Visor Dynamics · Geospatial Intelligence
All articles
Strategy 13 min read

Collapse the OODA loop

Whoever observes, orients, decides and acts faster than the opponent sets the pace. Where each step slows down, and how software is shortening the loop on the battlefield and in business.

What the OODA loop is

The OODA loop is a decision cycle described by John Boyd, a US Air Force fighter pilot and strategist. It has four steps:

  1. Observe: take in what is happening.
  2. Orient: work out what it means, using what you already know and believe.
  3. Decide: pick a course of action.
  4. Act: carry it out, then observe the result and go round again.

Boyd argued that the side that goes round this loop faster and more accurately than its opponent wins. Do it well enough and the opponent is always reacting to a situation that has already changed.

How the military uses it

Boyd took the idea from air-to-air combat and later applied it to strategy as a whole. It shaped US Marine Corps doctrine on maneuver warfare.

Militaries use the loop in two ways. The first is to run their own loop faster, getting from a sensor spotting a target to a decision and an effect in minutes instead of hours. The second, which Boyd cared about more, is to break the enemy's loop. Deception, surprise and constant change corrupt the enemy's Orient step. Their picture of the battlefield drifts away from reality, their decisions get worse, and eventually their command system freezes or falls apart. Boyd was after that kind of collapse, which can beat an army without destroying it.

Today's doctrine uses terms like "decision advantage", "compressing the kill chain" and "shortening sensor-to-shooter time". They all describe the same loop.

How business and startups use it

The OODA loop left the military decades ago. Chet Richards applied Boyd's ideas to companies in Certain to Win (2004), and the loop shows up wherever competition moves quickly.

It explains a lot about why small startups beat large incumbents. The incumbent usually has more money, more data and more customers. What it lacks is a fast loop:

  • Observe: customer feedback filters up through account managers, reports and quarterly reviews. By the time leadership sees it, it is months old.
  • Orient: the company's model of the market is built on whatever made it successful, and information that contradicts that model gets explained away.
  • Decide: decisions go through committees, budgets and sign-offs.
  • Act: a change ships in the next release cycle, if it ships at all.

A startup can go round the same loop in days. The founders talk to customers themselves, change their thesis when the evidence says so, decide in one meeting and ship that week. Each turn teaches them something. After a year they may have run fifty experiments to the incumbent's four. The Lean Startup's "build, measure, learn" cycle is the same idea under another name.

Some large companies work hard to keep this speed. In his 2015 shareholder letter, Jeff Bezos split decisions into one-way doors, which are hard to reverse and deserve care, and two-way doors, which are easy to undo and should be made quickly by small teams. Amazon used that split to keep the Decide step short wherever being wrong is cheap.

Agility matters more than raw speed

People often read Boyd as saying "go faster". Speed helps, but what wins is agility: changing your orientation when the world changes, and acting on the new picture before your competitor does.

A team that executes quickly on a wrong model of the market just fails sooner. A military that strikes quickly on bad intelligence hits the wrong targets sooner. A fast loop is valuable because it corrects itself. Each turn tests your picture against reality, so the faster you turn, the less time you spend acting on a mistaken view.

This is why Boyd treated Orient as the most important step. It decides what you notice, what you think it means and which options you even consider. In our experience, agile organizations, whether military units or startups, are the ones whose people are allowed to change their minds quickly.

Where each step gets stuck

Delays build up in different places at each step.

Observe: too little data, or too much

For most of military history the problem was not seeing enough. Now it is often seeing too much. One drone can record hours of video a day, and a fleet produces far more footage than any team can watch. Radar and RF receivers pick up thousands of signals, and satellite imagery arrives by the terabyte.

The number of drones is growing fast. Hundreds of thousands are already in service. The Pentagon alone aims to field more than 300,000 armed drones by 2027 (Defense One), and Ukraine's Delta system reportedly handles up to 10,000 drone video feeds a day. We expect the count to pass a million before long. At that scale, flying the drones is the easy part. The hard part is deciding which drone looks where, keeping track of what each one has seen, and analyzing the footage while it still matters. No team of analysts can watch a million feeds, so coordinating and analyzing drone fleets is going to be one of the biggest bottlenecks in the loop. We don't see a way to handle it without software doing most of the work.

Common failures at this step:

  • Sensors record data that nobody looks at in time.
  • Footage is archived but can't be searched, so finding last Tuesday's convoy means scrubbing through video by hand.
  • Imagery is too coarse, or too degraded by haze and compression, to answer the question being asked.

Businesses have the same problem in the form of dashboards nobody reads and customer interviews that never reach the product team.

Orient: where most loops slow down

Raw observations have to be combined into one picture people can act on, and they almost never arrive ready for that.

  • Video, imagery, maps, signals and reports sit in separate systems.
  • They use different formats, coordinate systems and symbol sets.
  • Analysts cross-reference them by hand, matching a radar hit to a drone frame to a grid reference buried in a text report.
  • Each unit keeps its own picture, so two commanders can be looking at different versions of the same situation.

Orientation also breaks down from the inside. People and organizations hold on to models that used to be right, and new evidence gets squeezed into the old picture instead of changing it.

Decide: approvals, overload and too few options

Even with a clear picture, decisions stall:

  • Authority sits too high, so every decision waits in a queue.
  • Decision makers are buried in information and can't tell which parts matter.
  • Planning a course of action takes so long that only one option gets developed, when several should be compared.
  • Doctrine, legal rules and past orders are spread across manuals that nobody can search under time pressure.

Act: the handoff

Once a decision is made, it has to reach whoever will carry it out. Coordinates get read out over the radio and typed back in by hand. A target found in one system is entered again in another. Communications get jammed. Each manual handoff adds minutes and another chance of error.

In software companies, this is the gap between deciding on a fix and getting it into production.

Most of these handoffs exist because systems aren't connected. Network-centric warfare, an idea from the late 1990s, was meant to fix that by putting sensors, command posts and shooters on one network so information moves as data. In practice this means central systems should connect to tactical data links such as Link 16 and to the wider battle network. When they do, a target confirmed in the command post can reach the pilot or the gun crew without someone reading coordinates over the radio. Each system left out of that network brings the manual handoff back.

The catch is that data links and fire-control networks are some of the most sensitive systems a military has. If an AI model, or anyone sitting at a connected workstation, can write to them directly, one stolen password could put false tracks or orders in front of forces in the field. We worry about AI in particular here, since an adversary can feed a model misleading data on purpose and wait for it to pass the error along.

So we don't think AI or ordinary users should ever touch these links directly. Their traffic should go through a secure middleware layer between the decision tools and the networks that carry orders. The middleware knows who or what sent each message and whether they're allowed to, and it rejects anything that breaks the message format or the rules of engagement. An AI can draft a message there, but anything with real consequences sits in a queue until a person releases it. Moving data between classification levels happens only through controlled gateways, and the middleware logs every message so it can be traced later.

This keeps most of the speed of a connected network while making sure no model or laptop has a direct line to the weapons. We'd rather give up a few seconds at that checkpoint than lose the commander's trust in what comes out of the loop.

How we compress the loop

We build software for every step of the loop, with most of our work in Observe and Orient, where we see the longest delays.

Observe

  • Elite SPRINT takes live drone video, orthorectifies it and drapes it on 3D terrain as it arrives. It records and indexes footage at scale, so analysts can find earlier video by place and time instead of watching it end to end. It works with various military and DJI drones out of the box.
  • Cognitive EW receives and profiles radar and RF signals in real time, with several layers of AI analysis across pulse, frequency-hopping and IQ data.
  • Airborne ESM processes direction-finding streams from fixed-wing, rotary and UAV platforms and shows threat locations as soon as they are found.
  • Super Resolution sharpens satellite imagery up to four times and reduces haze and compression artifacts. We designed the architecture to sharpen what is in the image without inventing detail that isn't.

Orient

  • Elite Mapper shows raster imagery, vector data and elevation together in a live 3D view, in the cloud or on-premises. Video, signals and imagery end up in one shared picture instead of three separate windows.
  • Command & Control combines electronic order of battle, direction-finding and jammer data from a distributed sensor array into a single operational picture.
  • Our platforms read and write standard formats such as STANAG 4609 and MISB KLV for video metadata, NITF for imagery and MIL-STD-2525 for symbols. Data from other systems arrives readable, so the Orient step doesn't have to wait for a conversion.

Decide

  • Elite AI is a reasoning harness built on defense doctrine and manuals. It helps plan missions, assess threats and search large document collections, so a commander can weigh several options instead of one. It runs fully offline in air-gapped environments, and sensitive data never leaves the network. It helps with the decision, but a person makes it.

Act

Each output uses standard formats and coordinates, so it can pass to the next system without being typed in again. That cuts delays and errors at the point where they cost the most. Our tools, Elite AI included, are meant to reach the battle network through a controlled middleware layer like the one described above, and never by writing to data links directly.

Where this is heading

The United States: Maven and CJADC2

The US Department of Defense's main effort is CJADC2 (Combined Joint All-Domain Command and Control), which aims to connect sensors and shooters across all the services. In April 2026, Deputy Secretary of Defense Steve Feinberg issued a directive calling AI-enabled decision-making the "cornerstone" of CJADC2, built on Palantir's Maven Smart System. The fiscal 2027 budget request includes more than $1.5 billion to expand access to Maven, which is becoming a formal program of record (DefenseScoop, DefenseScoop).

The Pentagon also folded its Replicator drone program into a new Defense Autonomous Warfare Group, which has a $54.6 billion request for fiscal 2027 (Defense One). Replicator's biggest reported difficulty was software to coordinate large numbers of different drones. That is an Orient and Decide problem more than a hardware one.

Ukraine is already doing this in practice. Its Delta system reportedly flags enemy equipment in drone video in about two seconds and passes targets to artillery planning tools without anyone retyping them. A person still approves each strike.

China: intelligentized warfare

The People's Liberation Army calls its version "intelligentized warfare". PLA writing treats it as a change in command and control more than in hardware, with machines carrying out tasks within objectives, constraints and escalation limits that humans set (FDD). Chinese state television has shown an AI-assisted system for planning large air strikes across hundreds of targets.

A 2026 US futures assessment gives the PLA roughly an even chance of reaching operational decision advantage over a peer adversary by 2035 (AY26 Futures Seminar). Analysts also point to limits. The PLA has less real operational data than the US, and its centralized command culture fits poorly with fast, decentralized loops (Defense News). We'd call that an orientation problem, and we doubt better algorithms will fix it on their own.

From a loop to a line

Both countries are moving in the same direction. Observe and Orient are being automated first. Decide is shrinking to a human approving or supervising a recommendation, and Act is being spread across large numbers of cheap autonomous systems. A recent US Army War College essay describes the end state as "Predict, Act", where models forecast what will happen and act once their confidence passes a threshold a commander has set (War Room).

We have mixed feelings about where that leads. Under time pressure, people tend to accept whatever the machine recommends, and human oversight can turn into a rubber stamp. Two automated systems reacting to each other could escalate a crisis faster than anyone can step in. A fast loop running on poisoned data is also easy to deceive, which is Boyd's idea of collapsing the enemy's loop turned against you.

Speed with judgment

Collapsing the OODA loop shouldn't mean taking people out of it. It should mean taking out the waiting, such as the hours of footage nobody watched, coordinates typed in twice, and systems that don't talk to each other. Machines can handle that work and leave people more time for orientation, which still needs human judgment.

The same holds for a startup. The teams that win are usually the ones that see their situation clearly and change course when the evidence says so, early enough for it to count.

References

More articles

View All Articles
AI

Is a military AI agent just a chatbot with tools?

Building an AI agent now takes minutes. Building one a military can trust with real missions is a different problem, and most of the hard parts have little to do with the model.
Strategy

Is mosaic warfare already here?

DARPA imagined a force of cheap, interchangeable pieces. Ukraine, Iran and the Pentagon suggest much of it is here, except the software to hold it together.
Engineering

Would you trust vibe-coded defense software?

AI can write code faster than anyone can read it. In defense, that raises hard questions about security, accountability and who can still fix the system when it fails.

© Visor Dynamics (Pvt) Ltd.

Visor Dynamics · Geospatial Intelligence