Companies continue to announce artificial intelligence projects with ambitious figures, but a substantial portion of those initiatives never produce a visible result. The discussion usually focuses on the chosen technology or the quality of the model. Sebastian Brickel, an AI governance advisor, points elsewhere: the internal power structure that decides, without saying it openly, which projects stay alive and which ones fade away.
Brickel is an AI Governance Advisor at The Remote Consultant. Before dedicating himself to corporate artificial intelligence governance, he worked in computational chemistry, a field in which he completed a doctorate focused on simulating the behavior of molecules. That scientific background, he says, has given him a different perspective than the typical business consultant: less attention to corporate discourse and more to what actually happens within an organization when an AI tool is introduced.
The moment an AI project dies
According to Brickel, most artificial intelligence initiatives do not fail due to model limitations, but due to a lack of internal support. And the way they die depends on who promoted them.
Projects developed internally, without an external partner, rarely have a clear end: they simply stop receiving resources until they fade out. Projects with a strong internal sponsor, on the other hand, tend to survive well beyond the initial scope of the pilot, sustained by the logic of the cost already invested. No one officially declares them finished.
"The moment is rarely technical. It is internal politics," summarizes Brickel.
The only projects that end in a visible way, he explains, are those that had weak internal support from the beginning: sooner or later, someone with authority decides to cut them.
What an AI governance audit really examines
For Brickel, an AI audit does not primarily analyze a system, but the habits of an entire organization. The first step consists of mapping all existing artificial intelligence uses in the company, separating them into regulated and unregulated uses. The next step is to identify the person responsible for each system.
Brickel insists that AI "is a reasoning machine, not a moral agent" and that, therefore, it cannot assume the authority to decide on its own. Each system needs, according to him, a specific person to answer for it, not a committee or a junior employee who signs off without real decision-making power.
AI that enters without permission
A good portion of what is used daily in companies, Brickel points out, has never undergone a formal review. A team adopts a tool because they find it useful, without asking for authorization, following the usual pattern of what is known as shadow IT.
There is an additional nuance that, according to Brickel, is often overlooked: the artificial intelligence that is already integrated into SaaS programs contracted for another purpose and that performs AI tasks without anyone having explicitly pointed it out. That hidden layer is, for him, what makes any subsequent attempt to regulate the actual use of the technology more difficult.
The warning sign in an executive
Brickel identifies a pattern that, in his experience advising management teams, anticipates problems: expecting results that are too ambitious, too fast, and with too little budget. The more enthusiasm a person shows for the technology, the more likely the project is to have difficulties, he notes.
For Brickel, AI belongs to the same category as automation, and in many workflows, deterministic rule-based automation remains the correct answer. The phrase "put AI on everything" is, in his opinion, the best indicator that a project is going to go wrong. The sequence he recommends is different: analyze the workflow, optimize it, automate what can be automated, and only then incorporate artificial intelligence where it offers something that automation cannot provide.
What the EU AI Act really produces
Regarding the European AI Act, Brickel offers an unusual reading. In his view, the regulation only really improves systems in companies that already intended to do the work well and use the regulatory deadline as an internal lever. In the rest of the cases, he maintains, the result is a system with the appearance of compliance, which retains the same underlying problems, now simply documented.
Brickel recalls that regulations are usually announced well in advance, but that most companies wait until the deadline approaches to act, which generates last-minute documentation designed to satisfy an auditor and not to change how a system is built or supervised.
To support that idea, he cites a 2023 analysis by appliedAI of more than one hundred enterprise AI systems, which concluded that 40 of them could not be clearly classified into the risk levels provided for by the regulation, a fact that, according to Brickel, reflects the true level of preparation of many organizations.
High-risk systems: an important exception
Brickel clarifies that, in systems classified as high-risk under European regulations, the system itself also has direct obligations: conformity assessment, documentation, and human oversight, not just control of the data that feeds it.
In the rest of the cases, he argues that the decisive factor is not so much the conformity of the model itself as that of the data pipeline that feeds it. A practical alternative to building a fully compliant proprietary model, he points out, is to govern that data flow: feed the model with anonymized and correctly processed information. If the pipeline is poorly designed, he warns, no quality of the model can compensate for it.
When a failure is rebranded as a pilot
Brickel observes that companies rarely openly admit that an AI project has failed. The usual thing, he says, is to discreetly rebrand it as a "pilot" or "proof of concept."
In his experience, failure is rarely due to a technological problem: it is, above all, a failure of change management, something that is repeated in both cheap and expensive projects. He compares the situation to that of a new CRM or ERP that no one ends up actually using. The result, he points out, is usually summarized later as "we tested AI," instead of admitting that "we couldn't get people to use it."
A different look from science
Brickel attributes part of his approach to his scientific background. Having worked with machine learning before it became a buzzword, he says, gave him a technical foundation on how artificial intelligence reached its current state. Studying the behavior of the natural world, he adds, teaches that some rules are followed regardless of trends or the internal politics of an organization, a different perspective from that formed solely within a corporate career.
Frequently asked questions
Who is Sebastian Brickel?
He is an AI Governance Advisor at The Remote Consultant. He arrived at artificial intelligence governance from computational chemistry, a field in which he completed a doctorate dedicated to simulating the behavior of molecules.
Why do AI projects fail according to Brickel?
He maintains that most fail due to a lack of internal support and change management, not due to technical limitations of the model. Political support within the organization, he says, determines whether a project survives beyond its pilot or fades out without noise.
What does he say about the European AI Act?
He considers that the regulation only produces truly better systems in companies that already intended to do the work well. In the rest, he states, it generates systems that appear compliant but maintain the same underlying problems, now documented.
What is the AI that enters a company without authorization?
Brickel refers to two phenomena: AI tools adopted by teams without undergoing a formal review, and artificial intelligence already integrated into SaaS programs contracted for another purpose, which performs AI functions without anyone having explicitly pointed it out.
What sign does Brickel detect in an executive that anticipates problems in an AI project?
The expectation of results that are too ambitious, in a short time, and with a very reduced budget. For him, excessive enthusiasm for technology is usually accompanied by more difficulties in the project.
What did the 2023 appliedAI analysis cited by Brickel reveal?
According to that analysis of more than one hundred enterprise AI systems, 40 of them could not be clearly classified into the risk levels provided for by the regulation, which for Brickel reflects the real level of preparation of many organizations.
