In June 2025 Saab and Helsing, the German Münich-based AI specialist, cooperating under a contract from FMV, the Swedish Defence Materiel Administration, announced that a Gripen E fighter flew with a Helsing’s air combat AI agent on board. Known as Centaur, the AI agent executed complex manoeuvres such a beyond-visual-range engagement. Sixteen months later Saab announces has developed its own AI agents and unveiled the AI FPA, an acronym standing for AI Fighter Pilots Academy
Typically, fighter pilots build-up syllabus sees them going to flight school, often referred to as Air Force Academy; Saab virtual “fighter pilots elements”, as we might define AI agents aimed at fighter aircraft, will follow the same path, “attending” the newly formed AI Fighter Pilots Academy. The AI FPA was announced in early November at an on-line briefing by Peter Nilsson, Head of Advanced Programmes at Saab.
According to Nilsson, to create a real AI capability three conditions must be met: first, knowing the problem, “A precise problem gives us meaningful objectives, relevant scenarios and measurable evidence.” Secondly comes domain knowledge, “In our case, which means combining AI experts with combat air experience, aircraft, and mission system engineering, testing and operation analysis. We have all that in Saab. We are working close with the customer. We have been working with AI, which we called machine learning before.” Finally, the capacity to build for scale, “You need computational power, you need to be able to calculate, you need to be able to scale. It does not matter if you have the two first and then you have your desktop computer. You will not get it. And this is not just for defence domain.”
Swedish capacity is based on what is defined the triple helix, government, academia, and industry.
In the Advanced Programmes department around 40% of personnel has or is on his way to a PhD, the latter doing their last assignments at Saab. “I have had people working with machine learning in our battle lab for 10, 15 years, and now the tools are coming,” Nilsson underlines.
Computational power availability being one of the three main conditions to create a real AI capability, Saab partnered with other Swedish companies to create Sferical AI. In August 2025 Saab, together with AstraZeneca, Ericsson, SEB and Wallenberg Investments, with SKF as a strategic partner, announced the creation of the Sferical AI company, and on 15 October 2026 in Linköping the company inaugurated Sferical One, a sovereign, shared AI supercomputer, and Sferical Labs, where the partner companies and Sferical jointly develop AI solutions and expertise. Pooling investments allowed to generate greater computing power, Sferical One being an NVIDIA DGX SuperPOD built on 16 NVIDIA DGX GB300 systems, which started operating in summer 2026 and enables the companies to train and run AI models at a larger scale on a sovereign platform.
“The AI agents I did last year in house, using CPU-based algorithms, I could scale 1,200 times, which means that when I flew a BVR or some other agent one hour in simulator, it meant it was being tested and flown 1,200 times, something quite impressive. Now, with Sferical one hour will be equivalent to one billion hours,” Nilsson said; himself a fighter pilot with a career in the Swedish Air Force, he underlined how no human pilot would ever reach the same level of training.
Considering that AI agents will soon become standard issue on board crewed and uncrewed platforms, and that they will be modular or configurable, the capacity to train them in a very short time with a number of iterations that no human operator would ever be able to reach, will allow to reduce the time elapsed between experience, learning and fielding a capability, what is usually referred to as the OODA (Observing, Orienting, Deciding and Acting) loop.
Saab has been developing AI agents for two years now, Nilsson highlighting that this is not another isolated experiment, but it is the establishment of a true industrial capability. “The AI Fighter Pilot Academy connects four activities in one repeatable loop: we develop and adapt Saab’s AI agents around defined operational needs, we train them in large-scale high-fidelity simulations, we challenge them across demanding and degraded conditions, and we assure the result by building evidence before we test them in aircraft. Industrial capability means repeatability, governance, evidence, and continuous improvement, Saab’s Head of Advanced Programmes stated.
Saab’s ambition is to have a capability platform where it can train AI agents, stress them, and involve them before they reach operational platforms. “We can run high fidelity multi-domain simulations in parallel and iterate behaviours with operation specialists, with real pilots or real C2 operators, or whatever we need. It also allows data and evidence to move between simulation, training, validation, and focused test flights,” Peter Nilsson said. The Academy provides the capability evolve continuously and responsibly, moving from proving the concept, to learning from an iterate, to industrializing the training, assurance, and release, he added.
In the AI FPA, operational scenarios translate experience and emerging needs into relevant training situations. Curriculum turns those scenarios into measurable competence. High fidelity simulation creates scale, repeatability, and demanding condition. Trust is built through evidence across training, testing and validation. And at the centre is a transferable AI fighter pilot core, supported by human-machine interaction before and during flight. “Transferable does not mean copied and changed between aircraft and platform. It means that the methods, evidence, interface, and engineering assets can be reused and learned within a governed environment,” Nilsson pointed out.
Today evolving at speed to adapt to the quickly changing threat is a must. “The closer we are to the operational problem, the faster we can develop a relevant solution. Today activities can be sequential. Flight test capacity is scarce, and learning may remain isolated. With AI Fighter Pilot Academy, simulations can run in parallel. Evidence can be created earlier. And flight validation can focus on the most important uncertainties,” the Saab representative explained, underlining that flight validation will remain essential.
Saab is of course focusing on its own AI products, however AI FPA fives the company the capacity to test and stress third-party AI agents, therefore having an in-house development capability does not rule out cooperation with other companies.
While the Gripen E software driven fighter is the obvious customer for Saab’s AI agents, unmanned platforms such as the Autonomous Collaborative Programme (ACP) A3 prototype unveiled in August 2026 will also be the recipients of those AI agents, unchanged copies or modified and adapted ones, depending on the situation, “however the core will be the same.”
Looking at last year AI insertion into the Gripen E, which considered a BVR situation, this AI agent was bounded to tactical autonomy within a well-defined frame, however in perspective, as evidence and trust grow, the scope of competence will expand. “This is a capability development logic, not a deployment timetable. The next level adds teams’ awareness and coordination, tactical actions, mission level capability, and coordinated execution under commando intent. The longer-term direction is integrated force capability across platforms, mission, and domains. This is where large language models supports the reinforcement learning agents a lot,” Nilsson explained.
IA FPA will allow testing scenarios in which manned and unmanned teams will be fitted with AI agents, with variable success rates depending on the many variables it is possible to play with, the super-high repetition rate allowing verifying complex scenarios in very short time. Should the threat evolve, it will be possible, once again in limited time, to adapt AI agents knowledge to the new situation, once again reducing the OODA loop.
According to Nilsson, in recent years much of the development in AI have shifted from rule-based system to machine learning and to foundation models, and now it is increasingly moving toward word models. The aim is to create word models with the data, which does not only mean recognizing the word, but rather try to build their own words. “And we believe that Saab is potentially well positioned to take on a leadership, a global leadership in this field for defence applications,” the Saab Head of Advanced Programmes said, underlining how the company is operating in a very broad spectrum of physical systems, such as combat aircraft, airborne surveillance, ground combat system, underwater systems, sensors, submarines, command and control system, weapons system, training and simulation environments.
Asked when we can expect seeing AI agents operational on fighter aircraft Peter Nilsson said it might me months or low single digit years, possibly in a Swedish operational test squadron.
Photo ocurtesy Saab
