Most Companies Have a Strategy. Few Have One Their Systems Can Run.

A logistics company chartered a VP-Agent to "minimize shipping costs while maintaining 98% on-time delivery." Clean objective, measurable, sensible. At the Plan stage, the review took 20 minutes — a rubber stamp. The constraint document specified cost and delivery, and on both counts the agent performed beautifully.

What the document didn't specify was minimum carrier diversification. Six months into execution, the VP-Agent had concentrated 73% of volume with the single carrier that offered the best cost-performance ratio. Then that carrier's port operations shut down for eight days. With no alternative routing in place, the company missed $40 million in seasonal commitments. The damage wasn't a model failure. It was a Plan-stage constraint gap — one that two hours of genuine review would have caught.

The translation problem nobody owns

Most organizations have a strategy. Very few have a strategy their systems can act on. The reason is a gap that's invisible until agents are doing the executing.

A CEO articulates a goal: "grow enterprise revenue 40% this year while maintaining margin above 18%." That sentence is perfectly clear to humans. It is not actionable by agents. An agent needs a precise definition of "enterprise revenue" — which segments, which contract types, which geographies — plus measurable leading indicators it can optimize in real time, explicit constraints on what it must not sacrifice in pursuit of the goal, and a definition of success precise enough to evaluate whether a candidate action advances the objective or violates it.

The conventional response is to assume the translation happens somewhere downstream — that someone, somewhere, turns the goal into actionable terms. In a human organization, that someone exists: a chain of managers each interpreting and refining. In an AI-Born organization, if you don't build the translation explicitly, the agents perform it themselves, badly, by filling the gap with their best guess. Nobody owns the translation, so the optimizer owns it.

The reframe: a pipeline from intent to execution

The IPRE Pipeline is the four-stage process that closes the gap: Intent → Plan → Run → Evaluate.

Figure: Inside the Run stage, each agent executes the PRAL Loop — perceive, reason, act, learn — escalating when it fails repeatedly or drifts from the intent the Plan stage encoded.

Intent is the human expression of what the organization wants — the outcome, not the method. It's natural language, but precise about priorities, constraints, and acceptable trade-offs. "Grow enterprise revenue" becomes "increase ARR from enterprise customers (>500 employees) by 40% before December 31, within a gross-margin corridor of 18–24%, prioritizing three-year contracts over annual."

Plan translates intent into machine-executable structure. A planning agent decomposes the intent into objectives, identifies which VP-Agents own each, defines the KPIs they'll optimize, and specifies the constraints — the lines the system must not cross even in pursuit of the goal. This is where the Human Cortex reviews and approves before execution begins. It is the last purely human checkpoint before the system runs.

Run is execution. VP-Agents dispatch specialists, coordinate across functions, allocate resources, and respond to real-time signals — all within the Plan's parameters. The system does not improvise the objective. It improvises the method.

Evaluate closes the loop. Evaluator agents monitor outcomes against the intent definition, detect drift toward proxy optimization (winning the metric rather than the goal), flag anomalies beyond a VP-Agent's authority, and surface what the Human Cortex needs to adjust Intent in the next cycle.

The mechanism: the checkpoint that pays for itself

Of the four stages, the Plan-to-Run boundary deserves the most attention, because it's the last moment human judgment is applied before the system runs autonomously. Once execution begins, the system runs largely on its own until evaluation surfaces a problem. If the Plan was wrong — if a constraint was omitted, if success was mis-specified — the damage accumulates at machine speed before evaluation catches it.

The logistics failure is the canonical illustration, and its lesson is almost embarrassingly cheap: spending two hours at the Plan checkpoint routinely beats spending two months in post-mortems. The constraint that would have prevented a $40 million loss was a single line about carrier diversification. The cost of catching it was attention — the willingness to treat the Plan review as genuine oversight rather than a rubber stamp.

The pipeline also determines your learning speed. A well-structured IPRE — clear evaluation criteria, enforced constraints — enables rapid cycle completion, because the Evaluate stage feeds directly back to Intent revision. That feedback is what produces a short You Didn't Lose the Decision. You Lost the Iteration.. A poorly structured one, where success is ambiguous and constraints go unenforced, generates iteration that produces noise instead of signal — motion without learning.

Sierra runs the full loop in production. At Intent, Bret Taylor and Clay Bavor encode enterprise customer outcomes with explicit quality constraints — resolution completeness, escalation-rate ceilings, satisfaction floors that can't be traded for throughput. At Plan, those become the objective function the VP-Agent layer carries into execution. At Evaluate, if resolution speed improves while repeat-contact rates rise — a sign the agents are closing tickets superficially rather than solving problems — the evaluator flags the tension before it compounds into a trust problem. The Human Cortex reviews and decides whether Intent needs revising. That review takes minutes, not meetings.

What to do about it

1. Translate one goal all the way down. Take a real strategic objective and write its Intent in agent-actionable terms — segments, indicators, constraints, success definition. The act of doing it once reveals how much your current strategy was relying on humans to fill the gaps.
2. Make the Plan checkpoint real. A 20-minute rubber stamp is where $40 million losses are born. Treat Plan review as the genuine human checkpoint it is — the question to ask is "what must this system never sacrifice?" and the answer belongs in the constraint document.
3. Staff the Evaluate stage with teeth. Evaluator agents that only produce green dashboards are mis-calibrated. The Evaluate stage exists to catch proxy optimization — to notice when the metric is winning and the goal is losing.
4. Close the loop back to Intent. Evaluation that doesn't feed intent revision is a report, not a pipeline. The learning lives in the cycle completing, not in any single stage.

The principle

A strategy your systems can't run isn't a strategy — it's a wish that agents will quietly translate on your behalf, in directions you didn't choose. The IPRE Pipeline turns the wish into something executable, and it concentrates the human judgment exactly where it's cheapest and most consequential: at the Plan checkpoint, before the machine starts running flawlessly toward whatever you forgot to forbid.