The AI Reckoning by Dr Kate Barker

What You Need to Know

When you buy “more AI capability,” do you actually gain durable power-or do you just increase your dependence on someone else’s system? That question determines whether AI becomes a source of strength or a new lever held by vendors, platforms, and regulators.

AI changes the distribution of leverage because it moves judgment into systems that you do not fully control. Capability sounds like the right measure-models that can draft, classify, predict, and summarize. But capability does not tell you who owns the dependencies, who can change outcomes, or what breaks trust when something goes wrong. Boards and executives need a different scoreboard: power, dependency, and legitimacy measured as properties of the institution, not as features of a demo.

You will use the Power Stack™ Leverage Map in this chapter. It helps you translate “AI capability” into concrete questions you can govern. Two terms anchor the framework:

• Power: the ability to choose your actions and keep the option to change course without losing control of outcomes.
• Dependency: the degree to which your results depend on conditions you cannot set-like a vendor’s model updates, a platform’s access rules, or a data supply you cannot replace quickly.

A third term matters because boards do not run on performance alone:

• Legitimacy: the organization’s right to act in the eyes of customers, employees, regulators, and the public-backed by evidence that you manage risk, explain decisions, and protect trust.

Elena Park, the CFO of a mid-market insurer, felt the disconnect the hard way. Her team proved the model could accelerate claim triage and reduce handling time. Then risk stepped in: audit trails were incomplete, decision boundaries were unclear, and the “improvement” depended on one external service the insurer could not easily swap. The board did not ask whether the model worked. They asked whether the insurer still owned the decision system once the vendor controlled the updates and the explanations.

This chapter gives you a way to answer that question before you scale.

Breaking It Down

AI shifts power because it bundles three things into one operational stack: compute access, model behavior, and data pathways. When you buy capability, you usually buy the first two only partially, and you often inherit the third as a dependency. The result is that your organization can spend to get faster while losing the ability to steer.

The Power Stack™ Leverage Map starts with a simple move: treat AI as an institutional system with inputs, outputs, and control points. Then score each control point for power, dependency, and legitimacy risk.

1. Separate “showing intelligence” from “owning outcomes.”

A demo proves the system can produce useful outputs. Governance requires you to prove that you can reproduce results, explain them, and keep them stable when conditions change. For Elena’s insurer, triage speed improved, but the explanations and audit trail lagged behind underwriting and claims governance needs. Speed did not equal control.

2. Map the control points where leverage concentrates.

Leverage concentrates at the interfaces where someone else can change your results. In practical terms, look for control points like these:

• Where you fetch model behavior (hosted model vs. on-prem or self-managed).
• Where you transform data before it enters the system (feature pipelines and labeling rules).
• Where you store results and evidence (logging, retention, audit access).
• Where you route decisions (human-in-the-loop thresholds, escalation rules, overrides).

Boards often focus on the model and miss the interfaces. The insurer focused on triage accuracy, but the risk team blocked scaling because the evidence chain did not support review.

3. Score dependency by swap difficulty, not by contract language.

Contracts matter, but swap difficulty determines real leverage. Ask: if the vendor changes access rules or deprecates the capability, how quickly can you replace it with comparable behavior? Elenas team learned that even if they had a contract, replacing the decision support workflow meant rebuilding feature transformations, revalidating outcomes, and retraining evidence processes. That takes more than “a quarter” and it stresses people, not just systems.

4. Measure legitimacy risk as a controllable asset.

Legitimacy risk often looks like “compliance work,” but you should treat it as a design constraint. Identify what regulators and auditors will ask for when the system makes a wrong call:

• What data drove the decision support?
• How did you record the inputs and the model outputs?
• How did you handle uncertainty and edge cases?
• How did you keep the system from drifting without detection?

If you cannot answer those questions with evidence, you do not have legitimacy. You have an ungoverned experiment....