any part of a professional service, check whether your colegio has published guidance or standards relevant to chatbots and language models: in an expert service, the judgment you bring to your area of practice is what clients pay for and trust.
The risks of over-automation
A person makes a mistake once; a machine repeats it in every operation, at full speed, until someone notices. And the more systems and workflows you chain together, the harder it is to see where an error comes from.
- Errors at scale: one badly configured rule sends the wrong invoice to your entire client list in seconds.
- Loss of control: if nobody understands what the setup does any more, your business depends on a black box.
- Dependence on one platform: if the provider changes its pricing model or closes down, your operations stop with it.
- No way out: a chatbot that never hands over to a person frustrates customers just when they need help.
- Automating for the sake of it: doing it because other businesses do, without a clear problem to solve, is a waste of money. Sometimes the simplest fix is to drop the task altogether.
Treat every automation you create as a project: give it someone in your team who checks it regularly, a log of what it does and an easy way to switch it off. One that nobody watches can pile up errors in silence for weeks. Other, more technical pitfalls are covered in the costliest small business automation mistakes.
How to decide, task by task
A good way to start is with two key questions: what does the task lose if a machine does it, and what do you gain in time, efficiency or fewer errors? If it loses little and the saving is obvious, go ahead without a second thought; if it loses something your customers care about, keep doing it yourself.
The one-question test
Imagine a customer found out that a machine handles the task: would they be reassured or disappointed? With the filing of your invoices, they wouldn't care. With the reply to their delicate enquiry, they would be disappointed, and rightly so.
A quick risk checklist: questions to ask first
- Is the right answer always the same, or does it depend on the situation?
- If the system gets it wrong, what is the impact, and can it be undone?
- Does the task happen often enough to justify the effort to build and maintain it?
- Is the process stable and repeatable, with clear rules and accurate inputs?
- Does it involve personal details or payments?
- Do your customers expect a person to do it?
If one of these questions is hard to answer, it isn't all or nothing. There is a middle way: the machine prepares the job, and a person steps in at the point where the decision is made. The following table shows where the line falls for some everyday jobs:
| Hand over with confidence | Hand over with a review | Keep in human hands |
|---|---|---|
| The acknowledgement of an enquiry | Sorting enquiries by topic | The reply to a delicate enquiry |
| Reminders for overdue invoices | The final notice before you claim an unpaid invoice | Negotiating with a customer in difficulty |
| Moving files between your tools | Assigning each request to the right person | A price rise or an exception |
| Filing invoices and documents | Category suggestions for each expense | Responding to a formal request from Hacienda |
| A first draft written with AI | A suggested reply to a routine question | The final version that carries your name |
For the middle column to function, decide in advance which details each request must include for the machine to handle it alone; anything that arrives incomplete goes to the person who decides.
AI suggests, you decide
Artificial intelligence widens what automation can do without changing the main rule: the tool suggests and a person decides. A language model can draft an email or summarise a long document in seconds, and it can also state an invented fact with complete confidence, which is why it shouldn't run your customer support unsupervised. Use it mainly in three ways:
- preparing drafts of emails or posts that you then review;
- summarising and organising long notes or documents;
- flagging what looks out of the ordinary.
Don't publish what it produces without reading it, and don't let it reply on its own to a customer with a serious complaint: reviewing its work is how you get the benefits of the technology without risking your reputation. To get started, see this selection of AI tools for freelancers.
How can you test AI before relying on it?
Try it on a sample of past enquiries or documents from your own firm, compare its results with what you would have done and keep evidence of what you checked. The UK government calls this kind of exercise AI assurance: measuring, evaluating and communicating how far it deserves trust.
The government's introduction to the subject, issued by the Department for Science, Innovation and Technology as the first in a series on assurance and governance, describes several techniques that organisations can combine; they include an assessment of the risks, an impact assessment covering wider effects, bias audits, compliance audits and conformity assessment. A sole trader or an SME doesn't need a formal audit, but it can borrow the same logic:
- Research and risks: check where the provider keeps your files and whether they are used to train its models, then ask what the worst mistake could be and who it would affect.
- Performance testing: run it on a range of past enquiries, invoices or CVs, focus on the ones where a mistake would hurt most and compare its output with your own choices.
- Bias check: look for results that consistently put one group at a disadvantage, one of the known issues with CV filters.
- Compliance check: ensure it meets your legal requirements and any professional standards that apply, above all on data protection.
- Evidence: keep a short report of what you tested and what you observed, so you can explain later why you trusted it.
The same guidance points out that no single technique settles the question: checks are combined and repeated, in proportion to the risks, and shared technical standards make them consistent. Base your decision on these checks rather than on the sales pitch, and test again when the provider releases the latest version of its model, when you change the way you use it and as part of an annual review.
Should an AI agent act on its own?
Only for low-stakes, tightly bounded jobs that are easy to undo. Anything with a lasting impact, such as payments, messages to your whole customer list or changes to customer files, should wait for a person's approval.
An AI agent goes further than a chatbot that answers questions: it can consult your files, use other applications, pursue a goal and take action with little or no supervision, such as replying to messages, updating a spreadsheet or placing an order online. In guidance published in 2026, the UK's NCSC urges organisations to do some careful thinking before the adoption of agents, and explains that this autonomy is the source of both their benefits and their extra risks. Agents receive wider access to your accounts and files, they can interpret a goal in a literal or unexpected way, they can act more quickly than anyone can review, and it can be hard to explain afterwards why they did something. They also inherit the weaknesses of language models, such as prompt injection, where instructions hidden in a message or a web page push the agent to do something you never asked for.
Three levels of oversight
The NCSC distinguishes three ways of keeping a person involved, depending on the stakes and on how much trust the agent has earned:
- In the loop: a person approves each action before it happens. For a sole trader or an SME, that is the right level for payments, refunds, contracts and anything with an impact on customers.
- On the loop: the agent acts, and a person monitors it and can step in. It suits internal work that is easy to reverse, such as sorting incoming messages.
- Out of the loop: the agent acts without any review. Keep this for jobs where a mistake would be trivial.
Limits to set before an agent goes live
- Give some thought to whether you need an agent at all: an honest assessment of the benefits may show that simpler options, such as a written procedure or an ordinary set of fixed rules, do the job more safely.
- Begin with a pilot: one specific, well-defined task with a single purpose and a narrow focus, and widen its scope only once it has proved reliable.
- Grant the minimum permissions: ensure it can reach only the folders, accounts and services needed for the task, for as short a period as possible, and never open-ended rights over your bank account, your customer database or anything else critical.
- Keep logs: a record of what the agent did, so you can trace an error or a security incident back to its source.
- Give someone the lead role: a responsible person who approves its permissions, checks its output and has the authority and capability to switch it off at once.
The NCSC's rule of thumb applies to a one-person firm as much as to a major organisation with formal governance: if you can't understand, monitor or contain what an agent does, it isn't ready to be used.
Agree the ground rules with your staff
As soon as you have staff, deciding what a machine may handle also becomes a question of roles, responsibilities and trust. The CIPD advises organisations to agree a short list of shared principles before they scale up their use of AI: which tools are approved and for what purpose, how personal details are handled, where AI must not be used and what happens in case of misuse. In a firm with two or three staff, you will probably lead the work yourself, and this code of conduct fits on one page. It should include the following points:
- the applications everyone may use, and the accounts they run under;
- the information that must never be pasted into a chatbot, such as medical files or customers' bank details;
- the choices that stay in a person's hands: hiring, performance reviews, promotions and dismissals;
- the role of each person: who checks the results, and who to tell when issues arise;
- the conduct expected when a chatbot produces something doubtful: stop, check and report it.
It doesn't call for an expert: it is basic governance, agreed as a group, written in plain words, put into practice from day one, kept up to date and open to revision when the rules, the benefits or the available options change. Keep your staff involved from the outset and develop the rules with them, because they have knowledge of exceptions that never appear in a written procedure. In the case studies behind a CIPD research report on responsible adoption, top-down mandates often bred mistrust, while changes shaped together with staff were more successful.
Watch the hidden burden of checking, too. Reviewing what an assistant produces demands attention, and the CIPD warns about the fatigue that sets in when staff spend their days verifying machine output. Give staff the chance to develop the knowledge and skills to spot a mistaken answer, protect room for thinking, and focus the hours freed up on the work that calls for judgment rather than squeezing more into the same day.
Official guidance and resources
You don't have to settle these questions alone, and help is available: several public bodies publish free guidance online, and some of it is aimed at SMEs and the self-employed.
- In Spain: the AEPD publishes practical guides on applying the GDPR, among them a guide to impact assessment, and INCIBE, Spain's public cybersecurity institute, offers a range of free resources and support for SMEs and the self-employed, with alerts on the latest threats and a helpline for incidents.
- At EU level: the AI Act, which also applies in Spain, lists AI used to recruit candidates or to decide on workers' promotion or dismissal among its high-risk uses, with extra requirements for the organisations that use it, such as ensuring oversight by persons with the competence, training, authority and support to intervene.
- In the UK: the NCSC publishes cyber security advice for organisations, the ICO explains UK privacy law, the government's assurance guidance is available on GOV.UK and the CIPD publishes resources to support HR professionals in their roles, including notes on employment law.
Treat the UK texts as a source of principles and good practice; for your own obligations, your options are the relevant authority, such as the AEPD, your gestor or another professional adviser.
What you can safely hand over to a machine
Once you set all that aside, plenty is left: repetitive, predictable jobs with no delicate choices, necessary chores that nobody misses doing. It covers back-office functions, the operational side of your business, and a good place to start:
- filing invoices and updating a spreadsheet you share with your team;
- reminders for appointments, payments and deadlines;
- moving customer data between your systems, for example from a contact form into your CRM;
- backups of your files;
- internal reports and the admin you repeat every week.
There, machines help without getting tired: they remove bottlenecks, reduce manual work, improve productivity and save you time for higher-value jobs. To choose between a person and a machine, see should you hire or automate?
Knowing where to stop is worth as much as knowing where to begin. Our free automation diagnostic puts a figure on the annual cost of a task in about two minutes: four questions, the result on screen and no commitment. If you answer six more questions, it tells you whether the task is worth automating, and has no trouble concluding that it isn't.