AI Mission Planning Lessons from Ukraine’s Last Stand

Ukraine’s hardest battlefield decisions in late 2025 aren’t about bold offensives. They’re about when to stop paying for ground with people.

The dilemma described by Lawrence Freedman and Ryan Evans—whether to commit forces to a desperate defense (a “last stand”) or preserve combat power by withdrawing to the next line—shows up in every modern army’s planning cycle. What makes Ukraine different is the intensity: dense drone reconnaissance, constant artillery risk, rapid attrition, and a logistics fight that never pauses. That combination turns “hold or pull back” from a commander’s intuition into a problem that begs for data-driven decision support.

This matters for anyone working in the AI in Defense & National Security space because it’s a real-world case study in what AI is actually good at: not replacing command, but improving mission planning, resource allocation, logistics forecasting, and risk estimation under time pressure.

The “last stand” dilemma is an optimization problem in disguise

The core point: A last stand is justified only when the strategic value of time and terrain outweighs the long-term cost of losing trained force. If the cost curve is wrong—even by a little—your “heroic defense” becomes self-inflicted operational collapse.

Freedman’s framing (as captured in the podcast summary) highlights two competing imperatives:

• Hold ground to protect key nodes, buy time, and deny the enemy momentum.
• Preserve the force so you still have capable units for the next defensive line and future operations.

What commanders are really trading: time, cohesion, and optionality

This decision isn’t just “land vs. lives.” It’s a three-part trade:

1. Time: How many days (or even hours) does the defense buy, and what does that time enable? Reinforcements? fortifications? political decisions? evacuation of civilians? repositioning of air defense?
2. Cohesion: A unit that retreats intact can fight again. A unit that’s attrited, fragmented, or cut off can’t—even if it technically “survives.”
3. Optionality: Preserved combat power gives you choices later (counterattack, plug a breakthrough, rotate exhausted brigades). A destroyed brigade removes options and creates cascading risk across the front.

AI can’t decide which value matters most—only leadership can. But AI can measure, simulate, and make the trade explicit instead of implicit.

Where AI decision support helps most: forecasting attrition and breakthrough risk

The fastest way to lose a war of attrition is to keep treating attrition as an after-action report instead of a forecast.

AI-enabled decision support can reduce that blindness by building predictive risk estimates around the “hold or withdraw” choice.

1) Predicting the probability of being fixed, flanked, or encircled

In places like Pokrovsk (highlighted in the RSS summary), what kills defenders isn’t just direct assault. It’s loss of mobility:

• key roads under persistent drone observation
• bridge and culvert chokepoints
• artillery “fire sacks” on predictable withdrawal routes
• electronic warfare effects that reduce friendly situational awareness

AI can support by fusing:

• drone feeds and change detection on terrain and routes
• signals data (where available) for enemy maneuver cues
• historical patterns of enemy assault tempo after successful probing

The output shouldn’t be a mystical “AI says retreat.” It should be something command staff can use:

• encirclement risk score over the next 24/48/72 hours
• estimated time-to-interdiction for primary and alternate routes
• confidence intervals that show what’s known vs. guessed

2) Estimating marginal cost: “How many troops per kilometer per day?”

Here’s a practical planning lens: What is the marginal cost of holding this line for one more day?

A mature AI mission planning tool can estimate that marginal cost by modeling:

• casualty rates by unit type and defensive posture
• expected artillery expenditure and resupply feasibility
• drone attrition and replacement cycles
• medevac and evacuation capacity under fire

When you see marginal cost spike—because routes are compromised or the enemy has gained fire-control—you don’t need a philosophical debate. You need a decision.

3) Stress-testing assumptions with simulation, not optimism

Most “last stands” start with an assumption: reinforcements will arrive, the enemy will pause, our fires will disrupt their assault, the weather will reduce UAVs.

AI can help by running Monte Carlo-style simulations against those assumptions:

• What if reinforcements slip by 36 hours?
• What if UAV losses are 2× higher than planned?
• What if the enemy commits a fresh battalion from reserve?

Good simulation doesn’t predict the future. It reveals which assumptions you can’t afford to be wrong about.

AI logistics and resource allocation: the unglamorous part that decides “hold” vs. “go”

The blunt truth: you can’t defend what you can’t supply. In a high-tempo fight, logistics isn’t a support function—it’s a constraint that shapes tactics.

AI in defense logistics becomes most valuable when it answers a simple operational question:

“Can we sustain this position for the next 72 hours without burning the force?”

Ammunition, drones, and spares are now “front-line” consumption

Ukraine’s war has highlighted a reality many militaries are still adjusting to: consumption isn’t just shells and fuel.

• FPV drones are consumed like ammunition.
• EW systems require spares, power solutions, and constant repositioning.
• Vehicles degrade quickly under rough movement and near-constant exposure.

AI supply chain analytics can forecast:

• expected daily consumption rates by sector
• route viability under observed interdiction patterns
• inventory positioning to minimize “last mile” exposure

That directly influences the last-stand dilemma. If an AI model shows that resupply will fall below a survivable threshold by tomorrow night, “hold at all costs” stops being bravery and starts being negligence.

Force preservation is also a personnel and training pipeline question

A last stand spends more than manpower. It spends experience.

AI workforce analytics (already common in the private sector) has a defense analog:

• How many trained specialists (UAV pilots, medics, EW techs, NCOs) are at risk in this sector?
• What’s the replacement time for those roles?
• Which units are “brittle” due to personnel churn?

If holding one town risks the loss of a unit’s veteran core, the long-term cost can exceed any short-term terrain value.

AI for intelligence analysis: detecting the enemy’s theory of victory

Freedman and Evans point to “theory of victory” and shifting battlefield realities. That’s the right frame. Wars aren’t just exchanges of firepower—they’re competing plans to make the other side crack.

AI-enabled intelligence analysis can help identify which plan the enemy is executing by detecting changes in pattern:

• recon intensity and axis selection
• artillery allocation (harassment vs. preparation fires)
• tempo of assaults after probing actions
• reserve commitment indicators

From “what happened” to “what’s the enemy trying to force us to do?”

The best use of AI here is intent inference support:

• If the enemy is pushing hard on a sector that seems tactically mediocre, AI can help ask: is this a shaping attack to pull reserves? Is it designed to trigger a politically painful retreat? Is it a logistics interdiction play?

This is where human judgment stays central. AI surfaces candidate interpretations and relevant anomalies; commanders decide which narrative matches broader context.

The guardrails: what AI must not do in high-stakes defensive decisions

AI can help commanders avoid catastrophic misreads. It can also amplify them.

Three guardrails matter in a “last stand” decision environment:

1) Don’t let dashboards replace ground truth

If your model says a route is viable but troops report it’s under constant FPV threat, the model is wrong. Full stop. AI outputs need feedback loops from frontline reports, not just sensor feeds.

2) Treat adversarial adaptation as normal, not exceptional

Russia (and any capable opponent) adapts quickly. If an AI model becomes predictable—what it weights, what it recommends—an enemy can shape behavior. Assume adversarial machine learning pressure as a baseline.

3) Make uncertainty visible to decision-makers

A clean number without uncertainty invites overconfidence. Good AI mission planning tools show:

• confidence bands
• missing data warnings
• “what would change this estimate?” sensitivity prompts

If leadership can’t see uncertainty, they can’t manage risk.

Practical takeaways: how defense teams can build AI that helps, not distracts

If you’re building or buying AI for national security missions, the Ukraine “last stand” dilemma points to a concrete roadmap.

Build for decisions, not demos

Start by mapping the actual staff questions:

• What’s our expected casualty rate if we hold 24/48/72 hours?
• What’s the probability our primary withdrawal route is cut?
• What’s the earliest time we become unsustainable on ammo/medevac?
• What indicators would tell us the enemy has committed reserves?

Then build AI around those outputs.

Use a “decision rehearsal” loop

The teams I’ve seen succeed treat AI as a rehearsal tool:

1. Define the decision (hold, delay, withdraw, counterattack).
2. Run multiple simulated futures with clear assumptions.
3. Pre-commit triggers (“If Route Bravo is interdicted for 4 hours, execute withdrawal plan C”).
4. Update continuously as new ISR and reports arrive.

That structure reduces panic decisions—and it reduces the odds of a last stand happening because nobody wanted to be the first to say “we should pull back.”

Make interoperability and speed your design constraints

In a high-intensity fight, “good but slow” is often useless.

• prioritize edge-friendly models where possible
• design for degraded comms and partial data
• exportable summaries for commanders (one page, not 40 charts)

Where this goes next for AI in Defense & National Security

Ukraine’s war keeps reminding planners that the decisive edge isn’t a single platform. It’s the ability to sense, decide, and move faster than the opponent while keeping your force intact.

The “last stand dilemma” is exactly where AI should earn its keep: clarifying tradeoffs, forecasting risk, and improving logistics realism—so leaders can choose when to hold and when to preserve the force.

If you’re responsible for mission planning tools, analytics, or operational AI, a useful next step is straightforward: audit one recent defensive scenario and ask what you would’ve needed to decide earlier—route risk, attrition forecasts, resupply viability, or enemy intent signals. Build there.

The forward-looking question that matters for 2026 planning cycles is uncomfortable but necessary: When the next “Pokrovsk moment” arrives, will our AI systems help commanders avoid a doomed last stand—or will they just generate prettier maps?