The Next AI Problem Isn’t Adoption. It’s Decision Governance.

Over the past several weeks, I have been looking at AI from a few different angles. The questions seemed separate at first.

• Why can the same AI task cost dramatically different amounts?
• How much should a company pay for a better answer?
• Why can average AI cost hide significant operating risk?
• Why is workplace AI adoption broad but still shallow across actual job tasks?
• What happens when AI makes marketing cheaper—but the advertising performs worse?

Different studies. Different business problems. But they keep pointing toward the same conclusion.

The next enterprise AI problem may not be getting companies to use AI. It may be deciding where AI should be used, how much intelligence should be purchased, what performance is good enough, and when the economics justify continuing.

That is a management problem. And increasingly, it is a decision-system problem.

---

AI Is Creating More Decisions, Not Fewer

The early AI question was straightforward: “Can AI do this?” Then came: “Can we put it into a workflow?” Those questions still matter. But once AI begins operating throughout an organization, a much larger set of questions appears:

• Should this task use a frontier model or a cheaper model?
• Should the system retry, and should it spend more computation trying to improve confidence?
• Is the improvement worth the additional cost?
• Should this workflow be automated at all?
• Should a human remain in control?
• Is the AI actually improving the business outcome—or merely making the production process faster?

These are not model questions. They are allocation and governance questions.

Cost Is Only One Part of the Decision

Agentic AI can behave differently from one execution to another. A task may require more reasoning, more context, more tool calls, more retries, or more verification. So the real cost of completing the work may not be known when the work begins.

Managing only average cost is not enough. Companies need to understand the full cost distribution:

Mean Cost → P95/P99 Tail Cost → Retry Behavior → Model Escalation → Quality Improvement → Cost per Successful Outcome

At some point, another retry, another reasoning pass, or another premium-model call may cost more than the improvement is worth.

How much reliability should the system be allowed to buy? Maximum quality is not necessarily the economically optimal quality.

Adoption Creates the Same Problem

Workplace AI adoption has spread rapidly, but when measured at the task level, the picture is much shallower. AI has entered many jobs without yet entering most of the work inside those jobs.

The traditional adoption metric: “How many employees use AI occasionally?”

The economic metric: “Which economically important tasks have actually changed because of AI?”

A company can have high adoption and modest value realization—the activity metric can improve while the business metric barely moves.

Marketing Makes the Problem Concrete

A recent advertising study found AI-generated ads were faster and cheaper to produce and often looked polished enough to be indistinguishable from human-made work. But on average, the AI-generated ads performed worse on short-term sales potential and long-term brand equity.

AI can improve the efficiency of producing an output while reducing the effectiveness of the outcome.

A CFO can report lower production cost, a CMO can report faster cycles, and a CEO can report increased adoption—and the business can still be worse off.

Output Metrics Are Not Outcome Metrics

AI governance must connect the activity to the outcome. We cannot rely solely on output metrics.

Output Metrics (Activity)

• Tickets Completed (Service)
• Code Generated (Software)
• Analysis Produced (Finance)
• Cost per Asset (Marketing)
• Model Usage (AI)

Outcome Metrics (Business Value)

• → Customer Retention
• → Production Reliability
• → Better Capital Allocation
• → Persuasion & Equity
• → Greater EBIT Contribution

The Role of a Decision System

This is the gap I have been building Decision Systems AI around. A Decision Systems Orchestrator sits between AI capability and business action. Instead of simply asking AI to produce outputs, it evaluates incoming data against explicit rules to turn signals into governance decisions:

Scale | Continue Measuring | Route to Cheaper Workflow | Redesign | Add Governance | Escalate to Human | Pause | Stop

The purpose is not to make AI less capable—it is to make the organization more deliberate about where capability is worth paying for.

The Cheapest Sufficient Intelligence

Some work deserves premium AI, some work can use smaller models, some should remain deterministic software, and some should be routed to a human.

Marketing gives a simple example. If AI can effectively execute a straightforward product demonstration or a localized variation, use it. If an important brand campaign depends on distinct strategic insight, emotional tension, or surprising point of view, do not optimize for cost per asset.

The question is not: “Can AI do the work?” It is: “What is the right combination of intelligence, cost, quality, and human judgment for this outcome?”

From AI Adoption to AI Management

As AI spreads, the first phase of access gives way to a phase of discipline. Companies must decide:

• Where to allocate expensive intelligence vs. where cheaper intelligence is sufficient.
• How much reliability to purchase, and when retries should stop.
• Which workflows deserve redesign vs. which should be stopped.

Those decisions cannot live inside a prompt—they need explicit metrics, thresholds, ownership, and governance.

The real measure of AI maturity is not how much AI a company uses. It is whether the company knows where AI creates value, where it creates risk, and what to do next.