AI Decision Support for Ukraine’s Last-Stand Calls

A “last stand” is rarely cinematic. It’s usually a spreadsheet of shortages, a map full of red arcs, and a commander trying to decide whether holding a town for 72 more hours saves a brigade—or destroys it.

That’s why the debate highlighted by the fighting around Pokrovsk isn’t just about one city or one defensive line. It’s about the hardest operational decision in modern war: when to commit scarce forces to hold ground you’ll probably lose, and when to preserve the force to fight the next battle. Lawrence Freedman and Ryan Evans framed this as an agonizing strategic dilemma. They’re right—and it’s also a clean way to see where AI in defense and national security can help, and where it can absolutely mislead.

Here’s the stance I’ll take: AI won’t “solve” last-stand decisions, but it can make them less blind—by improving real-time intelligence analysis, stress-testing mission planning assumptions, and clarifying the tradeoffs between time, terrain, and troop survival.

The last-stand dilemma is a force-management problem, not a morale problem

A last stand is often described in emotional terms: courage, grit, national will. Those matter. But the operational heart of the issue is simpler and colder:

The core question is whether time gained is worth combat power lost.

Holding a threatened position can buy time to:

• complete a withdrawal elsewhere without panic and traffic jams
• bring up reserves and anti-armor assets
• finish engineering obstacles and minefields on the next line
• protect a logistics hub or key road junction a little longer
• keep civilian evacuation routes open

But paying for that time can mean losing the exact assets you can’t replace quickly—trained infantry, experienced NCOs, artillery tubes, air-defense batteries, EW teams, or bridging equipment. In a war defined by drones, artillery, and attrition, the “irreplaceables” are often people and competencies, not just hardware.

Why Pokrovsk-type fights create brutal incentives

Battles around nodes like Pokrovsk are uniquely punishing because they compress multiple strategic values into one place:

• Terrain value: high ground, cover, lines of sight for drones and artillery
• Network value: roads/rail/logistics flow; a hub’s loss ripples outward
• Psychological value: symbolism and domestic/international signaling

This combination pressures leaders into “one more push” thinking. The trouble is that modern surveillance and fires can turn “one more push” into a predictable casualty event.

Decision-making under drone saturation: the battlefield moves faster than staff cycles

The Russo-Ukrainian war has accelerated the speed at which tactical situations become operational crises. Persistent ISR from UAVs, rapid kill chains, and wide-area artillery coverage mean:

• exposed movement is punished quickly
• units can become fixed, then destroyed, in hours
• logistics routes can flip from “usable” to “suicidal” between morning and afternoon

Traditional decision cycles—briefings, staff estimates, confirmation from multiple channels—can lag behind reality. That’s the opening for AI-enabled tools, but only if they’re designed to support humans, not replace them.

What AI can do well here (and what it can’t)

AI is strongest at pattern extraction across too much data for any staff to digest. It’s weaker at judgment, deception resistance, and political context.

AI can meaningfully assist by:

1. Fusing ISR at speed: correlating drone video, acoustic sensors, SIGINT-like cues, and battlefield reports into a coherent picture.
2. Estimating enemy intent: not “mind reading,” but probabilistic forecasts—e.g., likely axes of advance given prior behavior and logistics signatures.
3. Running fast mission planning iterations: generating multiple withdrawal/hold options with resource requirements and risks.

AI cannot responsibly do (by itself):

• decide acceptable losses
• adjudicate propaganda vs. reality
• weigh political imperatives (aid negotiations, coalition cohesion, public morale)
• guarantee truth when the enemy is actively manipulating inputs

A memorable rule for planners: AI can compress uncertainty; it can’t eliminate it.

“Hold” vs. “withdraw” is really four decisions, not one

Operationally, “last stand or pull back” sounds binary. In practice, it’s a set of linked decisions that can be supported by AI decision support systems.

1) How long can you hold before the position becomes unrecoverable?

The key variable is often not whether you can hold today—it’s whether you can still exit tomorrow.

A position becomes unrecoverable when:

• primary routes are under consistent observation and fire
• alternate routes lack bridging, cover, or engineer support
• ammunition and medical evacuation collapse
• the unit’s cohesion drops below a threshold (fragmentation, loss of leaders)

AI-enabled route risk models can help forecast when roads will cross from “high risk” to “near-certain loss,” using inputs like observed drone density, artillery activity, crater reports, and EW effects.

2) What’s the marginal value of time gained?

Time is only valuable if it enables something. That “something” should be explicit:

• finishing a new trench line
• moving air defense to cover a new logistics route
• rotating a battered brigade out
• positioning reserves for a counterattack

AI-supported planning can quantify time’s payoff by linking it to measurable readiness outcomes (e.g., “72 hours enables emplacement of X obstacles and movement of Y tons of ammo”). If the payoff is vague—“we need to show resolve”—you’re in political territory, not operational optimization.

3) What’s the expected cost in irreplaceable capabilities?

Attrition isn’t linear. Losing a handful of specialists (JTAC equivalents, EW operators, drone pilots, experienced medics) can collapse a unit’s effectiveness.

Good decision support should model capability loss, not just casualty counts.

A practical approach I’ve seen work in defense planning contexts is a “capability weighted loss” score:

• weight personnel by role scarcity and training time
• weight platforms by availability, maintenance pipeline, and repairability
• include “enablers” (comms, EW, drones) because they multiply everything else

This creates a clearer comparison between “hold 48 hours” and “withdraw tonight” than raw casualty estimates.

4) How does the choice affect your theory of victory?

Freedman and Evans tie the dilemma to each side’s theory of victory. That’s the right framing.

If your theory of victory depends on preserving a trained force for a future operational shock (a counteroffensive window, a surge in Western munitions, a change in Russian constraints), then force preservation is strategy, not retreat.

If your theory of victory depends on denying the enemy key logistics corridors long enough to exhaust their offensive capacity, then buying time can be strategy, not stubbornness.

AI can support this by connecting tactical outcomes to strategic assumptions: “If we lose this hub, what downstream changes occur in supply flow, artillery tempo, and reserve mobility?” That’s systems analysis—an AI-friendly problem.

How AI changes last-stand scenarios (and the three traps to avoid)

AI changes last-stand decisions by shifting what’s knowable, what’s fast, and what’s auditable.

The best-case outcome: commanders get earlier warning, clearer options, and better timing—especially for withdrawals, which are hard to execute under fire.

But three traps are common.

Trap 1: Treating AI outputs as truth instead of a forecast

A forecast is not a fact. The enemy’s job is to break your forecast.

Build decision support so it always shows:

• confidence ranges (not single numbers)
• what data it relied on
• what could plausibly be missing or spoofed

Trap 2: Optimizing a metric that doesn’t match reality

If the system optimizes “terrain held” or “enemy losses inflicted,” it may recommend holding too long.

Better operational metrics include:

• probability of successful withdrawal by time window
• projected capability retention after 7/30/90 days
• logistics continuity (fuel, ammo, medevac throughput)

Trap 3: Forgetting that autonomy creates policy problems

Autonomous systems—loitering munitions, automated targeting queues, robotic ground vehicles—can speed up fights. They also increase escalation risk and raise legal and accountability questions.

For national security leaders, the question isn’t “Can we automate more?” It’s:

Which decisions must remain human-owned because they carry irreversible political and moral consequences?

That boundary should be defined before a crisis, not improvised during one.

Practical playbook: AI-enabled decision support for contested withdrawals

A withdrawal under drone observation is one of the hardest operations in war. If you’re building or buying AI mission planning tools for defense organizations, focus on these capabilities first.

Build a “withdrawal viability dashboard” (not a generic COP)

A common operational picture is useful, but last-stand calls need specific indicators. A viable dashboard prioritizes:

• route exposure scores (by hour)
• enemy drone density trends and likely handoff to fires
• availability of smoke, EW coverage, and counter-UAS assets
• medevac queue times and trauma capacity
• ammunition burn rate vs. remaining stocks

The intent is blunt: prevent the moment when a unit is still alive but no longer extractable.

Use AI to generate options, then force a human to pick the assumption set

The best workflow I’ve found is “AI drafts; humans decide what’s true.”

For each course of action (COA), require explicit assumptions:

• “enemy artillery resupply remains constrained”
• “EW coverage holds for the first 6 hours”
• “bridge X remains intact”

When assumptions become visible, leaders can argue about the right things.

Run red-team models continuously

If AI is supporting operational decisions, it must also support deception resistance.

That means:

• anomaly detection for spoofed drone feeds and fabricated reports
• adversary intent modeling that includes deception incentives
• regular “model drift” checks as tactics evolve

In Ukraine, tactics adapt quickly. Any model not updated frequently becomes a liability.

People also ask: Does AI make commanders more likely to choose a last stand?

It can, but only if the system rewards the wrong outcomes. If AI tools emphasize territory and enemy attrition, they’ll bias leaders toward holding. If they emphasize capability retention, extraction probability, and logistics continuity, they’ll bias leaders toward timely maneuver.

A clean principle for AI in national security planning: design the tool around the decision you want humans to make under stress. Metrics are policy.

What this means for the AI in Defense & National Security series

The fight over places like Pokrovsk puts a spotlight on how war actually works in 2025: dense sensing, fast fires, contested logistics, and leadership decisions made with incomplete information.

AI’s real contribution isn’t bravado about autonomy. It’s the quieter work: turning scattered battlefield signals into decision-quality intelligence, and turning mission planning from one plan into many options with explicit tradeoffs.

If you’re responsible for defense innovation, procurement, or operational experimentation, your next step is practical: audit your planning and intelligence workflows for where time is lost and where uncertainty is hidden. Then build AI support around those choke points—route viability, kill-chain warning, capability-weighted attrition, and assumption tracking.

The hardest question will remain human: When does holding ground stop serving strategy and start consuming the force? AI can’t answer that for you—but it can make sure you’re not answering it in the dark.