Skip to content

AI-transform your organisation

AI guide for business leaders

Written

Author: Teemu Malinen

Contents

1 Why does AI stir up so much emotion?

In a survey run by the US organisation IIHS, 48% of more than 2,000 respondents considered it safe to take their hands off the wheel when the system was called “autopilot”. For four other driver assistance systems, at most 33% thought the same.[2] The participants had been told nothing about the systems beforehand apart from the terms themselves.

Language shapes the mind.

People intuitively read the word “smartphone” as nothing more than a handy device. The term “artificial intelligence” easily reads as an agent that thinks in a highly intelligent and often malevolent way, as science fiction has taught us. Before AI came into everyday use, the term already carried a threatening image.

We are also used to software working deterministically, producing the same answer from the same input. Generative AI, by contrast, draws its answer from a probability distribution, so the same input produces a different answer on different occasions. The language model behind AI therefore breaks our sense of control. The machine may not behave the way we wanted or expected.

Emotion is hard to shift with facts

Researchers ran an experiment with about 3,000 Italian companies, telling each of them how many of their competitors already use AI. The information did not change their intentions to use AI. The share of companies planning to adopt robotics did move: it rose from 37.6% to 44.3% once they heard the real level of use among their competitors.[3] The same kind of information shifted intentions on robotics but not on AI.

In a McKinsey study published in August 2026, 32% of the organisations that use AI regularly expected staff reductions because of AI within a year. The measured outcome was only 14%, less than half the expected share.[4] Even so, 39% of the same respondents expected staff reductions for the following year as well. Recent and comparable facts were available, and they still did not change what companies expected.

The dimensions of AI are hard to grasp

AI is not a single technology (such as an electronic signature) or a project (such as an ERP programme). It changes at least an organisation’s processes, learning, management, data and information security, all at the same time.

In one company, generative AI is already in daily use, but without proper governance, metrics or management.[5] In another, governance is strong, but use may be minimal (let alone the value it creates).

Something that is hard to grasp precisely is harder to measure. One consequence may be that the topic appears larger than it is. Even solid experience does not by itself make an assessment reliable (chapter 4.2).

No wonder the AI transformation of an organisation looks difficult when you do not know where to start or how to measure the benefits.

Summary for decision-makers

Single questions such as “do you make use of AI?” or “how far along are you?” rarely give a reliable picture of the real situation.[4][6] Several perspectives are needed, and together they give an overall view of where an organisation stands with AI.

The guide also introduces a maturity model that lets a decision-maker assess the AI maturity level of their organisation through concrete observations.

2 How big a wave is AI really?

New technology is readily declared revolutionary: cloud, mobile, the internet of things, blockchain and so on. The term “revolution” does not, however, classify the scale of the change. The deciding factor is the breadth of the economic impact.

2.1 Technology waves come in two sizes

In this guide I divide technology waves into two size classes. An economy-sized wave changes the principles of value creation broadly across industry boundaries. The internet, which later developed into digitalisation, was an example of this phenomenon.

An industry-sized wave, by contrast, changes a single industry or function fundamentally. Cloud services revolutionised procurement and maintenance in IT, for example, but this did not force engineering workshops to change their business models from the ground up.

The difference is decisive. An economy-sized wave sweeps over every industry in time. An industry-sized wave can be handled as a more limited decision, unless your own organisation happens to fall within that very industry. For a SaaS company, for example, moving to the cloud became the default model, but a traditional company could justifiably leave its servers in its own data centre.

The decision-maker’s task is to work out whether a wave is industry-sized or economy-sized. With the former, action is not necessarily needed (at least not immediately), but with the latter the impact is inevitable over some time horizon.

How much did the internet ultimately affect the economy?

In 2011 McKinsey carried out a study on the economic impact of the internet. In the study, the internet accounted for 3.4% of the gross domestic product of the thirteen countries examined. In developed economies, the internet explained 21% of GDP growth over the five years preceding the measurement.[7]

From the decision-maker’s point of view, though, the more interesting point is that 75% of the benefits produced by the internet arose in traditional industries, not in internet companies.[7] The technology’s greatest value came from applying it to business.

Adoption speeds up with every wave

According to research by Comin and Hobijn, it historically took an average of 45 years from the invention of a technology to its adoption, but the lag keeps shortening. Each new generation of technology is adopted faster than the previous one.[8]

ChatGPT reached a billion monthly users in about three years; the previous record was eight years.[9] In spring 2026, technology investment accounted for more than a third of economic growth in the United States.[10]

In June 2026 the BIS, the cooperative organisation of central banks, estimated that AI investment corresponded to roughly one percentage point of US GDP growth in 2025. The figure describes the demand effect of investment spending. The BIS reported time savings on the order of 20–50 per cent at task level.[11]

Even so, nearly nine decision-makers out of ten have seen no change in productivity over the past three years: so far, growth appears to have been generated by the investing itself.[12]

What is being forecast for AI?

McKinsey estimated in 2023 that generative AI would create 2.6–4.4 trillion dollars of new value annually.[13] Goldman Sachs estimated in 2023 that AI would eventually raise global GDP by about 7 per cent and annual productivity growth by about 1.4 percentage points.[14] The estimate describes a ten-year cumulative effect from the point at which the technology has been adopted by about half of all companies.[14]

The OECD and the IMF, among others, approach AI from a macroeconomic perspective and build global growth scenarios around it.[15][16] AI naturally cannot be compared directly with the internet, but the scale is visible, even though forecasts of its magnitude vary.

Summary for decision-makers

Is AI an economy-sized or an industry-sized wave? If it is an economy-sized wave, which part of the business will it affect first?

The premise of this guide is that AI is an economy-sized wave. The reasoning rests on a broad body of research, but the final magnitude will only become clear in hindsight.

2.2 How much does hype colour the AI debate?

When a new technology raises high expectations, the market may price the future over-optimistically. A bubble can then form on the stock market, and it may later burst as valuations collapse. A fall in share prices is not, however, directly tied to the technology or to the value it produces in organisations’ operations. In public debate, the technology and the bubble become confused.

The following sections give two concrete examples of how the market prices technology.

The dotcom bubble 2000–2002

The Nasdaq Composite rose to 5,132 points on 10 March 2000 and fell to 1,185 points on 23 September 2002. Around 4.4 trillion dollars of market value disappeared.[17] Pets.com ceased operations in November 2000 and eToys went bankrupt.[18] The crash is usually remembered for its failures.

The survivors have received less attention. Amazon’s share price fell by about 90%, but the company survived, as did eBay and Cisco. Google only went public in 2004, after the bubble had burst.[18] One of the biggest winners of the internet wave had not even listed yet when the whole phenomenon was already declared a failure.[18]

According to peer-reviewed research, the five-year survival rate of dotcom companies was nearly 50 per cent.[17] The figure is lower than in traditional manufacturing industry, where about two out of three new companies survive five years.[17] It does, however, correspond well to the early-stage development of other fast-growing industries.

A more precise interpretation is therefore that about half of the companies disappeared, which is an ordinary phenomenon when an industry is born. The stock market merely repriced expectations; the technology did not go anywhere.

SaaSpocalypse 2026

On 12 January 2026 Anthropic released Claude Cowork, a product capable of doing office work fairly autonomously directly on staff members’ own machines and work files.[19]

An interpretation emerged that the business model of per-user priced SaaS software was about to disappear. Why buy SaaS software aimed at a narrow task when a tool that can be personalised for your own company is available for a few tens of euros a month?

As a result, a stock market basket tracking well over a hundred US companies lost about 25% of its value in under six weeks. LegalZoom, for example, fell by nearly 20%.[19] The phenomenon was given the name SaaSpocalypse.

The mood changed very quickly. On 24 February 2026 Anthropic announced that it had partnered with precisely those software companies that had been feared to suffer most, such as Salesforce, Intuit and DocuSign.[20] Their share prices recovered quickly, and analysts judged the original wave of selling to have been overdone.[20]

In May 2026 the software sector had its best month since 2001: the iShares Expanded Tech-Software ETF rose 21% in a single month.[21]

So in four months the market priced the effects of the same technology twice, in two different ways. First strongly pessimistically, and then almost equally strongly optimistically.

Summary for decision-makers

It is worth keeping the valuation level and the benefit level of a technology separate. The former is determined by the market, the latter by each organisation itself. The benefit level is found by examining your own organisation’s business and processes rather than share prices.

The purpose of this guide is to help you understand AI and to encourage you to build a relationship with it that suits your organisation.

2.3 How does AI differ from earlier megatrends?

AI has been in use in the core business of pioneering companies for several decades already, in machine learning. Generative AI is new, but the mechanism is the same: data is accumulated and a model produces a prediction that guides decision-making. That sequence is repeated thousands or millions of times a day.

Amazon introduced its “item-to-item collaborative filtering” recommendation system in 1998. The scientific paper describing the method documented the system’s use and has since received the IEEE Test of Time award.[22][23]

RankBrain, published by Google in 2015, was in use in all of its searches by 2016.[24] In peer-reviewed research on Netflix recommendations published in 2026, switching to a plain top list would reduce the probability of viewing by 12% and switching to an algorithm ten years older by 4%.[25]

Summary for decision-makers

AI differs from earlier economy-sized megatrends in that there are already decades of evidence for how it works. Pioneering companies have been building AI into their core business since the 1990s.

The question is not, however, how to build the next Amazon or Google, but: “Which decision recurs in our organisation so often that improving or speeding it up by even a few per cent would produce significant business value over the long term?”

3 AI basics for decision-makers

3.1 Basic AI concepts

In the author’s experience, AI becomes easier to grasp when it is set alongside digital transformation, both in concept and in vocabulary. See the clarifying table below.

Digital transformation vs. AI transformation

Digital transformation vs. AI transformation A table-like figure in four columns: the digitalisation term, what the term means, the scale and the AI term. First row: digitisation means converting analogue information into digital form, the scale is the task, and the matching AI term is task automation. Second row: digitalisation means redesigning an organisation’s process or operations by making broad use of the possibilities of a new technology, the scale is the organisation, and the matching AI term is AI transformation. Third row: digital transformation means a broader societal change and phenomenon, the scale is society, and the matching AI term is AI shift. The column of AI terms is highlighted because it carries the conclusion of the figure: AI transformation is a change at organisation level, not task-level automation. Digital transformation vs. AI transformation AI transformation is a change at organisation level. Digitalisation term What does the term mean? Scale AI term Digitisation Converting analogue information into digital form. Task Task automation Digitalisation Redesigning an organisation’s process or operations by making broad use of the possibilities of a new technology. Organisation AI transformation Digital transformation A broader societal change and phenomenon. Society AI shift © 2026 Sofokus. AI-assisted visualization.

Digitisation means, for example, turning a paper form into an online form. Task automation is the analogy this guide uses from the world of AI. It automates a single task without using AI’s potential to redesign processes or ways of working. A meeting recording goes to a language model, which writes the minutes from it. The meeting arrangements, the distribution of the minutes and the decision-making do not change; only the writing work is automated.

Digitalisation is active work that the organisation does itself. Its aim is to redesign entire processes and thereby change the logic even at industry level, much as online services changed banking. As the counterpart to digitalisation, this guide proposes the term AI transformation.

AI transformation is a useful new term for a decision-maker. It does not describe a general societal phenomenon, but everything an organisation itself does to make broad use of AI in a way that produces measurable benefit. Other synonyms are the loanword forms AI transforming, AI-fication and AI-isation.

Digital transformation is the broader societal phenomenon, and in May 2026 the Ministry of Finance named its counterpart in the world of AI the AI shift.[26] A Sitra report from December 2025 speaks of an AI upheaval.[27] The AI shift is not something an individual organisation can choose. The phenomenon advances regardless of the measures taken.

As the above shows, Finnish-language AI terminology is very new and only now settling into established use.

3.2 The technological foundations of AI

The good news is that a decision-maker hardly needs to understand the technical details of AI. Grasping the following basic principles already takes you a long way:

  • What is generative AI?
  • What determines the quality of a language model’s answer?
  • What is an AI agent and what is the benefit of one?

At the end of this chapter there is a brief introduction to the interface standards a decision-maker may come across in, for example, requests for tender or technology discussions.

Generative AI answers on the basis of probabilities

A language model generates the most probable continuation for the input it is given. The model does not look the answer up in a database, nor does it know whether the answer it produces is true in the user’s situation.

Even when the model does not have enough information, it still produces the best answer it can. The same question can also produce a slightly different answer on different occasions, because the answer is chosen from a probability distribution. The less information a language model has on a subject, the more readily it hallucinates. Conversely, giving the language model more and better context is an effective way to improve the quality of an answer.

You can buy the thinking, you have to build the memory

The media focus on companies outdoing one another in the development of the leading models (ChatGPT, Claude, Gemini and so on). That focus can give a misleading impression of what really decides the outcome when AI is used.

The differences between models have narrowed quickly. Open language models are around four months behind the closed leading models.[28] McKinsey expects this development to lead to models gradually converging, with competitive advantage moving to data and context.[29]

In practice the quality of an answer usually depends more on what information the model is given at the moment it answers. A model responding to a request for tender needs the price list, the tender templates and the customer history, for example. Without proper grounding, even the best model’s answer falls short.

Organisations therefore build a shared context layer, known as organisational memory (“Organizational memory”), from which models and agents get the information they need. The brain (the language model) needs organisational memory so that what it does stays fit for purpose. This is why building organisational memory is a task for management, not an IT exercise.

Switching from one model to another is relatively easy and quick. Building high-quality context, on the other hand, is slow, and that is precisely why it turns into a competitive advantage. According to McKinsey, context accumulated over three years can set apart organisations that otherwise use the same software.[29]

An agent needs an understanding of the company’s data

A person knows without being told what a customer, an order or a margin mean in their own organisation. For a language model these have to be described separately. Without a clear and shared information foundation, agents draw different conclusions from the same data, and errors multiply as the system grows. According to McKinsey, eight out of ten companies regard gaps in data as one obstacle to adopting agents at scale.[30]

A chat answers, an agent acts

A chat answers the user’s question. An agent, by contrast, is connected to the company’s data and tools, which lets it carry out work steps independently: retrieve information, compare options, fill in forms, start processes or send messages.

AI chat vs. AI agent Two side-by-side connection diagrams with the same language model on both sides, so the difference comes from the connections rather than from the model. On the left, a chat: the user sends a question to the language model and gets an answer back, and there are no other connections; the model is not connected to the organisation’s data or systems. On the right, an agent: the same language model receives a task from the user and is also connected to the company’s data, tools and systems, which lets it carry out work steps independently: it retrieves information, compares options, starts processes and communicates. AI chat vs. AI agent Both have the same language model; the difference is the integrations. A chat answers User question answer Language model The model is not connected to the organisation’s data or systems. An agent acts User task Language model Company data Tools Systems An agent can retrieve information, compare options, start processes and communicate, among other things. © 2026 Sofokus. AI-assisted visualization.

The difference comes not primarily from the intelligence of the language model but from the systems, data and tools the model is connected to. In the AI maturity model presented later in this guide, agents running in production represent process-level capability, and in the chapter on secure use the central question is how extensive the permissions granted to an agent can be.

There is not yet an established practice for classifying agents. McKinsey, for example, uses different classification models in publications from the same year.[31] This guide keeps to a simple distinction between an AI chat and an AI agent.

How does AI change the way software is used?

Connecting software systems to one another is called integration, and what makes it possible is called an interface. Software interfaces have developed in three stages. In the first stage, every integration was built separately. Every pair of systems required a project of its own.

In the second stage, the API became a product. Salesforce, eBay and Amazon opened up their application programming interfaces in the early 2000s, and the value of software began to rest partly on how easily it could be connected to other systems.[32]

In the third stage, an AI agent uses the interface instead of a person. MCP is the current standard for this development. Anthropic published the standard in 2024, OpenAI and Google adopted it in 2025, and at the end of that same year stewardship passed to the Agentic AI Foundation under the Linux Foundation. According to Anthropic, at the time of the handover in December 2025 there were more than 10,000 active public MCP servers, and the protocol’s two software libraries were being downloaded more than 97 million times a month.[33]

If an agent can use systems directly through MCP, what happens to the per-user licences of a typical SaaS application?

Forrester estimated in February 2026 that the core operations of companies will continue to depend on SaaS software, but that pricing will shift gradually from per-user licences towards consumption- and outcome-based models. At the same time it forecasts SaaS spending to grow from 318 billion dollars in 2025 to around 576 billion dollars by 2029.[34] Gartner estimates that by 2030 agentic AI will apply to at most about a fifth of companies’ SaaS spending.[35]

Summary for decision-makers

Here are a few key questions for a decision-maker to put to their own organisation:

  • What information does a language model need so that it can give the best possible answers? Are those answers already within the machine’s reach?
  • Who is responsible for the accuracy and consistency of that information?
  • Which systems can agents be connected to, and how are they governed securely?
  • What off-the-shelf software do we have in use altogether? How would the pricing of that software change if an agent used it instead of a person?

4 How does the value of AI reach the bottom line?

4.1 Shipping containers and AI

In April 1956 the converted tanker Ideal-X left Port Newark for Houston carrying 58 shipping containers. The event is regarded as the starting gun for container traffic, but world trade did not change overnight. International containerisation only began to spread about ten years later, once container dimensions had been standardised and the various operators had committed to the same practices.[36]

The real benefit therefore emerged only once the whole system had been built around it. Ports needed new cranes, ships a new structure, lorries and trains compatible platforms, warehouses a new logic and dock work new ways of operating. The container was not merely a new means of transport but a shared agreement to change how everyone worked.

Research confirms the same phenomenon. Economists did not measure the effects of containerisation by looking at sea transport alone, but at the entire intermodal transport chain from the factory gate all the way to the customer. That is precisely where the productivity gain came from: goods no longer had to be unloaded and reloaded whenever the mode of transport changed.[36]

When the rest of the chain is stripped away and only port containerisation is left, the effect on world trade falls to roughly a third.[36] The same container in a different system delivers only a fraction of the benefit.

The effects kept growing for some fifteen years after both trading partners had moved over to the container system.[36] The largest benefits emerged gradually, as more and more operators adapted their own processes to the same operating model.

The same analogy applies to AI

If AI is used only to speed up tasks while the processes stay the same, the situation resembles port containerisation without railways or road transport. The new technology has been adopted, but the surrounding system remains unchanged.[37]

The decision-maker’s central question is not what AI could speed up, but which fundamental constraint it could remove.

Reckitt examined potential uses of AI in its operations, in customer service for example. Most of the candidates indicated returns and time savings, but management rejected them all the same. The reason given was that the individual cases were disconnected from one another and produced no strategic advantage.[38]

4.2 Where do the benefits of AI end up?

In a Danish study, employees’ survey responses were linked to nationwide register data, which made it possible to compare perceived time savings with the working-time and earnings effects measured for the same individuals. The employees estimated that they saved about three per cent of their working time. In the register data, however, the effect on earnings and hours worked was zero.[39]

Korean survey data showed working time falling by an average of 3.8%, but the amount of time saved was not connected to work output. Some of it moved into leisure within the working day, which the researchers read as a well-being benefit for employees rather than a problem.[40]

This is an important finding. Freed-up time can also create value through improved well-being at work, which may come, for example, from a greater sense of control or of getting things done.

There is, however, no broad research evidence so far on where the time freed up by AI goes. Freed-up time does not turn into business results by itself without management. Setting a meaningful and durable policy is therefore a shared challenge for the whole organisation.

The benefits are distributed unevenly

The benefit obtained from AI depends above all on the task.

AI raised the number of cases resolved per hour by an average of 15% in a customer service field trial, and by as much as 30% among inexperienced employees.[41]

In a trial with consultants, the results split in two. When the task stayed within the AI model’s area of competence, the quality of the work improved by 33.9%. When the task went beyond the model’s capabilities, the share of correct solutions fell by 19 percentage points.[42] Users do not usually notice that they have crossed this line.

Another organisation’s results do not predict yours

Even within the same occupational group, the results can be completely opposite.

Experienced open-source developers in the METR study slowed down by an average of 19%.[43] In another study, randomised trials at three companies showed developers completing 26% more tasks per week.[44]

The studies nevertheless do not cancel each other out. The tasks, the working environments and the participants were different.

Self-assessment is no substitute for measurement

In METR’s trial, experienced software developers estimated that AI had sped up their work by about 20 per cent, even though measurement showed the work had slowed by 19 per cent.[43] In the Danish register data, employees estimated that AI saved them about three per cent of their working time, but no effect on hours worked or earnings was observed.[39] In both studies the observed gap was systematic, not random.

Self-assessment vs. measurement Three rows in which estimate and measurement are set against each other. In the METR trial, experienced software developers estimated that their work had sped up by 20 per cent, but measurement showed a slowdown of 19 per cent; the gap is 39 percentage points in the wrong direction. In the Danish register data, employees estimated that they saved about three per cent of their working time, but the effect on earnings and hours worked was zero. In McKinsey’s August 2026 survey, 32 per cent of organisations expected workforce reductions within a year, and the outcome measured was 14 per cent. In every row the estimate is larger than the measurement, so the bias points in the same direction in three mutually independent datasets. Self-assessment vs. measurement The tendency to overstate in self-assessment shows in three studies. Estimate Reality Gap METR trial Experienced software developers, open-source tasks +20 % work sped up −19 % work slowed 39 %-points wrong direction Danish register data Employees’ own estimate matched against registers about 3 % work time saved 0 effect on earnings and hours 100 % of estimated saving not in the register McKinsey, August 2026 Organisations using AI regularly 32 % expected workforce cuts 14 % happened within a year >50 % fewer than expected © 2026 Sofokus. AI-assisted visualization.

Self-assessed time savings are not a sufficient basis for a business case.

Summary for decision-makers

The conclusions of this chapter can be summed up as follows:

  • Business cases should not be built on self-assessed time savings.
  • Process-specific before-and-after measurements give a more realistic picture of the potential time savings.
  • Directing the use of freed-up time is a management question. If it is not done, the time has already drifted elsewhere.

When the goal is business impact, what decides the outcome is not adopting the tool but changing the workflow.[45][46]

Scope note. The studies presented in this chapter are based on international data. No published impact studies of a comparable standard on Finnish organisations are available so far. The closest point of comparison is the Danish register study.[39]

5 What does AI require from an organisation?

5.1 The AI CODES model

The AI CODES model is the guide author’s application of what is known as the Owners–Directors–Executives chain (the ODE chain) to the AI era.[47] From the organisation’s point of view, the owner sets the ambition, the directors oversee and the executives implement. Together these form the traditional ODE chain. Although the chain alone does not determine an organisation’s AI maturity level, even a single missing link in the chain can make the organisation’s AI transformation significantly harder.

An organisation probably already has the traditional management codes in place, such as the owner mandate, the business strategy and the annual plan. The AI CODES model adds Customers and Specialists for an organisation aiming for high AI maturity. Although these roles are not part of the chain of command as traditionally understood, taking them into account over both the short and the long term is absolutely central to the success of AI transformation.

Below is an illustration of the AI CODES model together with a short description of how AI transformation typically looks to each of the roles.

The AI CODES model of a modern professional services company and its cross-cutting AI strategy At the centre of the figure is the Customer, with four roles around it: Specialists at the top, Executives on the left, Owners on the right and Directors at the bottom. Between each role and the customer in the centre there is a two-way arrow. The Customer covers expectations of the service level, customer need steering the service, and data steering the service and decisions; the code is the customer agreement, and the AI row is the impact of AI on the service level together with transparency of use and data terms. The Specialists cover values-based wishes and limits, working culture and community, psychological safety and meaningful work, and the chance to influence their own and the company’s development; the AI row is the ground rules of use and the responsibility for the quality of AI-assisted work, and the code is the culture handbook. The Executives cover a horizon of typically one or two years, execution and reporting to the directors, target setting and measurement, leading and coaching, and ensuring the internal flow of information; the AI row is AI goals, metrics and skills together with the allocation of the time that has been freed up, and the code is the annual plan, that is the budget. The Owners cover a horizon of typically more than five years, ambition and values, risk level, and growth or profitability; the AI row is the AI ambition, that is risk level and appetite for investment, and the code is the owner mandate. The Directors cover a horizon of typically two to five years, ensuring operational and strategic success, ensuring regulatory compliance, and ensuring the flow of information to external stakeholders; the AI row is the impact of AI on the business model and competitiveness together with oversight of regulation and risk level, and the code is the business strategy. The AI strategy is not a box of its own and not a member of the chain, but a cross-cutting layer: a dashed ring surrounds the whole figure, its interior is shaded as a single field, and in each of the five elements the AI row is a dashed box of its own at the bottom of the element. The AI CODES model AI strategy cuts across all roles in the AI CODES model. AI strategy Specialists Values-based wishes and limits Working culture and community Psychological safety and meaningful work A say in their own and the company’s development Code:Culture handbook Ground rules for use, quality responsibility for AI-assisted work Executives Typically a 1–2 year horizon Execution and reporting to Directors Goal setting and measurement Leading and coaching Ensuring internal communication Code:Annual plan (budget) AI goals, metrics and skills, allocating the time freed up Customer Expected service level Customer need drives service Data drives service, decisions Code:Customer agreement AI impact on service level, transparency of use, data terms Owners Typically a >5 year horizon Ambition, values Risk level Growth or profit Code:Owner mandate AI ambition: risk level and investment appetite Directors Typically a 2–5 year horizon Ensuring operational and strategic success Ensuring regulatory compliance Ensuring communication to external stakeholders Code:Business strategy AI impact on the business model and competitiveness, regulatory and risk oversight © 2026 Sofokus. AI-assisted visualization. The AI CODES model of a modern professional services company and its cross-cutting AI strategy The same figure for a narrow screen. In the wide figure the Customer is in the centre with four roles around it: Specialists at the top, Executives on the left, Owners on the right and Directors at the bottom. In this version the same five elements are stacked one below the other: first the Customer, then Specialists, Executives, Owners and Directors. The vertical line running along the right-hand edge corresponds to the two-way arrows of the wide figure: each of the four roles is in a two-way relationship with the customer in the centre. The Customer covers expectations of the service level, customer need steering the service, and data steering the service and decisions; the code is the customer agreement, and the AI row is the impact of AI on the service level together with transparency of use and data terms. The Specialists cover values-based wishes and limits, working culture and community, psychological safety and meaningful work, and the chance to influence their own and the company’s development; the AI row is the ground rules of use and the responsibility for the quality of AI-assisted work, and the code is the culture handbook. The Executives cover a horizon of typically one or two years, execution and reporting to the directors, target setting and measurement, leading and coaching, and ensuring the internal flow of information; the AI row is AI goals, metrics and skills together with the allocation of the time that has been freed up, and the code is the annual plan, that is the budget. The Owners cover a horizon of typically more than five years, ambition and values, risk level, and growth or profitability; the AI row is the AI ambition, that is risk level and appetite for investment, and the code is the owner mandate. The Directors cover a horizon of typically two to five years, ensuring operational and strategic success, ensuring regulatory compliance, and ensuring the flow of information to external stakeholders; the AI row is the impact of AI on the business model and competitiveness together with oversight of regulation and risk level, and the code is the business strategy. The AI strategy is not a box of its own and not a member of the chain, but a cross-cutting layer: a dashed ring surrounds all five elements and in each element the AI row is a dashed box of its own at the bottom of the element. The AI CODES model AI strategy cuts across all roles in the AI CODES model. Customer Expected service level Customer need drives service Data drives service, decisions Code:Customer agreement AI impact on service level, transparency of use, data terms Specialists Values-based wishes and limits Working culture and community Psychological safety and meaningful work A say in their own and the company’s development Code:Culture handbook Ground rules for use, quality responsibility for AI-assisted work Executives Typically a 1–2 year horizon Execution and reporting to Directors Goal setting and measurement Leading and coaching Ensuring internal communication Code:Annual plan (budget) AI goals, metrics and skills, allocating the time freed up Owners Typically a >5 year horizon Ambition, values Risk level Growth or profit Code:Owner mandate AI ambition: risk level and investment appetite Directors Typically a 2–5 year horizon Ensuring operational and strategic success Ensuring regulatory compliance Ensuring communication to external stakeholders Code:Business strategy AI impact on the business model and competitiveness, regulatory and risk oversight AI strategy © 2026 Sofokus. AI-assisted visualization.

Owners typically consider the development of the organisation on the longest time horizon. The owners’ most important task is to set the organisation’s ambition. Which maturity level are we aiming at, and how much are we prepared to invest in reaching those goals?

The directors‘ task is to secure the organisation’s competitiveness through the business strategy. Although AI is only one strategic theme, the organisation has good reason to take a position on it.

The executives‘ task is to carry out the chosen business strategy and to make sure that the things AI transformation involves happen in practice. This guide is meant to support that role in particular, because AI brings with it so many new things for decision-makers to deal with. One of the most central of them is the sensible management of the time that AI frees up.

The AI transformation of an organisation is above all an exercise in leadership.

A modern organisation takes its own specialists into account and involves them in AI transformation. It is safe to assume that the specialists are already using AI; the question is rather whether the organisation guides and supports them enough. Attitudes to AI range from fear of losing one’s job to burning enthusiasm and everything in between. AI may reshape familiar work roles at a rapid pace, and there is more on the subject in chapter 5.2. At the very least it is worth creating the ground rules together, choosing the tools and providing enough training.

From the organisation’s point of view, customers are linked to AI transformation along at least two routes, one of which the organisation can influence itself and the other hardly at all. The organisation can, for example, steer how the use of AI shows up in customer agreements or affects the service level. The market, in turn, decides how things such as customers’ buying behaviour change as AI becomes more widespread. What the organisation is left with is the choice of how to respond to that market change.

In the AI CODES model the AI strategy is deliberately placed as an element that cuts across every role, because most organisations already have the traditional codes in place. One practical way to go about AI transformation is to produce the first version of the AI strategy so that each role considers the themes that belong to its own field and the results are then drawn together.

In the target state there is no longer any need for a separate AI strategy, because its principles have been absorbed into the organisation’s normal management system. By then AI has moved out of separate projects and become a natural part of management practices, the customer promise, the culture handbook and so on.

Taken as a whole, far-reaching AI transformation is an operation to change the organisational culture.

5.2 How does AI change skills?

In June 2026 PwC published a study listing the tasks that were added to different professional roles between 2022 and 2025. The study scored those tasks by how much they rely on human capabilities such as empathy, vision, creativity or hope, and then grouped them by how strongly AI had affected the roles. The roles AI affected most gained 2.5 times more tasks than the professional roles where the effect was smallest.[48]

More interesting than the scores, however, was the logic by which AI affected the roles. The figure below sets out this guide’s reading of it.

Same automation, two outcomes Two rows, each showing the same span of tasks from routine to judgement-based work and the same skill gate in the middle. In a professionalising role, 22 per cent of advertised jobs, AI handles the routine and the judgement-based part is left to the human, so the bar rises and expertise separates performers ever more clearly. In a democratising role, 52 per cent, the human is left with work that more people can now do and AI handles the part behind the skill gate, so the bar falls and expertise no longer separates performers at all. Below the rows a brace: a design decision, the job designer decides which way the role develops. Same automation, two outcomes When AI handles the routine, the bar rises. When it handles the demanding part, the bar falls. Routine Skill gate Judgement-based work Professionalising role · 22 % of advertised jobs AI handles routine For the human judgement-based part Bar rises Expertise separates performers ever more clearly. Democratising role · 52 % of advertised jobs For the human work that more people can now do AI handles the part behind the skill gate Bar falls Expertise no longer separates performers at all. Design decision The job designer decides which way the role develops. © 2026 Sofokus. AI-assisted visualization.

How a professional role develops depends on where AI is put to work: on the routines or on the part that requires judgement. If AI handles routine tasks such as monitoring, the demanding decisions and the special cases are left to the human (compare the work of a radiologist). In practice the level of skill the role demands of the human goes up.

A skill gate is a threshold that has typically taken years of experience in the role to cross. If the work behind the skill gate is handed to AI, the level required to carry out a given task drops (compare the work of a translator).

For a decision-maker it is essential to take stock of the professional roles in the organisation and to consider which way each of them is worth developing with the help of AI. That view gives direction for training and supporting staff, and for reshaping career paths in a controlled way.

The human oversight burden

The relationship between automation and people has been studied for thirty years now. Under automation, human work shifts from an active role to a more passive supervisory one. If the system works correctly time after time, trust grows while attention weakens.[49]

In the laboratory, twenty minutes of other tasks is enough for supervisory alertness to fall.[49] Research on agentic AI also showed that the probability of failure grows with the length of the task.[50] In other words, as an error becomes more likely, spotting it becomes less likely.

When AI is introduced, it is important to plan which way the professional role will change, as set out in the previous section. As routine work decreases, the oversight load grows. If AI removes a demanding work step, the load falls and capacity is freed up elsewhere.[4]

5.3 Costs and budgeting

The Stanford AI Index tracks price at a fixed level of capability. In November 2022 the cheapest GPT-3.5 level model cost twenty dollars per million tokens, and in October 2024 seven cents.[51]

Although token pricing is relevant, from the point of view of AI transformation it pays to track other things as well. The sections below cover the different parts of the costs it involves.

Staff costs and AI

BCG offers a 10-20-70 rule of thumb for budgeting AI: a tenth to algorithms, a fifth to technology and data, and the rest to processes and people.[52] McKinsey, for its part, says that tokens take 20-25 % of agentic costs and human oversight the rest.[29]

AI cost distribution Two parallel distribution bars on a scale from zero to one hundred per cent. The upper bar is BCG’s 10-20-70 rule of thumb: a tenth to algorithms, a fifth to technology and data and the remaining 70 per cent to processes and people. The lower bar is McKinsey’s split of agentic costs: tokens take 20-25 per cent and oversight the remaining 75-80 per cent. In both splits human work is the largest cost item, that is, the costs of AI are mostly staff costs in high-maturity organisations. AI cost distribution Consultancies see human time as the largest cost item. BCG’s 10-20-70 rule 10 % 20 % 70 % algorithms technology & data processes and people McKinsey agentic costs 20–25 % 75–80 % tokens oversight © 2026 Sofokus. AI-assisted visualization.

In both, human work is the largest cost item. Somewhat surprisingly, then, the costs of AI are mostly staff costs in high-maturity organisations.

System costs

The maintenance costs of a modern IT system are estimated at 10-20 % of the cost of building it.[29] The recurring costs of generative AI solutions easily exceed the initial investment.[29] This is probably because taking AI deep into an organisation’s operations calls for integration-level solutions, and the products are developing fast. Tuning agents into working order is therefore not a one-off job.

As a yardstick for the investment, it pays to favour the cost of completed work. If it were, say, the number of agents or tokens in use, it would be easy to optimise the wrong things. In the end an organisation needs above all to understand its total costs and total returns.

The falling price of tokens may create the impression that costs are coming down, but a more realistic view is that all capacity will be used in any case and that those furthest along optimise their costs most effectively. Those who make the most effective use of agentic AI also use more tokens overall.[53]

Summary for decision-makers

On costs, a decision-maker should find out the following, among other things:

  • Have oversight and the AI transformation of processes been budgeted for separately?
  • What does the outcome of a given piece of work cost and what does it return? Which way is the cost moving?
  • Who decides on the AI budget?

6 Governed use of AI

“Can I feed this into an AI?” is a question asked more and more often in organisations. Behind it is the way the language models in AI tools work. User inputs are used to train the models, and as a by-product the data is stored on the vendor’s servers. With confidential information, or information covered by a duty of secrecy, you have to take the model’s runtime environment into account.

The diagram below sets out the main ways to govern how the information used by AI is stored. Note that in the diagram the organisation uses the AI model itself or takes it directly from the model vendor, not through a SaaS service, for example.

AI runtime environments A table-like figure with six rows stacked one above the other and three columns: the name of the runtime environment, a boxed action and an unboxed risk. The rows do not form an ascending order. Environment A, a public consumer service: an employee uses a public consumer service on their own account; what remains is that data may go into model development, there is no processing agreement, there is no log and nobody knows who inputs what. Environment B, a business account or API: a processing agreement, an agreed retention policy and no use in training; what remains are the data location and the subcontractor chain. Environment C, the EU region and a European operator: the model runs in the EU and a European company operates it; what remains are vendor dependency and a narrower model range. Environment D, an isolated environment of its own: the environment is the organisation’s own and isolated; maintenance shifts to the buyer. Environment E, a downloadable model in-house: a publicly downloadable AI model runs on the organisation’s hardware and data stays internal; compute, skills and update costs stay in-house. Environment F, off the network: the network is cut entirely; the model sees no external source while running. Environments B–E are surrounded by a dashed frame whose label is data encryption in use. AI runtime environments Runtime environment options for an organisation. Environment Action Risks Environment A Public consumer service Employee uses a public consumer service on their own account. Data may go into model development. There is no processing agreement, no log, no one knows who inputs what. Data encryption in use Environment B Business account or API Processing agreement, agreed retention policy and no use in training. Data location, subcontractor chain. Environment C EU region, European operator The model runs in the EU. A European company operates it. Vendor dependency and a narrower model range. Environment D Isolated environment Own and isolated environment. Maintenance shifts to the buyer. Environment E Downloadable model in-house A publicly downloadable AI model runs on the organisation’s hardware. Data stays internal. Compute, skills and update costs stay in-house. Environment F Off the network The network is cut entirely. The model sees no external source while running. © 2026 Sofokus. AI-assisted visualization. AI runtime environments The same content for a narrow screen: six sections stacked one above the other, in no ascending order. Each section has the number and name of the runtime environment, a boxed action block and, below it, the risk text pointed to by an arrow and shown without a box. Environment A, a public consumer service: an employee uses a public consumer service on their own account; what remains is that data may go into model development and that there is no processing agreement and no log. Environment B, a business account or API: a processing agreement, an agreed retention policy and no use in training; what remains are the data location and the subcontractor chain. Environment C, the EU region and a European operator: the model runs in the EU and a European company operates it; what remains are vendor dependency and a narrower model range. Environment D, an isolated environment of its own: the environment is the organisation’s own and isolated; maintenance shifts to the buyer. Environment E, a downloadable model in-house: the model runs in its own data centre and data stays internal; compute, skills and update costs stay in-house. Environment F, off the network: the network is cut entirely; the model sees no external source while running. Environments B–E each have a dashed frame of their own, and their shared label is data encryption in use. AI runtime environments Runtime environment options for an organisation. Data encryption in use Environment A Public consumer service Action Employee uses a public consumer service on their own account. Risks Data may go into model development. No processing agreement, no log. Environment B Business account or API Action Processing agreement, agreed retention policy and no use in training. Risks Data location, subcontractor chain. Environment C EU region, European operator Action The model runs in the EU. A European company operates it. Risks Vendor dependency and a narrower model range. Environment D Isolated environment Action Own and isolated environment. Risks Maintenance shifts to the buyer. Environment E Downloadable model in-house Action The model runs in its own data centre. Data stays internal. Risks Compute, skills and update costs stay in-house. Environment F Off the network Action The network is cut entirely. Risks The model sees no external source while running. © 2026 Sofokus. AI-assisted visualization.

Runtime environment A refers to an employee using a public AI model through their own account. That is typical of maturity level 1, and the risks make it inadvisable.

Runtime environments B and C are otherwise similar, but in the latter the model is run in the EU region and operated by a European AI company. Keep in mind that the location of the language model and its vendor can differ, for example when you use a SaaS product that runs on some model. Runtime environments D, E and F are options when the organisation does not want its own or its customers’ data to end up in the model.

Typically, runtime environments D and E are enough for organisations, because they can still be updated over the external network if desired. Runtime environment F is a rarer scenario, where the organisation wants to be absolutely certain that the environment has no connection to the external network.

The runtime environment does not on its own determine the security level

You cannot read the security level directly off the runtime environment, because in options E and F, for example, the organisation itself takes responsibility for setting it up and maintaining it. Self-maintained environments do not suit every organisation.

An organisation can use several runtime environments at once. Marketing data, for example, can be used with a business subscription to an AI service (B and C), while customer data is processed in the organisation’s own environment (D and E). Most organisations do indeed use more than one approach.[29]

Fin (formerly Intercom) trained its customer service model on its own customer service conversations. For that it used a 60-person team of specialists and its own computing infrastructure.[54]

The EU AI Act: what is in force now

The EU AI Act changed in 2026 (the “AI Omnibus” update, Regulation (EU) 2026/1744), so there is still plenty of out-of-date information about it online.[55] The AI transparency and AI literacy requirements are common to almost all organisations.

The requirements on staff AI literacy (“AI literacy”) came into force earlier, and the broadly applicable transparency obligations (Article 50) on 2 August 2026. They apply to every company that uses AI in its customer interface.

Chatbots and other systems have to show the user clearly that there is a machine at the other end. Videos, audio and images as well as text created with AI (synthetic content) must carry digital, machine-readable markings, for example watermarks.

Under Annex III of the AI Act, AI used in recruitment, such as the automatic screening of CVs or the analysis of video interviews, and in the management of employment relationships is classified as a high-risk system. Recruitment AI is therefore defined in law as a high-risk area, but companies now have until the end of 2027 to get their audits and technical documentation in order. The transparency obligation, however, is already in force.

Summary for decision-makers

Questions for a decision-maker in your own organisation:

  • What kind of data do we have, and which runtime environment would suit it?
  • Do we have the capability to maintain our own AI environment in-house?
  • Have we familiarised ourselves with the EU AI Act and made sure we comply with it?

7 The future of AI

Will the AI market concentrate like the platform economy?

In the assessment of Vipra and Korinek, the market for AI foundation models may develop in much the same way as digital platforms did in the early 2000s.[56] At first a large number of different players enter the market, and years of loss-making competition follow. In the end the market is whittled down to a few winners. An ever-accelerating flywheel (cf. the Amazon flywheel), economies of scale and end users’ limited willingness to switch all push the market towards concentration.

Platforms differ from AI, however. An end user gets the same benefit from their language model regardless of what others use. The result is a weaker network effect.[56] Markets may still concentrate, but that requires retaining users and accumulating data.

AI is moving into devices

An extreme example of concentration can be seen in SpaceX’s listing documentation, where the same person (Elon Musk) owns a cluster of companies developing AI models (Grok), telecommunications connections (Starlink), humanoid robots (Optimus), self-driving vehicles (Tesla), planned chip production (Terafab) as well as interplanetary data centres and habitation (SpaceX).[57][58] This is still only a vision, of course, but “Physical AI”, the concept the prospectus mentions, is worth keeping in mind. Smart lawnmowers are only the beginning.

Which resource will become scarce as AI spreads?

When a capability becomes cheap and generally available, it can no longer tell good companies apart from bad ones, whatever the starting point. Competitive advantage moves to wherever scarcity appears.[37]

Section 4.1 described the same mechanism through container shipping. The biggest change did not come from the container itself, but from the reorganisation of the system built around it. With AI, scarcity moves to coordination: to the ability to rearrange work, information and decision-making faster than competitors.

Which part of your own company’s expertise is becoming cheap and generally available, and where should the building of competitive advantage be moved instead?

The Finnish Parliament’s Committee for the Future has been developing the Radical Technology Inquirer (RTI) method since 2012 to anticipate how technological upheaval will change society over the next 20 years. The latest update to the report, from 2026, was produced with AI assistance and examines societal changes in the years 2026–2045.[59] Expert assessments may help a decision-maker grasp the changes in their own industry.

Summary for decision-makers

A decision-maker would do well to ask questions such as these:

  • Which AI-related decisions keep our operations as flexible and ready for change as possible?
  • If the AI market were to concentrate even more strongly, what would that mean for our organisation and its operating environment?
  • Which bottlenecks are likely to disappear from our own industry? When would that happen?
  • Which of our current capabilities will become commonplace and significantly cheaper?
  • What will this mean for our competitive advantage?

The task of decision-makers is to focus on identifying which way scarcity, value creation and dependencies will shift in the future.

Sources

  1. Tilastokeskus, Generatiivista tekoälyä käyttäneiden osuus nousi 41 prosenttiin (12.11.2025).
  2. Insurance Institute for Highway Safety (IIHS), New studies highlight driver confusion about automated systems, 20.6.2019; background paper What’s in a name? Drivers’ perceptions of the use of five SAE Level 2 driving automation systems (IIHS reference record 2190).
  3. Cullen, Faia, Guglielminetti, Perez-Truglia & Rondinelli, The Innovation Race: Experimental Evidence on Advanced Technologies. NBER Working Paper 34532, 12/2025 (PDF).
  4. McKinsey & Company, The state of AI in 2026: On the road to ROI, 25.8.2026.
  5. Gillespie, Lockey, Ward, Macdade & Hassed, Trust, attitudes and use of artificial intelligence: A global study 2025. University of Melbourne and KPMG, 2025.
  6. Bonney, Breaux, Dinlersoz, Foster, Haltiwanger & Pande, U.S. Census Bureau: BTOS AI supplement survey, CES Working Paper 26-25, 4/2026.
  7. McKinsey Global Institute, Internet matters: The Net’s sweeping impact on growth, jobs, and prosperity, 5/2011.
  8. Comin & Hobijn, An Exploration of Technology Diffusion. American Economic Review 100(5), 2010.
  9. Sensor Tower via Reuters, 2.6.2026: ChatGPT the fastest application to reach a billion mobile users, ~3.5 years from launch (Maps, YouTube and TikTok 5–8 years).
  10. ING, Q2/2026: AI investment accounted for 36–50% of quarterly US GDP growth depending on the method of calculation.
  11. BIS (Bank for International Settlements), Annual Economic Report 2026, Chapter I Progress and peril, 28.6.2026, Box A.
  12. NBER Working Paper 34836, 2–3/2026.
  13. McKinsey, The economic potential of generative AI, 6/2023.
  14. Goldman Sachs, Global Economics Analyst, 26.3.2023 (Hatzius, Briggs, Kodnani, Pierdomenico).
  15. Filippucci, Gal, Laengle & Schief (OECD Economics Department), Macroeconomic productivity gains from Artificial Intelligence in G7 economies. OECD Artificial Intelligence Papers No. 41, 6/2025.
  16. International Monetary Fund (IMF), World Economic Outlook Update: Global Economy — Steady amid Divergent Forces, 19.1.2026.
  17. Goldfarb, Kirsch & Miller, Was there too little entry during the Dot Com Era?, Journal of Financial Economics 2007 (working paper version 10.4.2006).
  18. Pets.com (11/2000), eToys, Amazon’s −90% share price collapse and Google’s IPO in 2004.
  19. The Batch (DeepLearning.AI), Claude Cowork plugins trigger a SaaS stock selloff, 27.2.2026.
  20. CNBC, Software stocks rebound as Anthropic announces new partnerships, 24.2.2026.
  21. CNBC, Software stocks wrap up best month since 2001 as talk of “SaaSpocalypse” subsides, 29.5.2026.
  22. Smith & Linden, Two Decades of Recommender Systems at Amazon.com, IEEE Internet Computing 2017.
  23. Linden, Smith & York, Amazon.com Recommendations: Item-to-Item Collaborative Filtering, IEEE Internet Computing 2003. DOI 10.1109/MIC.2003.1167344.
  24. RankBrain 10/2015: Search Engine Land and Google’s own statements.
  25. Zielnicki, Aridor, Bibaut, Tran, Chou & Kallus, The value of personalized recommendations: Evidence from Netflix. International Journal of Industrial Organization 108, article 103304, online 6/2026, print issue 10/2026. DOI 10.1016/j.ijindorg.2026.103304 · ISSN 0167-7187 · Elsevier. Open-access version arXiv:2511.07280 (v1 10.11.2025 → v5 8.6.2026, CC BY 4.0), also CESifo Working Paper 12257 and SSRN 5741805.
  26. The Ministry of Finance, press release Kunnat ja hyvinvointialueet kokoontuivat pyöreiden pöytien keskusteluihin: tekoälymuutos vaatii johtajuutta ja yhteistyötä, 27.5.2026.
  27. Ruokonen, Mika, Tehtävänä tekoälymurros. Sitra Study 254, 15.12.2025.
  28. Epoch AI, the capability gap between open and closed frontier models. Updated 29.5.2026.
  29. McKinsey (Hämäläinen, Patel, Blumberg, Catlin & Lala), Is that AI agent worth it? Agentic economics and the modern operating model, McKinsey Quarterly / QuantumBlack, 13.7.2026.
  30. McKinsey Technology (Tavakoli, Goodman, Rowshankish, Kumar, Soller, Barreto, Parekh & Litta Modignani), Building the foundations for agentic AI at scale, 2.4.2026.
  31. McKinsey, Seizing the agentic AI advantage (13.6.2025) and When can AI make good decisions? The rise of AI corporate citizens (4.6.2025).
  32. Anchor dates of the API economy: the Salesforce API in February 2000, the eBay API 20.11.2000, the Amazon Web Services API 15.7.2002.
  33. MCP timeline: launch 11/2024 (Anthropic), OpenAI 3/2025, Google DeepMind 4/2025, handover to the Agentic AI Foundation (Linux Foundation) 12/2025. Download figures: Anthropic, Donating the Model Context Protocol and establishing of the Agentic AI Foundation, 9.12.2025.
  34. Forrester, Kate Leggett et al., SaaS as we know it is dead: how to survive the SaaSpocalypse, 11.2.2026. Global SaaS spending 318 billion dollars (2025) → 512 billion dollars (2028) → 576 billion dollars (2029).
  35. Gartner, press release Gartner Says $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI, 1.7.2026.
  36. Bernhofen, El-Sahli & Kneller, Estimating the effects of the container revolution on world trade. Journal of International Economics 98 (2016).
  37. Sangeet Paul Choudary, Reshuffle: Who wins when AI restacks the knowledge economy, 2025.
  38. Challagalla, Khan & Beaulieu, Stop Running So Many AI Pilots. Harvard Business Review 11–12/2025.
  39. Humlum & Vestergaard, Still Waters, Rapid Currents (formerly Large Language Models, Small Labor Market Effects). NBER Working Paper 33777, 5/2025, revision 3/2026.
  40. Suh & Oh, Generative AI and the Reallocation of Time: Productivity, Leisure, and Fulfilling Work. arXiv:2602.12695, 13.2.2026.
  41. Brynjolfsson, Li & Raymond, Generative AI at Work. Quarterly Journal of Economics 140(2), 2025.
  42. Dell’Acqua, McFowland III, Mollick et al., Navigating the Jagged Technological Frontier. Organization Science 37(2), 2026. Pre-registered field experiment, 758 consultants.
  43. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10.7.2025. arXiv:2507.09089. 16 experienced developers, 246 real tasks in their own repositories.
  44. Cui, Demirer, Jaffe, Musolff, Peng & Salz, The Effects of Generative AI on High-Skilled Work. Working paper 6/2025, retrospective registration AEARCTR-0014530. Three randomised trials, 4,867 developers (Microsoft, Accenture and an unnamed Fortune 100 company). Standard error 10.3 percentage points; the authors’ own caveat “each experiment is noisy”.
  45. Dell’Acqua, McFowland III, Mollick et al., The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. NBER Working Paper 33641, 2025. Field experiment at Procter & Gamble, n = 776; professionals worked on the company’s own product development tasks.
  46. Jabarian & Henkel, Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews. arXiv:2607.28222, 7/2026.
  47. Haapanen, Mikko, Hiltunen, Arto & von Weymarn, Tom, OHJ – Omistajien, hallituksen ja johdon osaaminen, yhteistyö ja vallankäyttö. Boardman, 5/2022. ISBN 978-952-60747-2-9.
  48. PwC, Two futures for jobs in an AI era: 2026 Global AI Jobs Barometer, June 2026.
  49. Parasuraman & Riley, Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors 39(2), 1997 · Parasuraman & Manzey, Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors 52(3), 2010.
  50. METR, time horizon measurement · Ord, arXiv:2505.05115 (half-life model).
  51. Stanford HAI (Institute for Human-Centered AI), Artificial Intelligence Index Report 2025, 7.4.2025.
  52. Budget allocation principle. Boston Consulting Group, Closing the AI Impact Gap, 2025.
  53. Bai, Huang, Wang, Sun, Mihalcea, Brynjolfsson, Pentland & Pei, How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks. arXiv:2604.22750, 24.4.2026 (v2 29.4.2026).
  54. Fin (formerly Intercom), Announcing Fin Apex: the age of vertical models is here, 26.3.2026; background: VentureBeat interview 26.3.2026.
  55. Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 (Digital Omnibus, OJ 24.7.2026, in force 27.7.2026); consolidated text on EUR-Lex.
  56. Korinek & Vipra, Concentrating intelligence: scaling and market structure in artificial intelligence. Economic Policy (Oxford) 40(121), 1/2025.
  57. Space Exploration Technologies Corp., Form S-1 Registration Statement. U.S. Securities and Exchange Commission, 20.5.2026, registration number 333-296070.
  58. Tesla, Q2 2026 earnings call, 22.7.2026.
  59. Björkroth, Linturi, Jussila, Höyssä & Eriksson, Societal transformations 2026–2045. Tulevaisuusvaliokunnan julkaisu 2/2026, Suomen eduskunta, 27.5.2026.
  60. Valtioneuvoston kanslia ja valtiovarainministeriö, Talouspoliittinen ministerivaliokunta linjasi julkisen hallinnon tekoälymuutoksen kiihdyttämisestä ja organisoinnista (press release 25.6.2026).
  61. Elinkeinoelämän keskusliitto EK, EK:n kysely: Vain joka kymmenes yritys synnyttää uutta liiketoimintaa tekoälyllä (press release 8.9.2026).
  62. AI Finland ja Business Finland, Näin Suomi hyödyntää tekoälyä 2026 (2026; the review was compiled by Demos Helsinki on their behalf): PDF published on aifinland.fi in April 2026.
Sofokus founder

About the author

Founder & Chairman

Teemu Malinen

Teemu writes about digital business trends, modern company culture and startup investments.

Read more articles (157)