My AI Predictions 7 years ago, how did they pan out?

In 2019 I was a guest lecturer at the Nyenrode AI masterclass, where I presented “AI in practice”. At the time AI was still relatively obscure for most companies. The AI that was practical to use then came down to three things: classifiers, search, and natural language processing such as sentiment analysis. Some of the writing was on the wall for early adopters, but generative AI was not yet something most people had heard of. That changed when OpenAI released ChatGPT in November 2022.

I ended the talk with a slide on how AI would impact your company. Below I go through the seven points on that slide and look at how they have held up, based on public developments since then.

1. A boat load of resistance, both vocal and non-vocal

The vocal part was easy to see. AI was one of the central issues in the 2023 Writers Guild of America strike, and the agreement that ended it included specific rules on how studios may and may not use AI in writing. Similar discussions have taken place in many other professions.

The non-vocal part turned out to be just as real, though in a different form than I expected. I mainly thought of quiet non-adoption: people not using the tools or working around them. That exists, but there is also the opposite. In Microsoft’s 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work, and 52% said they were reluctant to admit using AI for their most important tasks. People use AI, but outside official channels and without talking about it. That is also a form of resistance, aimed at the organisation’s rules rather than at the technology.

Verdict: held up.

2. KPIs will break

When AI takes over part of a process, the numbers used to manage that process stop meaning what they used to. A clear public example is customer service. In early 2024 Klarna reported that its AI assistant handled two-thirds of its customer service chats in its first month, with resolution time going from 11 minutes to under 2. Once the routine questions go to the AI, the human agents are left with the harder cases, so metrics such as handling time or tickets per agent for the human team are no longer comparable with what they were before.

As a general observation, many organisations measure AI-assisted work with KPIs that were designed for the old way of working. The KPIs do not visibly break; they quietly become misleading. [Remco: add your own example here?]

Verdict: held up, although it mostly shows up as a slow drift in meaning rather than an obvious break.

3. Move people to creative jobs

This is the point I would phrase differently today. In 2019 the general assumption, which I shared, was that AI would take over routine work and that creative work would be the safe place to move people to. Generative AI turned that around to some extent: writing text, creating images and producing code are exactly the tasks it became good at. That is also why the writers’ strike happened.

What does seem to hold is that people move towards work that requires judgment: handling the complex and sensitive cases, checking and correcting AI output, and deciding what should be done in the first place. That is not the same as “creative jobs”, but it is in the same direction.

Verdict: partly.

4. Many jobs add very little; they just shuffle knowledge around

This held up more strongly than I expected, because moving knowledge around is precisely what large language models do well: summarising, drafting, searching, translating and routing information. A study of customer support agents by Brynjolfsson, Li and Raymond (Generative AI at Work, NBER) found that an AI assistant increased the number of issues resolved per hour by 14% on average, and by around 34% for novice and less-skilled workers. The assistant essentially passed on the knowledge of the best agents to the rest.

I would word the point more carefully now. Shuffling knowledge is not worthless; organisations depend on it. But if a role consists mostly of passing information from one place to another, it is the role most directly affected by current AI.

Verdict: held up.

5. Customers care about speed and convenience more than you think

The Klarna numbers above also support this point: in the same announcement the company reported customer satisfaction on par with human agents, while answers came much faster. For routine questions, many customers prefer an immediate correct answer over waiting for a person.

There is an important limit, though. In May 2025 Klarna’s CEO said that focusing too much on cost had led to lower quality, and the company started hiring human agents again so customers could always speak to a person (Fortune). Speed and convenience matter a lot, but only once the answer is good enough.

Verdict: held up, with quality as a precondition.

6. Give an opt-out to a human and things will be fine

The opt-out itself has become a common design choice, and the Klarna case shows a company going back to it after moving too far towards automation. Regulation is moving in the same direction. The EU AI Act, which entered into force on 1 August 2024, requires human oversight for AI systems classified as high-risk.

“Things will be fine” was too optimistic, however. A human option does not remove responsibility for what the AI says. In 2024 a Canadian tribunal held Air Canada liable for incorrect information its website chatbot gave a customer about bereavement fares, and rejected the argument that the chatbot was responsible for its own statements. The opt-out is necessary, but the organisation still owns the answers its AI gives.

Verdict: mostly held up.

7. Measure success and measure failures, and balance them

This still holds, and it is still not done well. In IBM’s 2025 CEO Study, CEOs reported that only 25% of their AI initiatives had delivered the expected ROI over the past few years, and only 16% had been scaled across the enterprise. The debate on how to measure the return on AI has become one of the main topics around enterprise AI.

Measuring failures is the part most often skipped. Wrong answers, escalations, rework and customers who give up are harder to count than time saved, but without them the picture is incomplete. [Remco: add your own example here?]

Verdict: held up.

What I would put on the slide today

Looking back, most of the points held up reasonably well. The biggest miss was the assumption that creative work would be the safe place to move people to. If I made the slide again today, it would look roughly like this:

• Use AI to enable things that were not possible or affordable before, not only to replace existing work. That is where I see the most value.
• Expect resistance, much of it quiet, including people using AI outside official channels. Give them sanctioned tools and clear rules.
• Redesign your KPIs before you deploy AI, and measure quality and failures, not just speed and volume.
• Keep a route to a human, and remember that you remain responsible for what your AI tells customers.
• Expect roles to shift towards judgment and verification rather than towards “creative work”.