About a decade ago, T-shaped skills were a familiar shorthand in Agile conversations: deep in one discipline, with enough breadth to work across team boundaries. The point was not that everyone could do every job. It was that specialists could collaborate without treating the edge of their own role as a wall.

I think AI is changing that balance. People can now use AI to get useful first-pass work from outside their deepest specialty. That can reduce how much adjacent expertise each person needs to carry personally. But the new bargain has a condition: you still need enough knowledge to tell when the answer is wrong, set the right constraints, and know when to bring in someone deeper.

AI can give us reach beyond our specialty. It cannot take responsibility for what we accept. The person using it still needs enough grounding to question the output and verify the result.

AI raises the floor, but only on some tasks

There is evidence that AI can help people do work that previously took more experience. A study of 5,179 customer-support agents found that an AI assistant increased issues resolved per hour by 14% on average, with a 34% improvement for novice and lower-skilled agents. The researchers suggest the tool helped newer workers draw on patterns used by stronger agents. That does not mean support expertise stopped mattering. It shows how AI can bring some of that experience into a workflow.

A Harvard Business School experiment with 758 consultants makes the other half of the case. On 18 tasks within GPT-4's tested capability frontier, consultants using AI completed 12.2% more tasks, finished 25.1% faster, and delivered higher-quality work. On one complex managerial task selected to be outside that frontier, AI users were 19% less likely to produce a correct solution.

That is the shape of the change: AI can extend what a person can do, but its capabilities do not form a smooth, dependable circle around every task. The boundary can be hard to see from the answer alone.

The new baseline is verification

A 2025 survey of 319 knowledge workers gathered 936 examples of AI use at work. Participants described critical thinking shifting toward information verification, integrating responses, and taking stewardship of the task. Because it is self-reported research, it does not prove that AI has permanently changed every job. It does point to a practical change in where human effort goes: less time producing the first draft, more time deciding whether the draft deserves to ship.

NIST's AI Risk Management Framework similarly emphasizes multidisciplinary perspectives, clear human-AI roles and oversight, and a critical-thinking, safety-first mindset. The framework is aimed at organizations, not a new job description for every employee. But its message fits: reliable AI work depends on people who understand the context, can surface risks, and know who owns the decision.

That is why "enough knowledge to call BS" is a useful starting point, but it cannot mean vague confidence. It means having a working mental model of the domain, knowing what a good result must satisfy, and being able to test the parts that matter.

What "enough to call BS" looks like

Imagine an engineer asking AI to write a database migration. The SQL may compile and look tidy. A reviewer still needs to ask: does it preserve existing data, can it be safely rerun, will it lock a large table during deployment, and is there a tested rollback path? The engineer does not need to know every detail of every database engine. They do need enough database knowledge to ask those questions, run the right checks, and know when a database specialist should review the change.

That is not a soft skill or a better prompt. It is judgment grounded in the work. AI can generate the migration, but it cannot decide whether the risk is acceptable for this system and this release.

Expertise still matters. Its leverage is changing.

It would be a mistake to treat AI as a universal shortcut around deep skill. In a randomized trial involving 16 experienced open-source developers and 246 tasks in mature projects they knew well, allowing early-2025 AI tools increased completion time by 19%. That result is limited to a small sample and a particular set of tools and repositories. It is still a useful reminder that generating an answer is not the same as making the work faster, safer, or easier to review.

Deep specialists remain essential for novel problems, high-consequence decisions, architecture, safety, and the failure modes that are invisible to a casual reviewer. What can change is how much of that depth every person must personally reproduce to make progress on bounded, lower-risk work. A team can let AI cover more routine execution while keeping specialists available for the questions that truly require them.

That only works if people know their limits. If you cannot explain what would make an output wrong or how you will verify it, you are not ready to approve it just because it sounds right.

The AI-era T

I do not think the T-shaped model is disappearing. I think its balance is shifting. The vertical stroke still matters, but for many roles it can be less about personally performing every specialist task and more about having a strong enough anchor to understand the work, direct AI, and recognize the boundaries. The horizontal stroke grows to include AI-assisted reach across more disciplines, plus the judgment to connect those outputs to a real outcome.

So yes, people may not need to be as deep in every adjacent area as we once expected. AI can supply part of that reach. But the bar for responsible breadth goes up, not down. We need enough knowledge to frame the problem, check the answer, and say, "This does not look right, and here is what we need to verify." The expert is still part of the team. The difference is that more people can get useful work started before they need one.

Sources