AI in the coming years: directions and limits

This study is a conceptual, partly speculative analysis of how artificial intelligence may reshape the knowledge professions and the service industry over the next five to ten years. I am less interested in the models and algorithms themselves than in the system-level problems that appear once AI is in mass use.

My premise is that the biggest problems of the coming period will most likely come from the infrastructure missing around AI, and much less from what AI can or cannot do. As AI gets built deeper into banking, legal, healthcare, engineering and public administration processes, four questions become pressing: whether the data is authentic, whether the context is structured, whether decisions can be audited, and whether AI systems are secured and kept under regulatory control.

The study takes these four problem areas one by one. The first is the authenticity of data on an internet where a growing share of digital content is artificially generated. The second is AI's information gap: in many cases these systems are not wrong, they simply work from incomplete context. The third is whether AI-assisted decisions can be reproduced and who is accountable for them. The fourth is the AI-specific security and governance layer that enterprise environments still lack.

I do not split the material into separate "problem" and "solution" blocks. Each topic follows the same order: where we are today, what exactly the missing layer is, what is at stake, what infrastructure could fill the gap, and what we would gain from it. The answers I sketch are a digital provenance and authenticity layer, a structured context management model, an auditable AI decision-support infrastructure, and an AI governance and validation framework.

The study is not a standards proposal and not an implementation guide. It tries to frame, at the system level, the questions that technology companies, businesses and regulators will probably have to answer in the coming years.

The purpose and genre of this paper

This is not an academic paper in the classic sense, and it is not futurism or technology hype either. It is a conceptual analysis of the structural changes AI may set off in the knowledge professions and in digital services over the next few years, and of the problems those changes bring with them.

I am writing for people who already use AI routinely, understand roughly how it works, and have it in their daily work, so I skip the basics. The question for these readers is where all this is heading and which missing systems they will soon run into.

I look at the likely directions of AI's development partly from a technology angle, partly from an economic one, and partly from the infrastructure level. The models themselves are in the background; in the foreground are the problems and missing systems that become visible as AI spreads.

My central claim is that the biggest AI-related problems of the coming years will most likely show up in the infrastructure and environment around the models, and less in the models themselves. For each topic I therefore also sketch a possible solution.

AI today: the big picture

Public talk about AI swings between two extremes. One side says that within a few years artificial intelligence will take over a large part of human work. The other says it is all overhyped and AI is one more tool among many. I think neither describes what is actually happening.

The change is happening somewhere other than where most people look for it. Benchmarks, spectacular breakthroughs and a few percentage points on a test get the attention, while the real shift is in documents, emails, analyses and case handling, the everyday material of knowledge work.

In a bank's back office, a law firm, a hospital's administration or a university department, the question is no longer whether they use AI, but how far it has already disappeared into their processes. A request like "summarize this", "check this" or "suggest an alternative" is routine.

This change is not spectacular, but I do not think it can be reversed. It also goes beyond individual tasks: it changes how professions and sectors are structured.

General-purpose language models have reached the level where, in certain tasks, they compete with the average professional and often do better. Generating structured text, summarizing documents, translation, finding patterns in data, generating code and running repetitive analyses are everyday tools now, not experiments.

This can easily give a misleading picture, though.

AI does not work the way a person does. It does not see the context a professional works in: it does not know the earlier decisions, the unwritten rules, the organizational culture or the compromises of the past. On every task it behaves like a new colleague who is technically skilled but knows nothing about the place.

This limitation will not go away with the next model. It follows from how the technology works.

Since the environment AI works in directly shapes what it produces, it will become especially important to feed it reliable, accurate, verified data. The quality of the answers will depend on the input data and the information the AI has access to, more than on what the AI "knows".

Validating the input data follows from this. An AI can produce an answer that looks logical or professionally sound at first glance, but it cannot judge whether that answer is the right decision in a given business, legal or human situation. A mathematically precise result can be irrelevant to a concrete organizational problem, and a well-drafted legal text can be useless in a particular case.

The third issue is responsibility. AI takes no responsibility for its decisions, and in the medium term it will not. If an AI-assisted decision turns out to be wrong, the consequences land on a person: the one who approved it, signed it or used the result.

The lack of context, the limits of validation and the absence of responsibility together will define the role of AI, and the position of the industry built on it, in the coming years.

The near future: when AI becomes a baseline

Over the next year or two the AI assistant will become a basic tool of knowledge work. Today nobody is competitive without email or spreadsheets, and before long a professional working without AI will be in the same position.

This does not mean everyone becomes an AI expert. It means that using AI becomes part of normal professional practice. The difference between people will be how well they use the tool they have: whether they can state the problem precisely and hand over the right data, whether they understand the result, and whether they can decide if that result is acceptable.

A less visible change is starting in parallel, and it is discussed almost daily: entry-level positions are changing. If AI can do a large part of the classic junior tasks, those roles gradually empty out. In the short term this is an efficiency gain. In the longer term it creates a structural problem.

If no junior layer forms, where will the senior professionals of the future come from?

A senior person's expertise is built over years of experience, mistakes and slow learning. If that path is cut off, the consequences will arrive late, but they will arrive.

The same pattern can appear in law, finance, healthcare and every other knowledge profession, not only in IT.

The medium term: trust becomes the bottleneck

Over a three-to-five-year horizon the structure of knowledge work keeps changing. Professionals will work with several cooperating agentic systems instead of a single AI system.

One collects data, another analyzes it, a third writes up the result, a fourth checks it. The professional's role gradually moves from direct production toward coordination, verification and final decision-making. They end up leading a kind of digital team, and their effectiveness will depend on how well they organize it.

At first this looks like an efficiency gain, but the consequences go further.

The language of the professions changes. Today a knowledge worker writes concrete documents or has AI write them, has spreadsheets filled in or edits processes by hand. Within a few years they will more and more often describe rules and intentions at a higher level of abstraction, and AI will translate these into concrete, executable behavior.

A lawyer will write a set of rules from which AI generates the concrete contract for a given client, instead of writing each contract. A doctor will define a protocol that AI adapts to the patient's current condition, instead of drawing up each treatment plan.

Look at what we use AI for today. Typically to produce output meant for people: a spreadsheet, a document, a presentation or code. We send it to someone else, who more and more often also processes it with an AI. So AI-generated content is read by another AI, and between the two sits an intermediate step formatted for humans, which is not the natural language of either machine.

The next step follows from this. At some point systems will stop handing data to each other through human formats and will do it directly, in a machine representation that a person may no longer be able to read, only interpret with the right AI tool. The first signs are already visible in the new agent protocols: Anthropic's Model Context Protocol (MCP) and Google's Agent2Agent (A2A) protocol standardize how one AI system hands context and tasks over to another.

Something similar may start at the level of programming languages. Today AI typically works in existing, human-oriented languages, but more and more languages and frameworks are built specifically for language models, such as LMQL, Microsoft Guidance or DSPy, which turn prompting into a programmable, structured operation. I can easily imagine that within a few years we will see formats and languages designed by AI and for AI, where human readability is not a primary concern.

This creates a new problem. Wrong decisions will still happen, and the typical cause will be that the AI worked from incomplete context, not that it reasoned badly.

Most human knowledge is implicit. It consists of experience, background knowledge, past compromises and unspoken assumptions, not of documented rules. AI only sees what it is explicitly given. If it does not get an unspoken rule, it does not take it into account.

That is why structuring context and deliberately building in human control points is becoming more important. A system can only decide well if the information it has is complete.

The four topics of this paper

Conversations about AI's development still focus mostly on the capabilities of the models: how accurate a system is, how much better than the previous generation, how many parameters it has, which benchmarks it passes. These are fair questions, but they are less and less where the real problem is.

The technology has reached the level where the biggest problems appear in the model's environment, not in the model itself.

Whether AI can give a good answer is now a secondary question. The primary ones are what data it works from, how complete its context is, whether its decisions can be traced back, whether its behavior can be manipulated, and who takes responsibility when it gets things wrong.

The next four chapters look at four structural problems that I expect to shape AI's next decade. I do not consider them technical bugs; they are signs of missing infrastructure. All four point to the same gap: today's systems lack the layer that would, in the long run, make AI reliable, verifiable and usable inside the institutions society runs on.

Each of the four chapters stands on its own and covers one problem area, in the same order: where we are today (situation), what exactly the missing layer is (the problem), what is at stake, what infrastructure could answer it (the proposal), and what we gain from it (why it works).

Situation

Modern AI systems are built on data. That is not news: machine learning has always been only as good as its training data. What is new is that until recently the vast majority of internet content was written or made by people, and that is changing fast.

The problem

The web is filling up with AI-generated text, images and video. Some of it is useful, some is mediocre, and some is noise. The fact that AI generates content is not the problem in itself. The problem is that the next generations of models will learn from this content.

The process feeds itself.

If the quality of training data drops, the models get worse. The weaker models produce even more low-quality content, which later ends up in the training data again. The system does not settle into a stable equilibrium; quality keeps decaying.

Underneath is one question: how do we tell authentic data from artificially generated noise?

What is at stake

This may look theoretical, but the practical consequences show up quickly. Medical research builds on published results. Legal analysis relies on precedents and documents. Financial models work from market data and analyses. If more and more of these are AI-generated, unverified or distorted, the foundation itself becomes uncertain.

For a long time the digital world assumed that whatever can be read or seen is somehow tied to reality. That assumption is fading.

Today images and videos can be convincingly faked, and so can voices, documents, conversations and entire online identities. Soon "I saw a video of it" or "I heard his voice" will no longer count as evidence.

In the future the information worth paying for will probably be the kind whose authenticity can be proven, not the kind that is easy to produce. That creates a new need: infrastructure for validated, traceable data authenticated as coming from a person.

The proposal: infrastructure for digital authenticity

Whether AI can produce realistic content will not be the hard question of the coming years. At this point it can be taken as given.

The hard question is how to prove that something is real.

In this environment digital content on its own stops being evidence. The emphasis moves from the content to its origin.

The existence of a video will be worth little. What will be worth something is whether it can be proven:

• when it was made,
• with what device,
• what modifications it went through,
• and who it can be tied to.

That requires different infrastructure from today's internet.

Today's web mainly stores and transmits content. One of the main infrastructures of the next era will probably be a provenance layer: systems that make the origin and authenticity of content provable, instead of trying to "protect" the content itself.

One of the first serious technical efforts here is the C2PA standard. In C2PA, digital content carries metadata that makes its creation and modifications traceable. For a photo, for example, it becomes possible to reconstruct what device took it, what software touched it, whether it was edited afterwards, and what steps happened before publication.

That alone does not solve the problem.

It is not enough to know whether a piece of content is authentic; we also need to know whether the person behind it is. I expect one of the biggest changes of the AI era to be that digital identity gradually comes apart from the real person. It will be less and less obvious that we are talking to whoever we think we are talking to. Soon a video call, a voice message or even an entire online meeting will not be proof on its own.

This is where a further infrastructure layer can come in: biometric and behavior-based validation combined.

Simple face recognition or voice identification will not do, because AI will learn to reproduce both fairly quickly. What I have in mind is a multi-layered authenticity model that looks at biometric patterns, behavioral traits, interaction patterns, device identity and the full digital chain of origin together.

Digital origin and authenticity look like a technical question, but they are at least as much an economic one.

On today's internet, producing content is cheap and verifying its authenticity is expensive. AI widens this gap dramatically. In the coming years the cost of producing digital content will approach zero, while proving authenticity gets harder and is worth more.

That will create a new market of its own.

I mean a market for trust infrastructure, whose job is to manage the authenticity of data, content and digital identities, and not just one more market for AI models or generation tools.

In this model the value lies in validated origin, not in the content.

The first layer of the solution I have in mind is a validated AI training data platform. The AI industry today needs enormous amounts of data, but high-quality, verified and legally clean data is very scarce. Most models collect content from the internet, where origin, quality and legal usability are often uncertain.

In the long run this cannot work.

As AI models learn more and more from AI-generated content, authentic, licensable and validated data sources become more important. For future AI systems the bottleneck will be the reliability of data, not its quantity.

The answer could be a platform where institutional data owners, for example healthcare organizations, engineering firms, legal databases, research institutes, universities and banks, publish their own validated data in a structured form for AI training. It would not be aimed at the mass market. The main target would be high-value professional knowledge bases, not bloggers or general educational content.

A university research institute, an engineering firm or a healthcare organization could, for example, publish authenticated data packages with provable origin, known authorship, a validation process and a settled legal usage model. That is very different from today's "scrape everything off the internet" approach.

The platform could accept validated data from individuals and from other data-producing companies as well as from institutions, and the trust placed in the data would depend on the level of validation. The other part of the system is a validator layer made up of users.

Verification would not be purely central; a community of professional validators would share it. Validators build reputation and check the quality, consistency and authenticity of the data, and the system keeps weighting the reliability of their past work. This resembles Stack Overflow's reputation model, the logic of rating systems like chess Elo, or even IMDb's scoring. One explicit goal is to prevent data influencers: a piece of data should not become authentic just because a validator with a high reputation approved it.

The validator community could also do more than check data inside the platform. It could examine the authenticity of any document, image, video or other digital content already on the internet. In such a model the technical integrity of content could be recorded with cryptographic hashes and timestamps, which proves that a given file has not changed since a given point in time.

That still does not mean the content is real. Whether it is real would depend on the validator network and on the validators being accountable.

Authenticity in this system is a trust score that builds up gradually, not a yes-or-no state. The platform keeps weighting the accuracy and reliability of past validations, so over time the authenticity of a piece of content rests on accumulated professional trust as well as on technical metadata.

Over time this adds up to a new kind of digital trust system.

If several validators rate a freshly published piece of content as real, it can quickly reach the level of trust at which the system treats it as authentic.

If it later turns out that the content was not real after all, the reputation of those who rated it authentic drops. That can also change how the other content they validated earlier is judged.

A validator account is tied to a specific person, verified through biometric and behavior-based methods, possibly with document checks as well. Over time a validator's reputation therefore becomes a metric with real value (see: Black Mirror).

The business model has several sides. AI companies or enterprise customers can buy validated data packages, and the revenue is split automatically between the platform, the original data owner and the validators. This gives an economic incentive both to publish high-quality data and to keep validating.

The technical layer of the concept is a biometric, KYC-based cloud key management infrastructure. In this model the C2PA signing key would never leave the cloud HSM. Registration would go through a bank-grade KYC process, and access would be tied to biometric validation. Biometric data must not be stored centrally, though. A model that holds up in the long run needs a zero-knowledge-proof approach, where the biometric data never goes into a central database. The key would then be tied to a validated human identity instead of a device.

This would be a new way of handling digital authenticity.

Why it works

The long-term value will be provable human origin, not the content. Oddly enough, being real may become one of the scarcest resources of the AI era.

The consequences may be economic and social as well as technical. As more and more of the digital space can be generated artificially, environments may appear where validated human presence is a premium category in itself.

Some online platforms will distinguish between authenticated and unauthenticated content. Companies and institutions will require that certain processes use only documents or communication of validated origin.

This may resemble how HTTPS became a normal part of the web. On the early internet, authenticated and encrypted connections were not the default. Today a service without one is practically unthinkable.

I expect the same to happen with digital origin and human authenticity. On the future internet the question will be less what we see online and more whether we believe it is real.

As the amount of AI-generated content grows exponentially, licensed and validated data of provable origin may be worth more and more.

Situation

One of the biggest limitations of AI systems is often a lack of context, more than a lack of reasoning ability.

Much of human communication is implicit. Inside an organization there are countless rules, pieces of background knowledge and compromises that nobody writes down, yet everyone takes into account when deciding, ethical rules or widely known social norms for example. AI only sees what it is given.

Today's prompt-based approach often assumes that all the relevant context of a problem fits into a single question or instruction. Most complex professional decisions have far more layers than that. They rest on rules, exceptions, organizational practice, history and human judgment at once.

An experienced lawyer, for example, does not decide on the basis of the law alone. They take into account the client's risk tolerance, case law, earlier cases, and often informal considerations that are hard to write down. The same is true of medical decisions, engineering design and the running of a company.

Most of this never reaches AI systems today. As AI becomes a decision-support tool in more and more areas, structuring context and deliberately building human control points into processes becomes more important.

The problem

AI rollouts are often discussed as if they were technology projects: a new model, a new Copilot, a few integrations, a couple of workshops, and the organization suddenly becomes more productive.

In practice they look much more like ERP rollouts did twenty years ago. On paper everything seemed logical there too. The system could handle purchasing, inventory, finance or manufacturing, and the demos worked. Yet companies ran into the same problem again and again: the technology worked, but the organization's real way of working collided with the system.

ERP rollouts were the first time it became really clear that a company runs in a far less structured way than it looks from the outside, or than the people working there believed from the inside. Departments called the same thing by different names, the official processes did not match actual practice, and many of the rules lived in people's heads and in unwritten customs instead of documents. The exceptions mattered more than the standard process. You could see this when a new colleague arrived: training took a long time even if they already knew the ERP system and had worked in a similar area, because they still had to learn the local habits and unwritten rules.

The main lesson of the ERP era turned out to be organizational, not technical: operations can only be automated reliably once they have been structured and defined as processes.

AI brings the same problem to the surface, much more harshly.

ERP systems could still live with operations based on unwritten custom. People made phone calls, coordinated, kept the exceptions in their heads or in Excel, or simply "knew" how the organization really worked.

AI cannot live with hidden context.

If the concepts are not consistent, if the rules are not clarified, if the same process exists in five documents in five different forms, the AI has no stable knowledge to work from. If many rules exist only as unwritten custom, the AI will try to infer them from fragments, and will most likely get them wrong. Then come the complaints from employees: "AI can't even do a simple task…" Of course it can't, if there are a thousand ways to do it and the right way has never been written down.

This is a different question from the previous chapter, which was about whether data is real. Company data today is overwhelmingly authentic data produced by people. The difficulty is whether it is consistent and complete, and whether it is used in the right context.

I think this is where one of the biggest misunderstandings of the AI era lies.

When an AI makes a mistake, people tend to read it as a human mistake: it misunderstood the problem, reasoned badly, it is dumb or not competent enough. The explanation is often much simpler. The AI works from what it sees, like a junior on their first job who can only work from what they have been told.

Structuring all this looks like a huge task that many will consider impossible, but it is not. It takes little more effort than training a new colleague. That sounds much easier, of course, because with a colleague I just tell them how we usually do things.

And you can't do that with AI? Yes, you can!

A simple example of missing context: if I tell an AI "write a program that adds two numbers", a few seconds later I really do get working software. But that program will most likely not do exactly what I want. The AI did not make a mistake; the request lacks the whole environment. It does not say what system the program is for, what error handling it needs, what platform it runs on, whether it gets its input through a GUI or on the command line, or whether it is a teaching example or part of a bank's transaction system.

The request does not contain enough information for a single correct solution to exist.

The same happens in almost every knowledge profession. When someone tells an AI "give me legal advice for my divorce", the system can give advice, but it does not see the human background, the financial situation, the custody conflict, the case law, or the informal considerations an experienced lawyer takes into account by instinct. A medical diagnosis may lack the full clinical picture in the same way, a corporate process the organizational culture, an engineering decision the history of past compromises.

The AI is reasoning from an incomplete model of the world, and its answer is only as good as that model.

What is at stake

AI is not dangerous in itself. The danger is that above a certain level of complexity, missing information can lead to failures across a whole system.

I expect this to be one of the main realizations of the coming years: the next phase of AI's development depends on better models and also on whether we can build knowledge and context infrastructure on which these systems can operate reliably.

If the problem really is missing information, then "better AI" will not solve it on its own. Past a certain point the bottleneck is how much the model sees of the world in which the decision has to be made, not how intelligent it is.

Most companies today are not yet ready for AI.

Most of them do have AI. What holds them back is that their operation still rests on implicit human context. Most corporate knowledge today lives in Word files, PDFs, emails, meeting notes, Excel files, or in people's heads. Employees can see the connections between these, assume context and fill in the missing pieces from experience. AI is far more sensitive to unstructured knowledge.

So in the coming years a new kind of organizational problem will probably appear.

Next to technical debt and data debt we may start talking about context debt: how much of an organization's operation still relies on implicit knowledge, undocumented exceptions, partly contradictory processes, or simply on the people who "know by heart" how the system really works.

This is more than a documentation problem. It will directly decide how reliably an organization can use AI.

Within a few years a company may be measured by a kind of context health score, next to digital maturity or cloud readiness. It would look at how consistent and structured the organization's knowledge is, how much ambiguity and contradiction it contains, and how well AI systems can interpret it, instead of the state of servers or infrastructure.

The proposal: infrastructure for structured context

If the bottleneck is the information and not the model, the core idea follows: instead of trying to make the AI ever "smarter", what if we worked on the environment it works from?

What if organizational knowledge existed as structured, connected, machine-interpretable context instead of isolated documents? This looks like a technical problem, but it is much more an organizational and cultural one.

It moves the attention from the model to the intermediate layer that turns human intent into something AI can process. A new infrastructure layer appears between raw human knowledge and the AI systems that execute.

I call it an AI navigator layer.

It does not make the decisions, and it does not execute them either. It is an intelligent intermediate environment that structures context, connects related concepts, detects contradictions, and shows the points where information is incomplete or ambiguous.

This layer is also AI, and it continuously monitors and analyzes the organization's data. It does not answer questions or propose decisions. Its job is to keep the data in order, and to find and flag anomalies, missing concepts and contradictory statements.

The decision-making AI then works from a prepared, structured knowledge space instead of the chaotic world of documents written for people, and with the right presentation, people can use the same knowledge space too.

This is needed because most human problems pass through several levels of abstraction before they reach an executable form. First there is the intent, the often imprecise human statement of what we want. That becomes a more structured set of requirements, then a formal model, and finally the concrete execution: program code, a workflow, a contract or an automated process.

This translation has always existed. A lawyer, a systems analyst or an engineer moves between levels of abstraction all the time. What AI changes is that the transition can now happen in a structured and traceable way, where until now it happened invisibly, in people's heads.

So I am not proposing another chatbot or developer tool. I am proposing a living context system.

The professional works in natural language, as today, describing the problem, the need, the exceptions and the environment. In the background several navigator AI systems work together to monitor and structure the process.

One navigator AI analyzes the text in real time. It recognizes concepts, notices contradictions, flags missing definitions and tries to structure the context. It does not decide for the professional; it works like an analyst or reviewer who is always there.

Another AI tries to build a structured set of requirements from the finished document. A third might look for conflicts, a fourth might check compliance or security issues. The system consists of several specialized AIs working on the same structured knowledge space, with no single "super AI".

Meanwhile the document and the knowledge being edited keep getting "smarter", because the AI makes visible the knowledge that today stays hidden between documents. The AI does not take control in the process.

In the background there are no traditional documents at all, only a shared semantic representation. The text is just one view of it. The real system is a web of relationships, a knowledge graph in which concepts, rules, actors and contexts are linked to each other. Human knowledge is not linear like a document either. The meaning of a rule can change with the environment, two statements can partly contradict each other, and the same concept can mean different things in different organizations or industries. People handle this by instinct. AI has to represent it explicitly.

I think this is where the real infrastructure of the next generation of enterprise systems begins.

Better models are only part of it. The other part is systems that connect human intent with machine execution without losing context between the layers. In the long run the most valuable AI will be the one that understands the most about the world in which the question came up, not the one that gives the best answer to the question alone.

Why it works

The idea is not new. People have handled complexity this way for thousands of years, by creating intermediate levels of abstraction.

Law does not describe reality directly; it builds a structured system of concepts, definitions and rules. Engineering design happens through models, drawings and calculations before it reaches the physical world. In software development we think in specifications, architectures, components and abstractions, not in machine code.

Complex systems work because there is an intermediate representation that both people and the executing system understand.

AI makes it economically feasible, for the first time, to maintain this layer continuously. Building such a knowledge system used to be very expensive and hard to keep alive. Models, ontologies and knowledge graphs required manual work, which is why they spread only in a few specialized areas, and continuously checking and updating their content was practically impossible.

AI can now do much of the structuring automatically. It can recognize concepts, uncover relationships, find contradictions and keep the representation of knowledge up to date. It does not replace human expertise; it extends it.

So the proposal is an infrastructure in which human knowledge exists as a shared, continuously updated context space that machines can also interpret, instead of a fragmented heap of documents, emails and individual experience.

This may turn out to be one of the main lessons of the AI era: the competitive advantage will go to the organizations that can structure their own knowledge so that AI actually understands it, and not necessarily to those that adopt AI first.

Situation

Much of modern society is built on auditability. A bank's decision must be traceable years later. For a medical diagnosis it must be possible to reconstruct what information the decision was based on. In a court case it is a basic expectation that the judgment can be justified and reproduced.

The way AI works today is hard to reconcile with this.

A model gives an answer today. A few years later the same model no longer exists, the training data has changed, the environment has changed, and the same question gets a different answer.

The problem

How serious this is depends on the stakes. In some areas it does not matter at all: when generating creative text, for example, it is fine for the result to differ every time. In a credit assessment, an insurance decision or a medical diagnosis it can have serious consequences.

Companies would already like to work through decisions with AI, such as credit assessment, legal judgment or medical diagnostics, but they don't dare, because the result is not reproducible, not auditable, and cannot be defended legally. An AI decision made in 2026 would come out differently three years later with the same question, because the model and the context have both changed. In front of a court, an insurer or a regulator that is untenable.

The problem is legal and organizational as well as technical.

What is at stake

Today responsibility always ends up with a person: the one who signed or approved the decision. In practice, though, in more and more situations that person can verify the AI-generated result only to a limited extent.

That combination is dangerous. AI's decision support keeps getting stronger, while reproducibility and auditability stay weak.

As AI gets built deeper into regulated industries, there is a growing need for systems that can reconstruct which model ran, what data it worked from, in what context it operated, and what human approval took place.

Otherwise decisions will keep getting faster while the chain of accountability behind them falls apart.

The proposal: infrastructure for auditable decision-making

The answer is an infrastructure that treats a decision as a process that can be followed from start to finish, not as a single moment. The starting point is the case.

A case lives for weeks or months, like a project: a credit assessment, a court proceeding, a patient's treatment path. It is one continuous conversation thread that everything goes into: documents, AI prompts, AI responses, queries, partial decisions. It is closed in milestones, following a preconfigured workflow with decision points.

When a new case is opened with the same client (an appeal, a complaint, a follow-up visit), the complete material of the previous case is available as read-only input. Context keeps building up, and old decisions cannot be modified after the fact.

All content that goes in or comes out is hashed with SHA256 and stored Git-style. If a law or internal policy has not changed, it is only referenced, not duplicated. The context is a hash list (manifest) pointing into the store. For external sources (legal databases, drug databases, BKR, Hungary's central credit register), the time of the query and its result are saved too. Storage is used efficiently, and everything stays reviewable.

For reproducibility, the model used must remain runnable in exactly the same way for years. Public cloud models (OpenAI, Anthropic API) are ruled out, because their stability cannot be guaranteed even by contract. The solution is frozen open-source models such as Llama, Mistral or Qwen, whose stored weights the company owns. A frozen model never changes, and five years later it still runs bit-for-bit the same way.

Reproducibility can be understood at three levels, and choosing the target level is a real decision.

Result reproducibility means that the same prompt with the same context and the same model gives the same answer three years later. It is expensive and technically hard, because it needs stored or runnable model versions, context snapshots and deterministic sampling.

Process reproducibility means that we can see what was asked at the time, what the AI answered, what the person decided, and on what data. A new AI would give a different answer today, but the process at the time can be reconstructed. In the target market (courts, law firms, banks) this is not acceptable: "today we would get a different answer" is not something you say in a courtroom.

Rationale reproducibility means that the reasoning behind the decision can be traced back: why the AI suggested this, and why the person accepted it.

In the target application areas the goal is result reproducibility and rationale reproducibility together. The first makes it possible to replay the same situation exactly years later, the second keeps the decision professionally understandable and defensible.

An auditable process is not enough on its own. Knowing exactly what happened guarantees neither that the result can be reproduced nor that the decision can be justified professionally.

Not every use case needs the same level of reproducibility guarantee. In a court proceeding or a medical decision, provability and replayability are hard requirements, while a business consulting or analysis task allows much more leeway. So the system can be implemented at several guarantee levels:

Level 1 – Full reproducibility, operated on the client side

The model and the full context stay in the organization's own environment. This is for courts, hospitals and pharmaceutical companies, where the highest level of provability is required.

Level 2 – Platform-guaranteed reproducibility

The system runs on managed, multi-tenant infrastructure, but the reproducibility of the model and the context is still guaranteed. This may suit banks, insurers and law firms.

Level 3 – Context guarantee only

The system does not guarantee that the model stays unchanged, only that the full context and the rationale remain traceable. This may be enough for business consulting, analysis or lower-risk use cases.

At every guarantee level the operating model follows the same principle: the system manages the full life cycle of a case instead of isolated question–answer exchanges.

Several professionals can work on one case at the same time. A medical case may involve a consultation panel, a legal case a team of lawyers, and a credit assessment may pass through several levels of approval. So the system models a shared context space, not the relationship between one user and one AI.

Before the system is put into use there is a preparatory phase, onboarding. Business analysts or domain experts map the organization's decision-making processes and define the milestones, decision points and approval steps that can occur for a given case type.

When a specific case starts, its initial context is created: the available starting information, the case's participants and their roles, and the specifics of the case are recorded. From then on the case exists as an independent context space that keeps growing.

In the active phase new information is recorded, the consultation with AI takes place, and partial decisions are made. Every event, question, answer and decision becomes part of the case's context and is kept in a searchable form.

When the final decision is made, the case is closed and the full context is frozen. From then on the information that influenced the decision can no longer be modified.

After closing it is still possible to analyze the case and ask explanatory questions. The system can answer why a given conclusion was reached or what information led to a decision, but these later interactions do not change the original context and cannot override the earlier decision.

Why it works

The approach can work because it does not try to make AI infallible. It keeps the decision process transparent, traceable and reproducible.

Regulated professions do not actually require decisions to be error-free; human decisions are not error-free either. What they require is that a decision can be reconstructed and defended professionally years later.

If we know exactly which model ran, what information it worked from, what context it saw, what reasoning led to its conclusion, and who approved the final decision, then AI-assisted decisions fit into the same accountability framework as any other professional decision.

Open questions and trade-offs

The approach does not answer every problem, and I leave several questions open on purpose.

One is how model quality changes over time. Frontier models will probably improve faster than frozen open-source models, so a system that is adequate today may look weak in a few years compared to the state of the art. I see this as the price of reproducibility, not a flaw in the approach: stability and continuous improvement pull in opposite directions.

Another is the long-term preservation and runnability of models. If a decision has to be replayed ten or twenty years later, the entire runtime environment has to be preserved along with the model weights. That makes it an infrastructure and archiving problem as well as a software problem.

A third is how to handle post-mortem analyses. In a later investigation, may new questions be asked? If so, which answer counts as authoritative: the original model's, a newer model's fresh answer, or both? Technology, regulation and industry practice will have to answer this together.

How these questions are answered will largely decide whether AI can become a provable, auditable and legally defensible decision-support tool in the coming years.

Situation

As enterprise AI systems spread, a new set of security and regulatory problems has appeared.

Classic IT security covers many areas: network protection, access management, encryption, vulnerability management. AI adds new attack surfaces.

One of them is the model itself.

The problem

Prompt injection, for example, is especially dangerous because for the system, instructions and data often look the same: both are natural language. An instruction hidden in a document or email can change how the AI behaves without the user noticing.

Meanwhile more and more organizations are trying to build their own closed AI systems. Banks, hospitals, insurers, government institutions and pharmaceutical companies cannot let their sensitive data end up in public models, so the number of on-premise AI deployments is growing fast.

That requires a new kind of expertise.

Classic IT security knowledge is not enough. Professionals need to understand how the models work, the risks of training data, prompt-based attacks, regulatory compliance and the governance problems of AI systems, all at once.

What is at stake

The market is not really prepared for this yet.

AI security is a fragmented field today. There are new standards, governance frameworks and technical recommendations, but no unified, mature infrastructure or professional practice.

As AI gets built into more and more critical systems, a new layer of the industry will probably form whose job is auditing, validating, overseeing the security of and checking the regulatory compliance of AI systems, rather than developing models.

The concrete difficulty sits at three points: data isolation (the AI must not look into another patient's chart, another client's loan or another court's case), resistance to prompt injection, and mapping regulatory compliance. Classic IT security expertise does not cover these today, and there is no established industry certification. The big consultancies talk at the governance level, but at the depth of an actual deployment there is a gap.

The proposal: AI governance and validation

This gap is where a new, highly specialized consulting layer could appear: providers that do nothing but security audits, validation and regulatory compliance for AI systems.

At first the business would rest mainly on consulting, but a proprietary, public and documented standard could take shape alongside it. The standard defines the requirements an AI system must meet in data isolation, protection against prompt injection and regulatory compliance.

As a first step the client receives a detailed gap analysis and an implementation proposal. Later they can also get certified through an independent audit. The two activities have to be separate: the same organization cannot advise a client and then certify its own work, because that would destroy its independence and the credibility of the whole scheme.

The proprietary standard is a transitional tool, not the end state. The industry will probably develop more mature, official standards and compliance frameworks in the coming years. ISO/IEC 42001 is evolving, the EU AI Act's conformity methodology is settling, and new industry practices are appearing. So the proprietary standard should be designed from the start so that its control points map easily to the later official requirements. The goal is to let clients move smoothly into the official certification schemes later, not to lock them into a proprietary ecosystem.

Certification can consist of several parts that support each other. The client receives an official certificate they can include in their compliance documentation, together with a publicly verifiable digital attestation and a registry entry. They also receive a detailed control catalog and scoring showing where they meet the requirements and where gaps remain. Behind all of this is a detailed audit report that can be used internally and for regulators.

Such a certification scheme would probably be built in several stages.

In the first stage the goal is credibility and reference clients. The standard and the certification are still entirely a private initiative, so the main task is to deliver the first successful projects and become known in the profession.

In the second stage partnerships with accredited certification bodies may follow, for example with TÜV, BSI or DEKRA. The specialized provider brings the AI-specific expertise, and the partner brings the accreditation and legal backing.

In the third stage such an organization could become an independent, accredited player itself and enter the official certification market for high-risk AI systems.

Every new certification scheme starts with the same problem: on day one nobody knows the stamp. Without recognition it has no market value, without market value it has no clients, and without clients it gets no recognition.

This has happened before. ISO 27001 and SOC 2 did not become baseline industry expectations overnight either. At first they were one new approach among many, and the market had to come to trust them gradually.

Several things help build trust. One is full transparency: the standard and the control points are public, verifiable and repeatable. Another is expertise with names attached, since the market trusts people and the experience behind them first. The first high-prestige clients count just as much, and so does having an independent party review from time to time how the certification works.

The public registry also gives legitimacy. At first it has only a few entries, but later it can become a trust infrastructure in its own right.

A certificate starts to count in the industry when outside players recognize it: when a regulator cites it as acceptable evidence, when an insurer offers a discount for certified systems, when clients start requiring it, or when competitors start building similar schemes.

From then on the certificate no longer depends on the credibility of the organization that issues it; it becomes a trusted institution of its own.

Why it works

The value of the approach is in what the big consulting firms and classic IT security approaches typically do not cover today: verifiable compliance at the level of the actual deployment. The client gets tangible proof that their AI system meets defined security and regulatory requirements, in addition to recommendations and governance principles.

This counts most where data isolation, protection against prompt injection or compliance with regulatory requirements has to be proven. The certification then becomes a shared language that the client, the compliance function, the regulator and later the official certification bodies can all read, and not just a document.

The approach also does not lock the client into a proprietary ecosystem. If the control system is built from the start to be compatible with the emerging industry standards, it gives a natural path to the official frameworks later, such as ISO/IEC 42001 or the EU AI Act's conformity scheme.

The risks should be stated openly too.

There is no guarantee that a proprietary certification scheme will ever be broadly accepted in the industry. Many similar initiatives stayed niche or disappeared within a few years. There are usually two reasons: either the market does not consider the problem serious enough to pay for compliance, or a bigger player builds a similar offering faster and organizes the market around itself.

In the most cautious scenario the proprietary standard remains mainly a methodological framework for the consulting work. It helps the work with clients, gives the audits a structure and standardizes the professional approach, but does not become a market standard of its own.

I would not call that a failure. A public and consistent control system that connects today's AI security practice with the official compliance requirements expected in the future is worth something on its own. If the market consolidates and the methodology is partly or fully absorbed into the official certification system, the original initiative can become one of the building blocks of later industry standards.

Where to invest

Suppose you have seed capital and want to invest in technology that will address a real market problem five years from now. These are the directions I would look at.

• Data provenance and validation. This is the subject of the chapter on data authenticity. As AI-generated content floods the web, the value of human-validated and certified data will grow exponentially.
• AI audit and explainability. Regulators already demand that AI decisions be explainable. In five years it will be mandatory in finance, medicine and law.
• Energy-efficient AI infrastructure. The energy demand of large models is unsustainable. Small, specialized models running locally (edge AI) will be a huge market.
• Digital twins for infrastructure. Real-time digital replicas of bridges, pipe networks and buildings. Whoever predicts failures saves lives and money.
• Cybersecurity against AI. AI writes attack code as well as it writes defensive code. Deepfakes, automated social engineering and infrastructure attacks are already here, and defense is in its infancy.
• Tools for human–AI collaboration. This is the subject of the chapter on structured context. Every industry will need them, and nobody has solved it well yet.
• Automating regulatory compliance. GDPR, the AI Act and sector-specific regulations keep multiplying and getting more complex. Large consulting areas will open up.
• Rethinking education in the AI era. I do not mean e-learning. The question is what a person needs to know when AI knows everything. Critical thinking, judgment and validating AI output have to be taught, and nobody knows how yet.

If someone held a gun to my head and I had to pick two, I would say data provenance and human–AI collaboration. The first because everything else rests on it, the second because it is the nearest real market.

Speculation: my own outlook for 2036

Ten years from now knowledge work will probably have changed beyond recognition. It will not disappear; it will become much more specialized.

Some professionals will structure business, medical or legal knowledge for machines. Others will build systems from this structured knowledge. There will be roles whose main job is validating and auditing AI systems and managing accountability for decisions. The professions in the traditional sense will not disappear either, but they will probably retreat into narrower, higher-value areas.

"Programmer", "lawyer", "doctor" or "accountant" as a general job title will probably become as imprecise as "IT person" is today, a label that hides more than it reveals.

Data may indeed be the new oil, but not in its raw form. In the oil industry the real value came from refining, and in the world of data the most important resource will be structured, authenticated and legally settled datasets. The winners will probably be those who can do this "refining".

Regulation will catch up with technology, but probably only partly. The AI Act and its successors will set frameworks, and the detailed rules will build up gradually from industry self-regulation, professional practice and court precedents.

The service industry will change a lot. Physical presence will become rarer, and in many areas the AI assistant will become the first point of contact. The bank branch, the customer service office or the insurance advisor will not disappear completely, but may become the exception.

Several risks are already visible.

One is hidden knowledge. If people no longer understand why an AI reached a conclusion and ask another AI to explain it, knowledge can gradually drift out of human oversight.

Another is resolving ambiguity too early. AI tends to settle uncertainty quickly and present a single solution as fact. In doing so it can lose the alternatives, exceptions and compromises that in many professions are the basis of a good decision.

A third is where responsibility sits. If the answer to every decision is "the AI suggested it", responsibility gradually evaporates. In the long run societies will have to define again who is responsible for an AI-supported decision, and under what circumstances.

Platform monopolies are a serious risk as well. Companies that own the user's data, the AI assistant and the service infrastructure at the same time may gain influence on a scale we have not seen before. A seemingly "friendly" digital assistant can easily become a gatekeeper serving commercial or political interests.

The digital divide will probably deepen further. Those who have no access to these tools, do not trust them or cannot use them well may drift further and further from the economic and institutional center of society.

What stays human?

Probably taking responsibility, deep understanding of context, seeking compromise, building trust and professional ethics: the ability to make a decision with incomplete information and then own its consequences.

AI may well help with these too. But in the end they are social and moral questions, and in a society responsibility is always borne by people toward one another.

This study made four claims.

First, within ten years data authenticity will become a legal and market requirement. The technical foundations already exist in the C2PA standard. What is missing is the community of validators, the mechanism for distributing royalties, and the layer of biometric human authentication. This is an ethical problem and a business opportunity at the same time.

Second, AI makes bad decisions mostly because it works from incomplete information, not because its reasoning fails. So "better AI" will not solve it alone; what is needed is a more structured handover of information: layered specifications, living documents, and explicit mechanisms for handling uncertainty and gaps.

Third, in certain regulated professions AI-assisted decisions must remain reproducible years later. This needs dedicated infrastructure: version-frozen models, content-addressed storage and auditable decision processes.

Fourth, the security problems of enterprise, on-premise AI systems (data isolation, protection against prompt injection and regulatory compliance) need a new kind of expertise and new certification models. Today neither classic IT security nor the big governance-focused consulting firms fully cover this area.

The proposals above are possible directions, not finished solutions. If I look back at this document in five years and find that some assumptions were wrong, that is fine. The real mistake would be not to start thinking about these questions in time.

The transformation of software, engineering, medical or legal work does not mean these professions disappear. In every knowledge profession the rule-following, repetitive activities recede, and understanding context, taking responsibility, systems design, validation and managing trade-offs gain value.

The physical presence of the service industry will probably shrink a lot, and digital assistants and AI-based customer relationships will take more and more of its place. That raises serious ethical and social questions, and we need to answer them before a real person in a real situation finds out that the human help they once took for granted is no longer there.

• C2PA (Coalition for Content Provenance and Authenticity). Open standard. https://c2pa.org
• IEEE 29148:2018. Systems and software engineering — Life cycle processes — Requirements engineering. Industry-standard guidelines for writing requirements.
• ISO/IEC 42001:2023. Information technology — Artificial intelligence — Management system. Certifiable AI management standard.
• ISO/IEC 23894:2023. Information technology — Artificial intelligence — Guidance on risk management.
• NIST AI Risk Management Framework (AI RMF 1.0), 2023, and Generative AI Profile, 2024.
• OWL 2 Web Ontology Language. W3C Recommendation. https://www.w3.org/TR/owl2-overview/
• RDF 1.1 Concepts and Abstract Syntax. W3C Recommendation. https://www.w3.org/TR/rdf11-concepts
• DMN (Decision Model and Notation). Object Management Group. https://www.omg.org/dmn/
• Open Policy Agent. Open-source policy-as-code framework. https://www.openpolicyagent.org
• OWASP Top 10 for LLM Applications. https://owasp.org/
• MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems). https://atlas.mitre.org/
• EU AI Act. The European Union's Artificial Intelligence Act (2024).
• PSD2 (Payment Services Directive 2). EU directive that opened up bank data to competitors; the basis of open banking.
• Anthropic Claude for Excel; Skills feature. Anthropic's application integrations from 2025 onward.
• Straits Research. Global AI Training Data Market Report, 2025–2033.
• Grand View Research. AI Training Data Market Analysis.

Most of the ideas in this study took shape in long conversations over the past year and a half with colleagues from the tech world, whom I do not name here. The claims, concepts and proposals are my own.