Executives & Directors · Module 01 · Chapter 009

Generative AI Changes the Game

Generative AI did not invent AI. It changed who can use it, by making ordinary language the way in. That widens opportunity and responsibility at the same time, and it adds to the AI an organization already runs rather than replacing it.

≈ 16 min read

After this chapter you can

Explain why generative AI felt sudden although the underlying capability was largely in place, by separating the interface from the model.
Distinguish traditional, task-specific AI from general-purpose generative AI by how people reach each, without technical detail.
Recognize that easier access widens both opportunity and responsibility, including duties to staff and customers.
Keep traditional AI in the portfolio and ask which kind of AI owns the high-stakes part of each job.

On 30 November 2022, OpenAI released a free “research preview” called ChatGPT. The model inside it was not new. The company said it had been fine-tuned from a model in its GPT-3.5 series that finished training in early 2022¹. A paper describing how to train such models to follow plain instructions had been public since March of that year², and developers had been able to reach the GPT-3 family through a programming interface since June 2020³. By the standards of the field, very little had changed.

Five days later, more than a million people had used it⁴.

Both facts are true. The capability was mostly old; the reach was entirely new. What changed in November 2022 was not chiefly what the machine could do. It was who could ask it to do something, and how. Why AI, Why Now? explained the forces that made the moment possible; what follows explains why it felt so different from everything AI had done before, and what that difference asks of leaders.

The front door changed

The one idea to carry from this chapter fits in a sentence: generative AI did not invent AI; it changed who could use it. Before late 2022, many people who benefited from AI never dealt with it directly. They used an application that had a model somewhere inside. After it, anyone with access could put a request to a capable model in the same words they would use with a colleague, and get back a draft, a summary, a comparison or an explanation.

The usability researcher Jakob Nielsen put the shift in historical terms. In his account, computing has had only three ways of being told what to do. In batch processing, from about 1945, people handed over a complete set of instructions and waited. In command-based interaction, from about 1964, person and computer took turns, one command at a time; the graphical interface most of us grew up with belongs to this paradigm. With generative AI, people state the outcome they want and leave the computer to work out the steps. Nielsen called it intent-based outcome specification, and argued that it reverses the locus of control: the user says what, and the system decides how⁵.

Figure 1.9.1 Nielsen’s three paradigms of computing. In his account, generative AI is the first new one in about sixty years.

That is why the same wave felt old to technology teams and sudden to everyone else. Specialists had been working with language models for years. For most employees, customers and board members, the first time they touched AI on purpose was the day they typed a question into a box.

Traditional AI is the engine inside the machine

To see what changed, start with what did not. As AI Is Already Inside Your Organization showed, many enterprises have run AI for years without calling it that. It forecasts demand so depots can plan staff, plans the sequence of stops for each van, scores card payments for fraud and sets markdowns in retail.

These systems share a shape. Data goes into a model built for one job. The model produces a prediction, a score or a plan, and that output feeds a decision inside a business application. The people who rely on it see a dispatch queue, a case screen or a dashboard. Many never know a model is underneath, and they should not have to. A fraud model that required every teller to understand machine learning would be a worse fraud model. Traditional AI is the engine inside the machine: built for one task and reached through the application that performs it.

That design has a consequence leaders tend to miss: the screen also limits what anyone can ask. The route planner answers routing questions and nothing else, and only the planners who use it can ask. The Major Types of AI, in Module 02, sets out the families of AI in more detail.

Figure 1.9.2 The difference that matters to leaders is less what each kind can do than who can reach it, and how.

Language became the interface

Generative AI turned that arrangement inside out. A person can now say: summarize this incident report; compare these two carrier contracts; draft a reply to this customer; explain this customs rule in plain words. They do not need to know how the model works, learn a query language or find the right screen. A finance manager can ask for a first-pass variance analysis without becoming a programmer. The person is no longer only using a system that contains AI. They are working with it directly, through language. How the model produces those answers is the subject of Large Language Models and How Generative AI Works, in Module 02; for a leader, the change in who can ask is the part that reshapes the organization.

A study by economists working with OpenAI shows how far that door has opened: by July 2025 the consumer version of one assistant had about 700 million weekly users, roughly one in ten of the world’s adults⁶.

Two cautions keep the picture honest. First, the same study found that work use was more common among educated users in highly paid professional jobs⁶. The door is open to everyone; the people best placed to get value through it are still those who already know what good work looks like. Second, Nielsen warned of an articulation barrier: describing what you want clearly in prose is itself a skill, and many people find it hard⁵. Language lowered the barrier. It did not remove it.

One interface reaches many kinds of work

The second reason the change felt large is range. Traditional systems usually solve one category of problem, and each new one is a project with its own data, build and integration. A generative model reaches many kinds of work through the same box: drafting, summarizing, comparing, translating, explaining, analyzing, planning and writing code.

The usage study shows what people actually do with that range. By purpose, across all messages, about 49 percent are asking for information or advice, 40 percent are doing, asking for an output such as a draft or a plan, and 11 percent are expressing⁶. These figures describe what people ask for. They do not show whether the work got better.

Figure 1.9.3 Most use is asking and doing: decision support and first drafts, the language work around the real job.

A useful way to picture this is the smartphone. It did not invent the camera, the map, the music player or the calculator. It put them behind one screen that almost anyone could use. It also did not make its owners photographers or navigators. Generative AI does something similar for knowledge work: many capabilities behind one conversation, and none of the judgment supplied with it.

General-purpose does not mean best at everything. It means a team can attempt a first version of many tasks without commissioning a new system for each. The Speed of AI Adoption showed what that does to the cost of a first try. The question it raises for leaders is better than “which specialist tool should we buy for this?” It is: can a general capability safely handle a first version of this work, and where do we still need a specialist system?

Easier to ask is not easier to be right

Easy to use is not the same as easy to deploy safely. The same ease that lets a thousand people try something new in an afternoon lets a thousand people make a fluent mistake. Someone pastes customer addresses and contract rates into a tool outside any approved environment. A confident error is copied into a customer letter or a board paper. An answer is trusted because it reads like the work of a capable colleague, when the system had no access to the facts it seemed to state.

Figure 1.9.4 The same ease drives both branches. Leaders cannot take one without the other.

The evidence on expertise points the same way. In a study of more than 5,000 customer-support agents, an AI assistant raised issues resolved per hour by 15 percent on average. The gains went mostly to less experienced and lower-skilled agents, while the most experienced saw small gains in speed and slight declines in quality⁷. As AI Is Changing Everything described, the frontier of what these tools do well is jagged, and people who rely on them outside it can do worse than people who do not⁸. Expertise is not made redundant by an easier interface. It becomes the thing that decides whether the output is any good, and far more people now need some of it.

Few leaders would send a junior colleague’s draft to a customer unread, and a fluent machine draft deserves at least the same review. Natural language lowers the cost of asking. It does not lower the cost of being right. The full treatment of these risks sits in Accuracy, Hallucination and Reliability and Privacy and Confidential Data, in Module 06.

Traditional AI is not obsolete: run both

If “the game changed” is heard as “our models are finished”, the wrong conclusion has been drawn. Prediction and optimization systems still do some of the most valuable work in the enterprise: credit decisions, fraud detection, pricing, inventory, routing and quality inspection. A language model can describe a delivery plan fluently. Only the system that sees the live route plan can make one.

A widely cited estimate of the economic potential makes the same point in numbers. In 2023, McKinsey estimated that generative AI could add 2.6 trillion to 4.4 trillion US dollars a year across the uses it studied. It set that against 11.0 trillion to 17.7 trillion dollars a year that non-generative AI and analytics could unlock, so generative AI would add about 15 to 40 percent on top of the older kinds rather than replace them¹¹. These are estimates of potential, not measured results, and they should be read as an order of magnitude. The order of magnitude is the lesson: the newer kind is an addition to the portfolio, not a successor.

Figure 1.9.5 In a widely cited estimate, generative AI adds to the value of older AI; it does not replace it.

The fashionable claim is that generative AI is AI and everything else is old. The accurate claim is that generative AI changed who can use AI and which tasks can be done through language. The first claim produces fashion. The second produces a portfolio. In practice the most useful habit is a question asked of every initiative: which kind of AI do we mean here, and which kind should own the high-stakes part of the job?

Story: the helper that learned to improvise

In 2023, a US nonprofit learned what it means when a change to its AI reaches everyone at once. The case involves people with eating disorders, which is exactly why its lessons matter: the people a public-facing system meets are not always the people its designers pictured.

The National Eating Disorders Association, a US advocacy nonprofit, had funded researchers to build a wellness chatbot called Tessa. The version the researchers tested and studied was rule-based: it could only give a limited set of prewritten responses. By design, one of its lead researchers later told NPR, it could not go off the rails¹². According to the association, Tessa underwent years of testing before it quietly launched in February 2022¹³.

The bot was operated as a free service by a mental-health technology company. Its chief executive told NPR that in 2022 the company had made a “systems upgrade” that included an enhanced question-and-answer feature using generative AI, giving the bot the ability to create new responses. He said the change was part of the contract. The association’s chief executive said it “was never advised of these changes” and would not have approved them¹². The two accounts have not been reconciled in public.

In spring 2023 the association announced that its human helpline, which had run for more than 20 years, would close, and promoted Tessa as a prevention resource; it later said the two decisions were separate and had been conflated. In late May, a consultant in the eating-disorder field tried the bot. It told her she could lose one to two pounds a week, eat no more than 2,000 calories a day and aim for a deficit of 500 to 1,000 calories a day. She shared the exchange on social media. On 30 May 2023, less than 24 hours after receiving her screenshots, the association took Tessa down “until further notice”¹². The association said the language had not been in the original “closed” product, and that the operator had reported a sharp surge in traffic, including attempts to trick the bot¹³.

Figure 1.9.6 A post-mortem in four steps. The system that was tested was not the system that answered.

It would be easy to tell this as a story about one bad bot, or about one side of the dispute. Read as a post-mortem, it is a story about the front door. The analysis below is about the pattern, not about either organization’s code.

Figure 1.9.7 The visible failure was alarming. The causes underneath are ones every leader shares.

First, anyone can reach it. A public chatbot meets everyone: the most vulnerable person it was built to serve, the curious, and people trying to make it misbehave. The language interface opens the door to all of them, not only to the users you planned for. Security and AI Attacks, in Module 06, covers deliberate manipulation in depth.

Second, a change let it improvise. The tested bot could only say what experts had written. The upgraded bot could compose. A generative component needs to be tested against what it must never say, and tested again whenever it changes, in the way a pricing engine is tested against what it must never charge.

Third, nobody agreed who owned the change. Whatever the contract said, the person reading Tessa’s answer saw the association’s name on it. When a supplier can change what your AI says, you need to know when it does, and to approve it before the public sees it.

Fourth, plausible is not the same as right. Advice to count calories reads as ordinary to most people. For someone with an eating disorder it can feed the illness. Fluent output carries no knowledge of who is reading it, which is why the cost of being right does not fall with the cost of asking. AI in Customer Service, in Module 05, takes up how to design a public-facing assistant, including the route to a person.

None of these is a reason not to use generative AI with the public. Every one is a reason to decide, before the door opens, which kind of AI owns which answer, what the AI must never say, who approves a change, and who catches what it gets wrong.

What this means for leaders

Generative AI hands leaders a real gain and a real duty in the same package. The gain is reach: work that once needed a specialist screen, a report request or a data team can now start with a sentence, and far more people can try new ways of working. The duty is that the same reach applies to mistakes, to data, and to people outside the organization who can now talk to its AI.

Three habits make the difference. Name the kind of AI in every initiative, and keep each kind in the job it does best: prediction and optimization systems for the high-stakes numbers, generative AI for the language around them. Widen access on purpose rather than by default, pairing broad learning with clear rules on data and review. And treat fluent output as a first draft, with a named person who checks it before it leaves the team. From Copilots to AI Agents, next, asks what changes when that output is no longer a draft but an action.

Check yourself

Generative AI is the same thing as AI.
The model behind ChatGPT’s launch was a major technical leap over anything developers could already reach.
In a study of customer-support agents, the AI assistant helped the least experienced staff most.
Because generative AI is easy to ask, its answers need less review.
Generative AI is likely to replace the value of traditional AI.
A chatbot that passed its testing will keep behaving the same way after a supplier’s upgrade.

Reflection: find the specialist screen

What comes next

Generative AI made AI accessible, and accessibility raises the next question. An answer, a draft or a summary is rarely the end of the work. Tessa only talked, and its words alone could do harm; most AI at work still stops at the answer. Someone still has to update the record, book the appointment or change the plan. The next chapter, From Copilots to AI Agents, asks what happens after the answer, when AI starts to act in the systems where work actually gets done.

Laws referenced

Not legal advice. Laws change; verify before relying on this, and consult counsel for decisions.

EU AI Act · EU

Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744

Risk-based rules. Prohibited practices include social scoring, untargeted scraping of facial images, and emotion recognition in workplaces and schools (with narrow exceptions). High-risk systems (Annex III: biometrics, safety components of critical infrastructure such as energy, water and traffic, employment and worker management, credit, education, essential services, law enforcement, migration, justice) need risk management, data governance, documentation, logging, human oversight, human oversight that keeps people able to understand the system, notice automation bias (over-reliance on its output), override it or stop it (Art. 14(4)), appropriate accuracy, robustness and cybersecurity (Art. 15), automatic logging of events (Art. 12), a provider quality-management system (Art. 17) and conformity assessment. An Annex III system is not high-risk if it poses no significant risk of harm, for example a narrow procedural or preparatory task that does not replace human assessment; systems that profile people are always high-risk, and a provider relying on this exception must document it and register (Art. 6(3)). Deployers of high-risk AI must use it as instructed, assign competent human oversight, monitor its operation, keep logs for at least six months and report serious incidents (Art. 26); employers must inform workers' representatives (Art. 26(7)). Public bodies, private providers of public services, and deployers of credit-scoring or life and health insurance pricing systems must carry out a fundamental-rights impact assessment before first use (Art. 27). Providers must run post-market monitoring (Art. 72). A deployer that puts its name on a high-risk system, substantially modifies it, or changes its purpose so that it becomes high-risk takes on the provider's obligations (Art. 25(1)). A substantial modification (Art. 3(23)) of a high-risk system needs a new conformity assessment, unless the change was pre-determined and documented at the first assessment, as with planned continuous learning (Art. 43(4)). Providers of general-purpose AI models (from 2 Aug 2025) must keep technical documentation, have a policy to comply with EU copyright law including text-and-data-mining opt-outs, and publish a sufficiently detailed summary of training content (Art. 53). Research, testing and development before a system is placed on the market or put into service is outside the Act, except testing in real-world conditions (Art. 2(8)). Since the 2026 Omnibus, the Art. 4 AI-literacy duty is an obligation of effort (take measures to support literacy), not of result. Fines reach EUR 35 million or 7% of global turnover for prohibited practices.

2024-08-01 — Entered into force
2025-02-02 — Prohibited practices (Art. 5) and the AI-literacy duty (Art. 4) apply
2026-07-27 — Omnibus softens Art. 4: providers and deployers must take measures to support AI literacy; no specific level must be guaranteed
2025-08-02 — General-purpose AI model obligations apply; governance and penalties regime in place
2026-08-02 — Transparency duties (Art. 50) apply: disclose AI interaction, label synthetic and deepfake content (marking for generative systems already on the market: 2 Dec 2026)
2027-12-02 — High-risk obligations for Annex III systems (e.g. hiring, credit, education, essential services) - moved from 2 Aug 2026 by the 2026 Omnibus
2028-08-02 — High-risk obligations for AI in products regulated under Annex I

Last verified 2026-10-06 · official text

References

OpenAI. Introducing ChatGPT. OpenAI. 2022.
Long Ouyang et al. Training language models to follow instructions with human feedback. NeurIPS 2022 (arXiv:2203.02155). 2022.
Devin Coldewey. OpenAI makes an all-purpose API for its text-based AI capabilities. TechCrunch. 2020.
Sam Altman. ChatGPT launched on wednesday. today it crossed 1 million users!. X (formerly Twitter). 2022.
Jakob Nielsen. AI: First New UI Paradigm in 60 Years. Nielsen Norman Group. 2023.
Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman. How People Use ChatGPT. NBER Working Paper 34255. 2025.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond. Generative AI at Work. The Quarterly Journal of Economics 140(2). 2025.
Fabrizio Dell'Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, HBS Working Paper 24-013. Harvard Business School. 2023.
European Parliament and Council of the European Union. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. 2024.
European Union. Regulation (EU) 2026/1744 (Digital Omnibus on AI) amending Regulation (EU) 2024/1689. Official Journal of the European Union. 2026.
McKinsey Global Institute. The economic potential of generative AI: The next productivity frontier. McKinsey & Company. 2023.
Kate Wells. An eating disorders chatbot offered dieting advice, raising fears about AI in health. NPR. 2023.
Daysia Tolentino. National Eating Disorders Association pulls chatbot after users say it gave harmful dieting tips. NBC News. 2023.