AI is changing the world, just not that fast

Be careful with bold statements

There's the real world, people, processes, regulation, liability, culture. And then there's the world many companies live in when they're fundraising or publicly traded: a world where a single sentence can move valuation, headlines, and stock price.

In that second world, speed is an asset. Hype is a strategy. And incentives matter.

If your product gets more valuable when the market believes a quake is imminent, why would you go out of your way to say, "This will take years, and it will roll out slowly"? You wouldn't. You'd say the loud part. You'd turn "possible" into "inevitable." You'd turn "some tasks" into "entire professions."

That's how you get claims like: "Lawyers are gone in six months." "Accounting is finished." "Support teams won't exist." "Developers are obsolete."

AI is changing the world. But not at the level, and not at the speed, that those slogans suggest.

And the reason is simple: most of those predictions forget the key variable.

The human factor.

The difference between can and will

A model can draft a contract. That's not the question.

The question is whether a company will ship that contract into production workflows, have someone sign it, accept the legal risk, update internal processes, train the team, pass audits, handle privacy/compliance, and survive the first incident without reputational damage.

In high-stakes domains, responsibility doesn't disappear because you used a tool.

And errors happen. Courts have already dealt with filings that included AI generated fake citations, which triggered new limits and stronger warnings to lawyers: verify your sources, own your submissions.

So AI isn't competing against "what a lawyer writes." It's competing against an entire accountability system: ethics, liability, insurance, precedent, professional discipline, and the cost of being wrong.

In tech right now, building has become increasingly permissionless.

A small team can ship a real product with multiple agents in Claude Code, or generate and iterate flows with tools like v0. In a free building environment, you can prototype, test, and deploy at a speed that would have sounded unrealistic not long ago. For early stage teams, the cost of failure is low, feedback loops are tight, and AI feels like a multiplier.

But the companies that move the most money do not live in that environment.

Large, historic enterprises live in a permissioned reality. Their constraints are not theoretical. They are compliance, governance, procurement, auditability, data residency, vendor risk management, and internal controls shaped over decades. Many cannot adopt modern AI tooling easily, not because the technology is weak, but because the institution is built to avoid avoidable risk.

And even when compliance is not the main blocker, politics often is.

In many organizations, automating a process does not just improve efficiency. It changes ownership. It removes influence. It breaks routines that became stable over decades. It can invalidate the informal agreements and internal relationships that keep departments functioning. A technology that removes an internal political business from someone who has been in the company for 30 years will not be adopted simply because it is better. It will be resisted, slowly and strategically, unless there is a cultural shift strong enough to override the immune system of the organization.

I have always been drawn to the frontier of innovation. Using the best tools, the newest patterns, the fastest ways to build. Over the last years, I have worked with many startups and with enterprise organizations: insurance, banking, energy, telecom. Some of them have existed since the early 1900s. The contrast is sharp.

Startups optimize for speed. Enterprises optimize for continuity.

I remember proposing a DevSecOps pipeline for a company running on premise servers, with years of manual operations as the default. After long discussions, the final verdict from the lead architect was that it would take around 18 months just to implement it for two applications.

From a purely technical standpoint, it could have been done in a fraction of that time.

But the real timeline was approvals, validation gates, internal agreements, budget cycles, shifting responsibilities, and the politics of change. The bottleneck was not execution. The bottleneck was human.

Some companies cannot use AI for coding today because compliance forbids sending proprietary repository context to external tools. If you cannot share code, you cannot get high quality context, and the assistant becomes far less useful. That is why many teams are exploring workarounds like local AI models, on premises deployments, or internal systems that generate more abstract representations of the codebase, such as summaries, dependency graphs, interfaces, and sanitized snippets, and only then pass limited context forward. It is not a capability problem. It is a governance problem.

So when someone says, "AI will replace this profession in six months," I do not ask whether it is possible.

I ask what kind of organization they mean, under what constraints, with what governance, and who loses power when it ships.

Because in the real world, the bottleneck is rarely the model.

It is the institution.

Reality looks like task change, not job deletion

If you want a grounded view, look at credible data.

• The IMF reiterated in January 2026 that nearly 40% of global jobs are exposed to AI driven change. Exposure is not replacement. It is mostly task reshaping plus pressure to reskill.
• Adoption is rising, but not instant. The OECD reported that in 2025, 20.2% of firms used AI, up from 14.2% in 2024 and 8.7% in 2023. That is strong growth, but it also shows we are still early in diffusion.
• McKinsey's 2025 global survey shows the same pattern inside big orgs: high curiosity, uneven scaling. They report 23% say they are scaling agentic AI somewhere in the enterprise, and 39% say they have begun experimenting. They also note most organizations are still not scaling AI across the enterprise.
• Productivity is real, but it is jagged. A Microsoft Research backed set of randomized trials in software development, published June 2025, evaluates effects in real workplaces and frames the impact as context dependent rather than automatic.
• Professional services show both adoption and inertia. Thomson Reuters' 2026 AI in Professional Services report says organization wide usage of AI rose to 40% in 2026 from 22% in 2025, and that many professionals use public tools for work while enterprise tooling adoption is also significant.

This pattern repeats across industries: AI shows up first as a tool individuals use, then as features, then as workflows, and only later, after friction, governance, failures, and learning, as real organizational redesign.

Founders, incentives, and narrative gravity

You need to read incentives.

Some leaders push the "fast, everything changes now" framing, partly because it might be true in pockets, but also because hype moves markets and capital.

Other leaders bring a more grounded constraint: show me the real world macro impact.

Satya Nadella has a brutally anti-hype yardstick: if AI is truly "like the Industrial Revolution," we should see Industrial Revolution like growth. Otherwise, it's a lot of benchmarks and capital expenditure without broad productivity showing up where it matters.

Andrej Karpathy put it in a way that matches this thesis: not "the year of agents", more like the decade of agents. Translation: this is real, but it's not instant.

Meanwhile, researchers like Arvind Narayanan and Sayash Kapoor warn about the hype cycle and "AI snake oil": the gap between what gets promised and what can be safely deployed in the real world.

The missing variable is the human factor

When someone says, "Accountants are finished," they're collapsing a profession into one function: handling numbers.

But a real accountant doesn't "handle numbers." They interpret them, apply standards, negotiate judgment calls, defend decisions to auditors, manage risk, and ultimately sign their name under responsibility.

AI can accelerate parts of that. It can draft, summarize, reconcile, spot anomalies, generate explanations.

But as long as there's regulation, auditability, and consequences, there's a human layer that doesn't go away. It changes shape.

The bottleneck isn't capability. It's trust.

Trust is earned through:

• predictable performance in messy reality
• explainability where needed
• governance
• quality controls
• a clear line of accountability

That's why "AI replacing entire professions in months" is usually not a prediction, it's a marketing line.

On Lenny's Podcast, Marc Andreessen argues that "the atomic unit of what happens in the workplace is the task," not the job. A job is just a bundle of tasks, so new technology usually reshuffles and replaces tasks long before it erases whole roles. He illustrates it with a simple secretary example: decades ago executives dictated and secretaries typed; today executives write their own emails, while administrative roles evolve toward coordination and planning rather than typing.

This is without even factoring in government policy. New regulations, liability frameworks, data sovereignty rules, and labor protections can all slow down adoption and the pace at which tasks get replaced, even when the technology is ready.

The most valuable skill right now is judgment

If you run a company or build a product and you're trying to adapt to AI shifts, there's one meta-skill you need more than anything:

Judgment.

Judgment is being able to separate:

• a task from a role
• a demo from a production system
• technical capability from human adoption
• short-term tooling from long-term organizational change

A practical way to think about it:

1. Identify the highest-friction tasks in your value chain.
2. Ask where AI can reduce cost or cycle time without increasing risk.
3. Add guardrails: review, logging, evals, audit trails.
4. Ship small. Measure. Iterate.
5. Expand only when you can defend reliability and responsibility.

The question is not "Will AI replace my industry?"

The question is: Where can AI give me a real advantage today without breaking trust, quality, and accountability?

Most importantly, focus on adoption and on learning to tell what is real from what is noise.

We are living in a period of constant alarmism and sweeping statements. It is normal to feel overwhelmed. You try to establish a process and something new drops. You start learning a tool and suddenly you are "behind" again.

So move toward adoption, but do not panic.

AI is changing the world.

Just not that fast.

Stay curious.