How to implement AI in a company in phases

If you are working out how to implement AI in a company without losing control, work in three phases: a diagnosis that measures the baseline and picks the first process, an implementation with human oversight, and ongoing support that compares the result with that baseline. No phase starts until the company has approved the previous one.

This article is part of the Legathum AI-First Hybridisation Method.

In short

AI is implemented process by process, one at a time, not tool by tool.
The diagnosis decides what to automate first and leaves the baseline measured.
Implementation writes down the company's rules and puts a person at every critical point.
Ongoing support measures against the baseline, corrects and adds the next process.
Each phase has its own approval and its own legal obligations.

How to implement AI in a company: in phases, one process at a time

How to implement AI in a company has a short answer: in phases, and one process at a time. Buying subscriptions for the whole workforce on day one usually achieves the opposite of what you want: everyone uses the tool their own way, nobody knows which criteria it answers by and there is no way to measure whether it has worked.

The destination is an AI-first company: one that decides, in every process, what AI agents do and what people do, with written rules. The route is company hybridisation with AI, and this article sets out its three phases.

In the Legathum AI-First Hybridisation Method those phases are diagnosis, implementation and ongoing support. Between each one there are approval points signed off by the company itself: nobody moves to the next phase without the process owner's go-ahead.

The three phases at a glance

Each phase answers a different question and leaves documents that stay with the company:

The phases for implementing AI in a company
Phase	Question it answers	What is written down	Who approves
Diagnosis	What to automate first, and what not to?	Decision map, baseline and the order of the processes	Management and the owner of each process
Implementation	How does the agent work and who reviews it?	Book of Criteria, test cases and checkpoints	The process owner, before it goes into production
Ongoing support	Has it worked, and what comes next?	Measurement against the baseline, real use and documentation to run what was built	Management, before the next process is added

Phase 1: diagnosis

The diagnosis says where AI makes sense and where it does not, before anything is built. It looks at how the work is really done, not how the manual describes it, and it leaves these pieces behind:

Decision map: in each process, who decides, with what information and by which rules.
Baseline: how long the process takes today, how much manual work it needs and which errors it produces.
Order of the processes: which goes first, which comes later and which stay with a person. We explain it in which processes to automate with AI first.
Legal risks: whether any process touches decisions about people, which the EU AI Act treats as high risk.
Owners: a named person for every process that goes in.

A diagnosis that does not leave a measured baseline is an opinion. The baseline is what later lets you say, with the company's own figures, whether the implementation has worked.

Phase 2: implementation

Implementation turns the chosen process into work shared between agents and people. It follows these stages:

Capture the criteria: write down the company's rules with the people who apply the process today, in a versioned Book of Criteria that a person reviews and approves.
Prototype in a day: a first agent running on those rules, and an explicit decision to carry on or stop.
Build with tests: real cases the agent has to handle correctly before it goes into production.
Set the checkpoints: what the person reviews, at what moment, and what happens if they disagree with what the agent proposes.
Deploy: the agent becomes part of the team's daily work.

During implementation, the first result in production arrives within 2 days, measured against the baseline. What is built works with the company's knowledge, held in its Master Brain, rather than with what a general model assumes.

Phase 3: ongoing support

Ongoing support checks that what was built is used and works, and decides the next step. It has three tasks:

Measure against the baseline, using the same measure as the diagnosis.
See whether the team works with the agent or around it. An agent nobody uses has solved nothing.
Correct the rules: every exception that comes up is written down, reviewed and added to the Book of Criteria.

Once the process works, the company keeps the documentation and knows how to run what was built, and the next process from the diagnosis is chosen. Cycle by cycle, AI reaches every area without any phase being skipped.

Which legal obligations come in at each phase

Implementing AI in a company in Spain or elsewhere in the EU touches several rules, and each one weighs on a different phase:

Diagnosis. GDPR, Article 35: if processing that uses new technologies is likely to result in a high risk to people, the impact assessment is done before you start. In Spain, the AEPD has guidance on processing that incorporates AI.
Diagnosis. EU AI Act, Annex III: systems involved in hiring or in decisions about employment relationships are high risk. Following the Digital Omnibus on AI (Regulation (EU) 2026/1744), their obligations apply from 2 December 2027.
Implementation. Article 26: for those uses, human oversight must be assigned to people with the necessary competence, training and authority.
Implementation. Article 50: if the agent talks to clients or to anyone else, it must be clear that it is an AI. It applies from 2 August 2026.
Ongoing support. Article 4: take measures to support the AI literacy of your staff, taking into account their knowledge and how the systems will be used. It has applied since 2 February 2025 and, after the Digital Omnibus, does not require you to guarantee any specific level of literacy for any individual.
Throughout, if you employ people in Spain. Workers' Statute, Article 64.4.d: the works council has the right to know the rules of 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. None of these obligations holds an implementation back if it is considered from the diagnosis onwards.

Common mistakes when implementing AI

Almost all of them come from skipping a phase or doing them all at once:

Starting with the tool and looking for the problem it solves afterwards.
Implementing in every area at once: if something improves, nobody knows why, and if it gets worse, nobody knows either.
Skipping the baseline and measuring only impressions.
Leaving the rules inside an agent's instructions, where the company cannot review them.
Naming someone to oversee without the time or the authority to correct the agent.
Leaving the EU AI Act until the end, when it shapes the design from the start.

If you want to see how the three phases would look in your company, the area-by-area work is that of the AI-First hybridisation consultancy.

Frequently asked questions

Where do I start implementing AI in my company?

With a diagnosis that maps the decisions in each process, measures the baseline and says which one to automate first. With Legathum, the first step is a 15-minute meeting with Deibson Silva.

How long does it take to implement AI in a company?

It is done process by process. With Legathum's method, the first result in production arrives within 2 days of implementation starting; the rest of the company moves forward in cycles, one process at a time.

Do I need an in-house technical team?

Not to start with. You need an owner for each process, someone who knows the rules and time to review what the agent does. At the end of the ongoing support phase, the company keeps the documentation to run what was built.

What is the baseline?

The measurement of the process before you touch it: how long it takes, how much manual work it needs and which errors it produces. It is what lets you compare afterwards using the same measure.

Which EU AI Act obligations already apply?

AI literacy (Article 4) since 2 February 2025 and transparency when someone talks to an AI (Article 50) since 2 August 2026. The obligations for Annex III high-risk systems apply from 2 December 2027.

Does implementing AI mean replacing people?

No. Agents do the repetitive work using the company's rules, and people approve, decide about other people and answer for the result.