The AI Tsunami

The risk is not the wave itself. The risk is treating the wave like weather — temporary, observable, something you batten down for and move on from. AI is not weather. It is the new coastline.

Most leadership teams I sit with are still asking the wrong opening question. They ask which tool to buy. They want a recommendation, a vendor, a benchmark, a slide. The opening question is harder and more consequential: which decisions in this company are about to become machine-mediated, which ones should never be, and what authority structure are we prepared to defend when both kinds start to look the same on the dashboard.

Software was a tool you bought to do a task faster. AI is becoming the layer the task happens inside. That is a different kind of dependency, and it requires a different kind of governance. The teams that confuse the two will inherit invisible decisions they did not architect, cannot audit, and will eventually have to defend in front of regulators, customers, journalists, and a board that is now realizing the model output IS the policy.

The companies that survive this decade will be the ones who decided, in advance, which questions a model is allowed to answer on their behalf — and which questions a human still has to look at and own.

What this letter argues

Three things. First, the AI tsunami is not a product cycle; it is a substrate shift, closer to electricity or the cloud than to a CRM upgrade. Second, the most expensive mistake leaders make right now is delegating model strategy to whichever vendor showed up most recently with the best slide deck. Third, there is a defensible playbook for moving with intelligence — choosing models, scoping agents, preserving human authority, and refusing to inherit the moral failures of the social-feed era.

I will publish the playbook in pieces in this letter series. This first letter is the framing: what the water is actually doing, why the standard corporate response will not be enough, and the five things leaders need to internalize before the wave hits the part of the business they thought was safe.

1. The wave is a substrate, not a product

We have been here before. Electricity did not change the factory by replacing the steam engine. It changed the factory by allowing the factory to be re-laid out — looser, taller, distributed, modular. The first generation kept the old footprint and just swapped the power source. The second generation tore up the floor and rebuilt the machine around what the new substrate made possible. That is the difference between adoption and adaptation.

AI right now is in the adoption phase for most companies. A chatbot gets glued onto a help center. A copilot gets stapled to the IDE. A summary widget gets bolted onto the CRM. The org chart, the workflows, the decision rights, the metrics — all unchanged. The next phase looks different. Whole functions get re-laid out around the new substrate. Agents replace ticket queues. Memory replaces meetings. Routing replaces management of routine work. The roles change. The headcount math changes. The advantage compounds for whoever moves second-fast, not first-loud.

The first generation kept the old footprint and swapped the power source. The second generation tore up the floor.

2. Model strategy is not vendor strategy

Vendors are not the problem. Vendors are doing their job. The problem is that “which AI vendor do we use” has become a proxy for “what is our AI strategy,” and the two are not the same question. Buying GPT does not give you an AI strategy. Buying Claude does not give you an AI strategy. Buying twelve overlapping enterprise copilots from three different platforms gives you neither a strategy nor a budget you can defend.

Model strategy is a thesis about which decisions in the business deserve machine help, which deserve human judgment, where the two intersect, and what your minimum acceptable failure looks like in each case. From that thesis, the model choice and the agent design follow. Without the thesis, every vendor demo looks compelling and every renewal cycle is a small heart attack.

3. The hallucination problem is a governance problem

Models will hallucinate. That fact is not going to change at the timescale you need to plan around. What changes is whether your organization has set up the conditions under which a confident wrong answer can become a real business decision. The hallucination problem looks technical and is actually structural: which steps require evidence, which require a human signoff, which require an audit trail, and which can run on autopilot because the downside is recoverable. Companies that conflate those four categories will lose money on at least two of them.

The hallucination problem looks technical and is actually structural. Which steps require evidence. Which require a signed human. Which require an audit trail. Which can run on autopilot.

4. Agents change the shape of work, not just the cost of it

The early conversation about agents was about cost — replace a task, save a salary. The more interesting conversation is about shape. An organization that delegates routine decisions to a fleet of agents runs at a different cadence than one that does not. Cycle times compress. Hand-offs flatten. Reviews concentrate at the consequential junctions instead of being smeared across every step. The teams that succeed will not be the ones with the most agents. They will be the ones who knew which ten decisions per week needed to keep traveling through the human spine.

5. The next layer cannot inherit the last failure

The social-feed era trained an entire generation to accept tracking, manipulation, behavioral prediction, dark-pattern dependency design, and surveillance as a price of being online. The AI era cannot inherit that pattern. Not because it is impolite, but because the next layer is more powerful and the abuses will be less reversible. If we re-create the engagement-metric, attention-extraction, consent-as-UX- trick playbook on top of LLMs and agents, we will hand future regulators a much heavier hammer than the one currently pointed at social platforms — and we will deserve it.

That is why this letter — and every letter that follows — comes from an explicit posture: privacy by default, human final authority, transparent agent action, no surveillance as a business model. Not as marketing. As architecture.

What I recommend leaders do this quarter

First, write down the ten decisions in your business that you would be embarrassed to discover were being made, even partially, by a model without a named owner. Write down who that named owner is. If you cannot name them, the model is not the problem; the org chart is.

Second, pick one workflow that is genuinely expensive in human time and routine in cognitive shape, and put an agent on it — with a real evaluation loop, a real failure budget, and a real off-switch. Not a pilot. A production system small enough to govern.

Third, identify the one person on the executive team who is going to own model strategy as a discipline, the way someone owns finance, legal, or security. If that person is “everyone,” it is nobody, and you are going to find that out at a quarterly review.

If you cannot name the owner of an AI-mediated decision, the model is not the problem. The org chart is.

The water is doing what water does

I do not think AI is going to make leaders obsolete. I think it is going to make unserious leaders obvious very quickly. The leaders who learn to read the substrate — who treat AI as the operating layer it is becoming, not as a vendor category to procure — will compound an advantage that is hard to describe and impossible to fake.

The water is doing what water does. The coastline is moving. The teams that already started building higher are going to look, retrospectively, like they were lucky. They were not lucky. They read the water.