I learned everything I know about space travel from business. I have worked on consumer data since AOL and Yahoo, and spent more than 20 years at companies now part of Omnicom, most recently as global chief technology officer and then chief AI officer. My work has been on consumer data, the trail people leave when they shop, browse, and click, and on restricted data in healthcare and financial services, where the law decides who may see a record and why. I never worked on a spacecraft. But as I began to understand NASA’s plans for a crew on its way to Mars, and the AI that crew will rely on, I kept finding problems I had already lived through.

The core problem is time. NASA plans for delays of "up to 22 minutes one-way for Mars." Researchers at NASA Ames, KBR, and San José State, reviewing 20 years of delay studies , concluded that “certain roles and responsibilities will need to shift from the ground team to the onboard crew, depending on latency duration.” The crew, and the systems it carries, will have to decide alone.

Businesses that run on consumer data, and the hospitals, insurers, and banks that hold restricted data, have been deciding on late, doubtful data for years. Here are six lessons they taught me.

Personalization that works for almost everyone can still hurt the one person it gets wrong. In December 2014, Facebook’s Year in Review greeted web designer Eric Meyer with "Eric, here’s what your year looked like!" above a photo of his daughter, who had died that year. Meyer called it " inadvertent algorithmic cruelty ,” the result of code that works in the overwhelming majority of cases. The product manager behind the feature apologized to him directly .

Space crews are no different. In the 520-day Mars-500 isolation study, researchers "observed substantial inter-individual differences in the behavioral responses of crewmembers." In NASA’s HERA habitat, where crews live sealed off to rehearse missions, they "ignored MCC messages if they were irrelevant to current task or outdated." MCC is mission control.

NASA already tailors some support to each crewmember. But a crew of four, the size of NASA’s yearlong Mars simulation crews , is too small for lookalike models, which guess what one person wants from thousands of people like them. Onboard AI has to learn each person against their own baseline.

Decide Before The Data Arrives

Marketers lost real-time measurement for iPhone users who declined tracking when Apple launched App Tracking Transparency in April 2021. Apple’s replacement reports results late on purpose: "The random delay is 24–48 hours for the first postback." Snap felt it within months. In October 2021, its chief executive told investors that Apple’s new measurement tool "did not scale as we had expected, making it more difficult for our advertising partners to measure and manage their ad campaigns for iOS."

A 2019 NASA technical memo proposes “assigning a latency factor to each data product” so medical information can be prioritized for the link back home. A 2015 study of delayed communication, summarized in a 2025 NASA review, found participants ”misapplied the conventions of synchronous communication to asynchronous contexts." Every data point needs a timestamp that says how stale it is, and a rule for what to do before fresher data shows up.

Two Kinds Of Lie Detection

The Media Rating Council, the US body that audits ad measurement, splits fake ad traffic (views and clicks no real person made) into two. The simple kind is caught by "routine means of filtration." The sophisticated kind needs "advanced analytics, multi-point corroboration/coordination, significant human intervention.” Financial services has a case where the routine check passed. When JPMorgan Chase asked to check the 4.25 million users a startup called Frank claimed, the founder hired a data scientist to fake a customer list ; the real number was about 300,000, and the bank paid $175 million. She was convicted in March 2025 and is appealing.

Spacecraft and aircraft face the same split. The European Space Agency calls fault detection “a critical function on board all spacecraft." Aviation regulators warn that detecting a spoofed satellite navigation signal "may be more difficult and not immediate for the flight crew." Cheap filters catch the obvious errors. Deliberate deception needs a second, slower check.

In automated ad buying, where software bids for ad space in milliseconds, buyers could not check who was reselling a website’s ad space. A small public text file, ads.txt, introduced in 2017, let publishers "publicly declare the companies they authorize to sell their digital inventory." In a December 2023 study of DSPs (the software advertisers use to bid for ads automatically), the ANA found "only 36 percent of every dollar invested by an advertiser that enters a DSP effectively reaches the consumer."

In space, two systems disagreeing about what a number meant has already cost a mission. NASA lost the Mars Climate Orbiter in September 1999 because “ground software used English units, while onboard software worked in metric.” I could not find a public spacecraft standard for declaring a data record’s chain of custody. Marketing had to build one. A Mars mission will need one too.

Consent Rides With The Data

In Europe, your answer to a cookie banner is coded into a short string that rides along with every ad call, so each company down the line knows what you agreed to. Then the courts looked at the string itself. In March 2024, Europe’s highest court ruled that a consent string like this “constitutes personal data” when it can be linked to an identifier such as a device’s IP address. In May 2025, a Belgian court upheld a €250,000 fine against the industry body that runs the system. The record of what you agreed to became protected data in its own right.

Healthcare goes further. HIPAA requires reasonable efforts "to limit protected health information to the minimum necessary to accomplish the intended purpose of the use, disclosure, or request." The purpose travels with the record, and so does the limit.

NASA applies a similar rule to its astronauts. "Crews must be 're-consented’” before their medical data is used for research. Onboard AI should carry consent as part of the data, and should ask again when the purpose changes.

Marketing learned to make a model prove itself before it spends. Google will not estimate the sales it cannot track for an advertiser with fewer than “700 ad clicks over a 7-day period.” Researchers who compared methods found that judging an ad by watching who saw it, with no comparison group, can mislead: such methods “often fail to produce the same effects as the randomized experiments.” Real estate shows what is at stake when a forecast spends real money. Zillow’s home-buying business depended on forecasting home prices three to six months ahead. In November 2021, Zillow wound it down , wrote down $304.4 million of homes it had bought above what it expected to sell them for, and cut about a quarter of its workforce. Its chief executive said "the unpredictability in forecasting home prices far exceeds what we anticipated."

Banks have it in writing. US bank regulators’ revised model risk guidance , issued in April 2026, calls for “effective challenge” from ”objective experts” and for outcomes analysis that" compares model outputs to corresponding real-world outcomes.” Healthcare has a model that proved itself first. Before an AI system was allowed to screen for diabetic eye disease with no doctor reading the images, researchers tested it on 900 patients at primary care clinics against bars set in advance: it had to catch more than 85% of the people who had the disease and correctly clear more than 82.5% of those who did not. It reached 87.2% and 90.7%, and in April 2018 it became the first autonomous AI diagnostic system the FDA authorized in any field of medicine.

The European Space Agency did the same thing in orbit. Its Φ-sat-1 satellite kept every image and recorded its AI’s cloud decision for each one, so people could check the AI before letting it discard anything. An onboard AI should earn trust the same way. Run it in shadow mode first: it makes its calls silently next to the crew and is scored against what people decided. It takes over only when the score holds.

The Crossover Already Happened

NASA Ames first applied the Kalman filter to navigate Apollo. Google researchers built it, decades later, into a tool that measures "the causal effect of an online advertising campaign on search-related site visits." The ideas have flowed from space to business before. Now they need to flow back.

What Any Business Holding Personal Data Can Do Now

  • Judge each person against their own baseline.
  • Stamp every dataset and report with how fresh it is, and a rule for acting before fresher data arrives.
  • Run two tiers of fraud detection, cheap and fast first, then slow and corroborated.
  • Declare where every record came from and who touched it.
  • Carry consent with the data, and re-ask when the use changes.
  • Run every new model in shadow mode, with as little data as possible, before it decides anything.

Apple makes advertisers wait a day or two to learn what their ads did. NASA plans for a crew to wait up to 22 minutes to hear from mission control. Business has been rehearsing for Mars for years, and the list above is what the rehearsal taught.