Why our predictions about work and AI fail

Economist John Maynard Keynes estimated in 1930 that within a few generations we would be working 15 hours a week. Technology was supposed to increase productivity so much that we could get by with far less work. Ninety years later, the reality is different: the average work week has barely changed since the 1940s and remains at roughly 40 hours.

According to Peter Evans-Greenwood, Peter Williams, and Kellie Nuttall from the Deloitte Center for the Edge, the problem is that we are using the wrong mental models when trying to predict how technology will change work. Every model — that is, the way we frame the present — determines which future paths we can even see, and which remain invisible.

The classic "product-process-task-skill" model assumes that automation gradually displaces lower skills and pushes workers toward higher ones. But reality is far more complex. Deloitte points to the example of radiologists: just a few years ago, it was widely predicted that AI would completely replace them. Instead, the field is thriving because the role of doctors has transformed — AI helps them locate suspicious findings, but the final decision remains with the human.

From task automation to behavior automation

The key shift in thinking that Deloitte proposes is that we should stop viewing AI as a tool for task automation and start understanding it as a tool for behavior automation. Behavior is a response to a change in the environment — the same stimulus in different contexts triggers different reactions.

Large language models (LLMs) are a prime example. When you prompt a model to "remember" something, the memory is actually created in the moment of interaction between the prompt and billions of neural network weights. It's not retrieval from memory, but the creation of a new response — much like when a scent triggers a long-forgotten memory.

If we automate behavior, Deloitte says we must track two dimensions: agency — the degree of freedom with which AI can act — and authority — who has the final say. It is precisely the intersection of these two axes that creates a spectrum of possible collaborations between humans and machines.

Nine collaboration models already working today

Deloitte has identified nine specific collaboration patterns that are already emerging across industries. For wealth management and investment management, the following are particularly relevant:

1. The Prioritizer

AI analyzes a list of investment opportunities, client requests, or business leads and ranks them by importance or potential value. The portfolio manager then processes items in the order suggested by AI — sometimes with recommendations on how to approach them. This model is gaining traction at large asset managers where data volume far exceeds human capacity.

2. First Pass

The machine performs an initial analysis — processes a life insurance application, evaluates a client's investment risk, or categorizes a portfolio under MiFID II regulations. A human expert then reviews and confirms or adjusts the output. This model is closest to what we already see today in private banking, where AI assists in preparing investment proposals.

3. The Collaborative Decision-Maker

Complex decisions — for example, changing portfolio allocation in response to a macroeconomic shock — are conducted through a dialogue between AI and the portfolio manager. AI lists available options, helps weigh them objectively, and suggests the highest probability of success. The final decision, however, remains with the human.

4. The Personal Coach

AI discovers the strengths and weaknesses of an investment advisor — for instance, during a phone or video call with a client. It then continuously coaches the advisor and improves their performance. This model could dramatically raise the quality of client service, which is a key differentiator in today's wealth management.

5. The Doppelgänger

The machine learns from an experienced portfolio manager or private banker and replicates their decision-making patterns. The experience of top experts can thus be digitally scaled across the entire organization. This is one of the most ambitious models — and also the one that raises the greatest concerns about replacing human work.

6. The Triage Nurse

AI assesses a client request and decides whether it requires a human consultation. If it's a routine query — for example, about the current portfolio balance or index performance — it handles it on its own. This model is already used today by robo-advisory platforms as a first-level response to client inquiries.

The remaining three models — the supervisor, the muse, and the subordinate — complete the spectrum from fully autonomous AI systems, through creative collaboration, to routine support tasks.

What this means for wealth management

Wealth management is an exceptionally interesting field for agentic AI. Unlike retail banking, it's not just about speed and volume — trust, personalization, and the ability to understand clients' complex life situations play a key role. That's precisely why Deloitte emphasizes that the future will not be determined by technology alone, but by our decisions about how we deploy it.

Firms that automate behavior without thought — for example, deploying AI as a "supervisor" that assigns tasks and evaluates the performance of private bankers — risk losing the trust of both employees and clients. In contrast, the "collaborative decision-maker" or "first pass" models preserve the human element and expertise while leveraging AI's power for data analysis and recommendation generation.

European and Czech context

In the European Union, the AI Act adds another layer of complexity, introducing strict rules for high-risk AI systems from February 2025. Wealth management — including creditworthiness assessment, investment advisory, or client scoring — may fall under this regulation. This means mandatory human oversight, algorithm transparency, and documentation of decision-making processes.

For Czech financial institutions, from the private banking divisions of large banks to independent investment advisors, this is a dual challenge: deploy agentic AI in a way that boosts productivity while simultaneously meeting regulatory requirements. Given that Czech clients traditionally prefer personal contact with their banker or advisor, the most suitable path appears to be models where AI assists but does not decide on its own.

Don't bet on a single card

Deloitte warns against trying to bet everything on a single prediction. History shows that even the biggest technological breakthroughs did not lead to predictable outcomes. Factory electrification, for example, initially only brought 20–60% fuel cost savings. The real productivity increase of 20–30% came 30 years later, when engineers realized that electrical power allowed for a complete reorganization of production lines — with the same number of machines, workers, and floor space.

The same applies to agentic AI. Initial deployments will bring incremental savings, but the true leap in wealth management productivity will come only when firms rethink entire work processes around the new capabilities that AI brings. And that requires experimentation, small bets on different collaboration models, and a willingness to learn along the way.

The future of wealth management with agentic AI is not predicted — it is created. And by every decision that banks, investment firms, and regulators make today.