By Daniele Bellomi, Chief Commercial Officer, Intella
Decision architecture, not detection accuracy, is where satellite operations actually breaks — and the data is starting to say so out loud.
Executive brief. The dominant cost in satellite operations is not the rare anomaly. It is the daily, decade-long confirmation that nothing is wrong. The industry has spent ten years optimizing detection sharpness; the next decade’s wins will come from automating the routine. The operators who recognize that first will scale at multiples the rest of the field cannot match.
For ten years, the conversation about AI in satellite operations has had the same shape. Detect anomalies faster. Reduce false negatives. Catch the rare failure before it becomes a $300 million headline. Conferences run sessions on it. Vendor decks open with it. Procurement RFPs centerpiece it.
This framing is wrong. Or, more precisely: it is right about something that is not the problem.
The actual cost center is the boring 80%. It is the engineer who spends most of their shift confirming that nothing is wrong. It is the team running the same nominal health checks for thirty years on a GEO bird. It is the alarm queue full of “expected, benign, repetitive” alerts that operators long ago learned to mute. It is the routine, the predictable, the thing nobody writes a conference paper about — and it is where the cumulative cost of running a fleet hides.
This is not a hot take. It is the increasingly explicit conclusion of the operators themselves.
The 80% nobody automates
A December 2025 industry survey of in-depth interviews with operators, operations managers, system engineers, payload specialists, and reliability experts — across LEO, GEO, institutional, and commercial missions — reaches an unambiguous finding: “Automating routine monitoring offers larger cumulative returns than optimizing rare anomaly handling.”
The reasoning is structural. Anomalies are episodic and focused. They get attention because they are stressful and visible. Routine monitoring is continuous, low-intensity, and cumulative — and it persists for the entire mission lifetime, often decades for geostationary assets. Seasonal effects amplify it. Eclipse seasons mean more analysis, not less. Engineers spend significant fractions of their shifts confirming that nothing is wrong.
The strongest hard-data confirmation of this comes from Planet’s SkySat team. In a 2019 SmallSat Conference paper, “Turning Off the Lights: Automating SkySat Mission Operations,” the operations group documented what happened as their imaging fleet grew from three satellites to fifteen in three years. The forcing function for automation was not anomaly frequency — it was the volume of routine triage work scaling faster than headcount could absorb. Their solution, automating both anomaly response and routine maintenance, produced a three-fold reduction in active operator hours per week while supporting a five-fold increase in on-orbit assets. They eliminated the need for 24/7 staffing at a dedicated operations center.
That is what an order-of-magnitude improvement in operations economics looks like. It did not come from better anomaly detection. It came from automating the work that nobody noticed was the work.
How the routine problem disguises itself as a detection problem
The reason the industry mistook one for the other is that the routine-monitoring problem wears the costume of an alerting problem.
Static thresholds are the standard tool for confirming nominal behavior. They generate large volumes of alerts that are, again in the words of the December survey, “expected, benign, repetitive.” Operators do what humans always do under high false-positive load: they mute, ignore, or filter. Over a long enough horizon, the alarm channel becomes background noise. Real signals get suppressed alongside the false ones.
When something does fail, the post-mortem reveals that an alert had fired, but no one acted on it. The natural conclusion, drawn by every operations review board in the industry, is that detection needs to improve. Sharper thresholds. Better filters. Smarter triggers.
That treats the symptom. The root cause is that the alerting system was generating too much routine work in the first place, and the human in the loop was rationing attention accordingly. You cannot fix that by making alerts smarter. You fix it by making the routine invisible — handled in the background, surfaced only when it deserves attention.
The complexity cliff is no longer theoretical
There is also a scaling threshold beyond which this stops being an optimization argument and becomes a viability argument.
Multiple operators, interviewed independently for the December survey, identified the same inflection. Below roughly five satellites, manual practices suffice. Between five and ten, complexity grows non-linearly — cross-satellite dependencies, scheduling conflicts, telemetry-volume increases that exceed what a small team can absorb. Beyond that range, manual operations become “unsafe or uneconomical.”
The upper end of the curve is now operationally proven. DLR’s German Space Operations Center, in collaboration with Telespazio Germany, published a paper at the 2024 SmallSat Conference describing the automation system that runs the European Data Relay System. EDRS-C is commanded for up to 200 inter-satellite links per day during nominal operations — a volume the authors describe bluntly as “beyond the capabilities of a classical operational concept where only manual operations are performed by a single satellite operator.” The system has logged four years of 24/7 operation and more than one million minutes of inter-satellite communication time.
The line between “manual works” and “manual collapses” is no longer theoretical. It runs somewhere between five and ten satellites, and a meaningful share of LEO operators are about to cross it.
Devil’s advocate
The honest counterargument is that anomaly detection still matters, and reframing the conversation away from it carries its own risk.
Geostationary assets routinely cost between two and four hundred million dollars. A single missed precursor — a thermal drift, a cross-parameter interaction that breaches a threshold no one set — can compromise the mission. Telling operators to redirect investment toward the routine because the routine is bigger by volume invites the obvious failure mode: the team that automates the boring 80% and lets the dangerous 20% degrade.
This is a real critique. It is also not in tension with the thesis. The argument is not that anomaly detection should receive less investment — it is that the next dollar of automation, in most operations centers, returns more if spent on the routine than on incrementally sharper detection. That is an allocation argument, not a substitution argument. Mature operators are not choosing one over the other. They are recognizing that the marginal return on detection improvements has plateaued, while the marginal return on routine automation has barely been touched.
Why this hasn’t happened yet
There is a second reason the industry has been slow to make this shift, and it is structural rather than technical.
The routine-monitoring cost is invisible at the leadership level. Anomaly response is dramatic and costable: there is a meeting, a report, a recovery plan, a number. Routine monitoring is a slow background drag on engineering capacity that never produces a line item. Executives consistently want, and consistently lack, what the December survey calls “high-level health indicators, risk and downtime exposure, productivity and capacity metrics.” Operations teams struggle to express their actual cost structure in terms that leadership can act on.
The result: the most expensive problem in satellite operations is also the least visible to the people who allocate budget. AI investment keeps flowing toward detection — because detection has narratives, victims, and dashboards. Routine automation has none of those, until it runs for two years and the operations-cost line falls sixty percent.
The frontier nobody is naming
Sharper detection will keep getting better. Detection is solvable, and the industry will solve it. But the operators who scale at multiples their competitors cannot match are not going to be the ones with the cleverest anomaly classifier. They are going to be the ones who recognized, earlier than the rest of the field, that the work that needed automating was not the work everyone was talking about.
The frontier in satellite operations is not detection accuracy. It is decision architecture — what gets surfaced to a human, when, and why. That is the conversation the industry has not fully had yet, and the operators who lead it will define the economics of the next decade.
About Intella
Intella is a deep-tech company building Mercury, a mission intelligence platform for satellite fleet operations. Mercury sits above existing ops stacks to automate routine monitoring and reduce operator decision load — handling the nominal in the background, surfacing only what requires human attention, and retaining operational context across shifts and missions.
