Most AI Strategies Fail at Second Order

Enterprises have poured tens of billions into generative AI. By MIT's 2025 count, roughly 95 percent of those initiatives have produced no measurable return. Not below target. No quantifiable business impact at all.

The usual explanations are that the technology is immature, the talent is scarce, or the tools do not integrate. I have come to believe the real reason is simpler, and more uncomfortable. Most AI strategies stop thinking too early. They stop at the second order.

The distinction is the difference between a pilot that impresses a steering committee and a transformation that changes how an institution actually works.

The four orders of an AI decision

Every significant technology moves through a causal cascade. AI is no exception.

First order: the technology does something. A model drafts a memo, summarizes a filing, flags an anomaly.

Second order: it changes what people do. The analyst reviews instead of writes. The reviewer supervises instead of searches.

Third order: the institution responds. Roles are redefined, controls are rebuilt, and decision rights move as the operating model adjusts to a world where a machine now holds part of the workflow.

Fourth order: the system reorganizes. Functions are restructured, the boundary between firms and their vendors shifts, and the regulatory perimeter is redrawn around the new reality.

Most strategy decks live entirely in the first two orders. They catalog use cases, which is first order, and estimate time saved, which is second order. Then they stop. That is the failure mode. Stopping at second order feels like progress, because the demos work and the productivity numbers look real. But nothing structural has changed, so nothing compounds, and eighteen months later the initiative is quietly folded back into business as usual.

Why everyone stops at second order

It is not a lack of intelligence. It is gravity.

Second order is where the comfortable metrics live. Hours saved, tickets deflected, documents processed. They are easy to measure and easy to defend in a budget meeting. Third and fourth order effects are harder. They require you to touch the operating model, to renegotiate who decides what, and to engage the regulator. That work is slower, more political, and less photogenic, so leaders bank the second order win and declare victory.

There is an honest intellectual trap here too. Productivity is genuinely valuable, so it is tempting to treat it as the prize. It is not the prize. Productivity is the floor. It is bounded by definition, because you can only remove so much friction from existing work. The unbounded upside sits higher up, in redesigned processes and reinvented revenue, and you cannot reach it without doing the third and fourth order work. An institution that frames productivity as the ceiling has quietly capped its own ambition.

The strongest case for stopping, and why it fails

It is worth taking the best argument for caution seriously, because in a regulated institution caution is often correct. Third and fourth order change means altering controls that exist for good reason. A board is right to be wary of anyone who treats a surveillance function or a credit process as a greenfield. So why not bank the productivity and wait for the technology to settle?

The answer is that the cascade does not wait for you. If your competitors are already redesigning, and your supervisor is already revising expectations, then a faster version of your old process is not a safe position. It is a slowly eroding one. The real choice is not between caution and change. It is between change you design and change that is imposed on you. Caution applied to the second order, while the third and fourth order move without you, is the riskiest posture of all.

What the higher orders look like in regulated finance

Consider compliance, a domain I worked in for years while leading an AI Center of Excellence in the compliance function of a global systemically important bank.

The first order story is familiar. A model triages alerts and drafts case narratives. The second order follows. Analysts move from writing to reviewing, and a queue that used to take hours clears in minutes.

If you stop there, you have a faster version of the old process. The third order question is where value is created. If a model now performs the first pass of judgment, who owns the model risk, how do the three lines of defense adapt, and have you begun to extend Know Your Customer discipline to the agents themselves: knowing what each autonomous system is, what it is permitted to do, and who is accountable when it acts. The fourth order question is larger still, and it is not hypothetical. Supervisors are already asking how institutions govern models that shape decisions, and the EU AI Act has written specific obligations for high risk uses in finance, such as creditworthiness assessment, into law, now scheduled to apply from December 2027. The definition of an adequate program is being rewritten while institutions are still running pilots. The only question is whether you help shape that definition or wait to be told what it is.

Governance is what lets you operate at the higher orders

Here is the part that surprises most boards. What lets an institution move confidently into third and fourth order territory is not a better model. It is better governance.

This inverts the usual assumption that governance is the brake. In regulated industries the opposite holds. Governance enables trust, trust enables speed, and speed enables scale. The mechanism is mundane but powerful. When you have pre-approved governance patterns for a class of use case, you redesign within known boundaries instead of renegotiating from zero every time. The first deployment might take six months. The tenth takes six weeks. That compounding is what makes operating at the higher orders possible at all. Where governance is absent, every deployment becomes a fresh negotiation, and the institution stays stuck at the second order, running pilots it can never industrialize.

This is the bridge very few occupy. Many people can tell you what AI does to a task. Far fewer carry the analysis through to how the institution and the system reorganize, and fewer still connect that to the governance that makes the move safe. That bridge, from the third order to the fourth, is where the durable advantage in regulated finance sits.

The test you can run this quarter

Take any AI initiative on your roadmap and ask four questions, one per order. What does it do. What will people do differently. What in our operating model, our controls, and our decision rights must change for that to hold. And what happens to our function, our vendors, and our regulatory posture if this becomes standard across the industry.

If an initiative has crisp answers to the first two questions and blank stares for the last two, it is a second order project wearing a transformation label. That is not a reason to cancel it. It is a reason to staff the third and fourth order questions before you scale it, rather than after.

The position I hold

Stopping at the second order feels like progress and is not. Productivity is the floor. The upside sits in the third and fourth orders, where the operating model, the controls and the regulatory posture change, and governance is what makes that move safe rather than reckless.

So the question for any board is not which model to choose, or how many hours a pilot might save. It is how much of its AI activity has been carried past the second order, into a real change in the operating model and the governance. If the honest answer is very little, the institution is not behind on technology. It is stopped where most institutions are stopped, and the opportunity is to be one of the few that keeps going.