Which processes to automate with AI first (and which not to)

Which processes to automate with AI first

The processes worth automating with AI first are the ones a person does many times, following rules they could explain to a new colleague, and whose errors are spotted and corrected without harm. They are not the most impressive or the largest: they are the ones that free up the most hours with the least risk.

Choosing the first one well is half the work of company hybridisation with AI, which divides each process between AI agents and people. If the first process goes well and can be measured, the team trusts the second. If it goes badly, the company concludes that AI does not work here, when what failed was the choice.

An AI agent is not a script that repeats clicks. It reads a request, looks up the data it needs, drafts a reply or a document and tells the person who has to approve it. So the question is not only which task repeats, but which part of the process an agent can do with your company's rules and which part a person has to review.

Five questions to choose the first process

Before you automate anything, ask these five questions about each candidate process. The more answers are yes, the better the candidate:

1. Does it repeat often? A process that happens daily pays back the effort sooner than one that happens twice a year.
2. Can its rules be written down? If the person who does it today can explain how they decide, an agent can work with those rules. If the answer is «it depends», you first need to write down what it depends on.
3. Is an error cheap to correct? Sorting a request into the wrong queue takes a minute to fix; sending a client the wrong terms does not.
4. Does it have an owner? Someone has to decide how the process should work, review what the agent does and answer for the result.
5. Can it be measured today? If you do not know how long it takes, how much manual work it needs and which errors it produces, you will not be able to tell afterwards whether automating it worked.

The fifth question is the one most often skipped. Measuring before you start, what we call the baseline, is what turns an impression into a result you can show.

A table to score each process

Put each candidate process in a row and give each criterion 0, 1 or 2 points. The one with the highest total, with the error criterion at 2, is usually the first.

A 1 means halfway, and it nearly always becomes a 2 with some preparation: writing down the rules, naming the owner or measuring for a few days. That work is part of the diagnosis, not a delay.

Examples of processes that usually go first

In services businesses, these processes tend to score well. They are examples, not a closed list: your company's order comes from its own table.

• Sorting the requests that arrive by email or web form and drafting a reply with the company's rules, which a person approves before it is sent.
• Organising incoming invoices and preparing drafts for the accounts, which someone confirms before they count.
• Finding companies that match your ideal client and drafting a first message, which a person reviews before it is sent. This is what the Legathum platform does.
• Preparing reports from data that already exists and flagging what falls outside the plan.
• Following up on pending tasks and reminding whoever has to act.

They all have one thing in common: the agent prepares and the person decides. That is not a limitation of the first process; it is how any hybridised process is designed.

Which processes not to automate yet

Some processes are best left out of the first cycle, or kept in but always with a person in charge:

• Decisions about people: shortlisting candidates, deciding promotions, allocating tasks based on individual behaviour or ending an employment relationship. The EU AI Act treats these systems as high risk (Annex III, point 4).
• Decisions with legal effects on someone, or that similarly significantly affect them, if they were taken without human involvement (Article 22 GDPR).
• Processes nobody can explain: the judgement has to be written down first, and that is already a project.
• Processes whose rules change often: the agent would be working with rules that no longer apply.
• Processes without an owner: without someone to review, an agent has nobody to pass on what it cannot resolve.

Not automating yet is not the same as not touching. A process about people can have an agent that organises the information and a person who decides; what it cannot have is an agent that decides on its own.

What the law says about the choice

Choosing the first process is also a compliance decision. These are the rules that weigh most in Spain and the rest of the EU:

• EU AI Act, Articles 14 and 26 and Annex III: for high-risk uses, human oversight must be assigned to people with the necessary competence, training and authority. Following the Digital Omnibus on AI (Regulation (EU) 2026/1744), these obligations apply to Annex III systems from 2 December 2027.
• EU AI Act, Article 50: if the automated process talks to clients, they must be told they are dealing with an AI. It applies from 2 August 2026.
• GDPR, Article 22: the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects someone, with the exceptions the article sets out. The EDPB guidelines explain how it applies.
• If you employ people in Spain, Article 64.4.d of the Workers' Statute gives the works council the right to know the parameters, rules and instructions behind algorithms that affect working conditions.

In the UK, the ICO's guidance on AI and data protection covers the same ground under the UK GDPR; the ICO says it is under review following the Data (Use and Access) Act. Starting with a low-risk process is not just prudent: it leaves time to design the oversight of the next ones properly.

How the decision is made in practice

In the Legathum AI-First Hybridisation Method, the choice of the first process comes out of the diagnosis, in this order:

1. Map the decisions in each candidate process: who decides, with what information and by which rules.
2. Measure the baseline of those that pass the first filter.
3. Score them with the table and choose one, with its owner.
4. Write its rules in a Book of Criteria that a person reviews and approves.
5. Implement it with human oversight at the critical points.
6. Compare against the baseline and choose the next one.

During implementation, the first result in production arrives within 2 days, measured against the baseline. The full phases are in how to implement AI in a company in phases, and the area-by-area work is that of the AI-First hybridisation consultancy.

Once every area has been through this cycle, the result is an AI-first company: one that decides, process by process, what agents do and what people do.

Keep reading

• What is company hybridisation with AI?
• How to implement AI in a company in phases
• AI-first company: what it is and how to build one