If you work in product development, as an engineer, designer, product manager, or researcher, you're watching AI reshape your profession in real time. The question isn't whether AI will change your work. It's when, how much, and what you should do about it.

The temptation is to assume AI will plateau. The current limitations will persist. The jobs being automated today are fundamentally different from yours.

That assumption is risky.

This article maps out four levels of work and estimates when AI is likely to reach each one. The timelines are opinionated estimates, not forecasts, because they assume breakthroughs that don't yet exist, and solving one problem often reveals the next. Treat them as a framework for thinking about sequence and relative difficulty, not as dates to mark on your calendar.

Early adopters typically lead mainstream adoption by years, so if you're at a company willing to experiment, you'll encounter these shifts sooner.

The Four Levels of Thinking

I’ve previously written about how job titles can be misleading and how using the four levels of thinking is a better reflection of job level. I’ll use these levels to track where AI is and how it will evolve in the future. To recap, the four levels are. 

Level 1: "Is it done?" 
You're given specific tasks. Your value is speed and reliability.

Level 2: "Is it solved the right way?"
You're given problems. Your value is your ability to choose the best methods and tradeoffs.

Level 3: "Is this the right problem?"
You're given outcomes. Your value is thinking in cross-functional systems to identify problems worth solving.

Level 4: "Is this the right direction?"
You set the frame itself. Your value is defining which outcomes to pursue under uncertainty.

Level 1: Task Execution

Timeline for AI replacement: 2027–2030 (mainstream).

Task execution requires consistency and accuracy.

AI can already design UIs, write code, consolidate data, and create reports. But not reliably. Output is inconsistent across disciplines and even across runs. Engineers get solutions that violate architecture. Designers get components that clash with design systems. Researchers get summaries that miss nuances.

The core problem: AI doesn't know your system, whether it is your codebase, brand guidelines, or user context, so it makes wrong assumptions. This stems from two limitations: the probabilistic nature of LLMs and their inability to learn.

The consistency problem
Run the same prompt twice, get different results. The emerging solution mirrors how we manage junior employees: create a plan first and then validate their performance against the plan. Most LLMs now have planning modes, and some tools implement "swarms" which are multiple LLMs with different roles (planner, architect, developer, tester) that validate each other's work. This improves reliability but consumes an enormous amount of tokens and still requires human oversight.

The learning problem
Swarms are a workaround for LLMs' inability to learn. We run multiple passes because the model forgets to follow patterns. But this is a difficult problem to solve because efforts often run into "catastrophic forgetting" - when LLMs learn new knowledge, they lose other information. True long-term learning remains elusive.

Breakthroughs Needed

• Continuous learning: Maintaining understanding across sessions

• Cost reduction: Making agent swarms economically viable

Confidence

Cost reduction is a safe bet because token prices dropped from $20/million (late 2022) to $0.40/million (August 2025), and there is no sign of that stopping soon. There is an incredible amount of research effort focused on the Continuous learning space as well, with a focus on three distinct areas: modular architectures, hierarchical memory systems (like short-term and long-term memory), and compositional approaches. 

I predict that frontier models will start releasing learning features in late 2026, with quality improving on every release. My optimistic view is that through 2027, people will continue delegating even more, and each model release will improve capabilities. By 2028, AI will be able to build complex features in existing complex applications, with minimal to zero human intervention. The remaining challenge is adoption.

Level 2: Solution Design

Timeline for AI replacement: 2028–2031 (mainstream).

Solution design requires understanding system requirements and constraints of different approaches.

LLMs are proficient sounding boards today. Engineers evaluate architectures. Designers explore directions. Researchers test concepts with synthetic users. But every conversation starts with exhausting context-setting where you need to reiterate your system, constraints, design language, and user learnings. You often spend more time explaining than exploring.

However, Level 2 work needs the same breakthroughs as Level 1. If LLMs can learn your application's context, they can produce quality plans without endless iteration. Once continuous learning works for task execution, extending it to solution design is primarily engineering, not research.

While level 1 tasks will be easier to offload, solution design will take a bit longer. Technical debt can grow quickl,y so there will be a higher quality bar required to offload the product designs and architectural decisions. Early adopters will have more appetite for risk, and the learnings from these iterations will improve quality across the industry. 

Level 3: Problem Framing

Timeline for AI replacement: 2030+ for AI-native companies. Unlikely for established organisations.

Problem framing means deciding on the right problem to solve. To reduce churn, do you focus on onboarding, the application UX, your support services, or something else?

The breakthrough needed isn't technical - it's organisational.

AI can help explore which customers to prioritise or what's causing churn. But it requires enormous context and understanding of the business landscape, not just the product code and design. It requires full systems thinking.

The fundamental barrier for AI delivering on this work is that this context doesn't exist in any system. 42% of institutional knowledge resides solely with individual employees. Organisations record what happens but rarely capture why decisions are made. Strategy, competitive dynamics, relationship history, and other tacit knowledge are all sparsely documented and rarely current.

Reverse-engineering this into existing organisations is extremely difficult due to the resistance to change and organisational inertia. New companies have an edge here because they can design processes from scratch to capture this context.

Breakthroughs Needed

• Decision capture systems: Recording why, not just what

• Organisational change: People actually documenting reasoning

• Infrastructure for "decision traces"

Confidence

For established companies, I see Level 3 replacement as unlikely because the behavioural change required is too large. But this isn't existential; replacing Levels 1 and 2 will make these established companies more profitable. 

New AI-native companies building with these principles will automate far more, and single-person unicorn companies will likely emerge in the early 2030s.

Level 4: Outcome Definition

Timeline: Indefinite human-in-the-loop.

Level 4 is where you set the direction for a company. Should we target a particular customer segment? A particular geography? Should we focus on cost-cutting or sales growth?

AI has access to huge amounts of information and can make strong recommendations. But a system cannot be accountable. There will always need to be a human in the loop to make the final decision.

Adoption Barriers and Strategic Risks

These timelines are more focused on the technical ability and do not factor in some very real blockers for organisational adoption. The key adoption challenges include:

Data privacy and platform dependency
Using AI for Level 1–3 work means feeding it your architecture, strategic reasoning, and competitive insights. That data flows to third-party systems, often in foreign jurisdictions. Access can be revoked, pricing can change dramatically, and features can be deprecated. Building core processes on capabilities you don't control creates vulnerability.

Geopolitical risk
If your AI provider is subject to foreign regulations or export controls, access isn't guaranteed. We've already seen technology access used as leverage in international disputes.

There is already a push towards more on-site solutions for AI to mitigate these risks, and I see that trend accelerating over time.

What This Means for Your Career

Honestly? I don't know.

The pessimistic view is that there will be mass layoffs as companies replace Level 1 and 2 workers with AI. Organisations operating with a fraction of the current headcount.

The optimistic view is that it has never been easier to start and run your own company. The barriers that required teams of specialists are dissolving.

Both are probably true, in different places, for different people.

What's certain is that you have time, whether five years or ten, to get ahead of these changes. Move from task doer to solution designer. Then move again to a systems thinker. Understand where your product sits in the business ecosystem. Learn why the strategy is what it is, not just what it is.

The companies that win won't just have better models because we’ll all have similar access to the models. They'll have better context around the decision traces and reasoning chains that most organisations never capture.

The people who win will be comfortable in ambiguity and relentless about their own growth.

The direction is clear. Plan accordingly.

This is the seventh and final article in a series exploring how AI is reshaping career progression. If you want to go deeper, check out our free ebook: Managing your Career in the Age of AI.