Hypothesis 2

A Falling Intervention Rate Can Be Manufactured Rather Than Earned

Intervention rate is useful but incomplete.

VerdictHolds in part

Georgian + NewtonX Physical AI Survey, Wave 1 · Reporting base: Buyers, n=85 unless a benchmark states otherwise · July 24 to Aug. 21, 2026

What we said

Intervention rate is useful but incomplete. A company can reduce interventions by narrowing the operating domain, slowing throughput, or moving edge cases to human supervisors, rather than by improving the underlying autonomy. If our hypothesis is right, buyers whose intervention rates have fallen should be able to name those mechanisms in their own deployments.

How we tested it

To test our hypothesis, we looked at four measures:

  • What buyers with falling intervention rates say contributed to such decrease.
  • How confident they are that the fall in intervention rates reflects real autonomy.
  • Whether the range of tasks and conditions the system handles widened or narrowed over the same period.
  • Whether buyers accept the general principle that demonstrations are not a reliable indicator of production readiness.

What we found

Of the 65 buyers whose intervention fell, 88% had narrowed what the system attempted, 75% had started routing harder cases to people first and 71% had slowed the system down for reliability. Buyers also accept the general point that intervention rates are useful but incomplete: 69% agree a demonstration does not tell you a system is ready for production. But 89% are confident their own drop reflects real autonomy rather than a smaller job, and nobody in any group chose “not at all confident.”

What Contributed to Falling Intervention?

88%

of the robotics buyers whose human intervention had fallen name narrowing the range of tasks the system attempts among the three things that contributed most to falling intervention rates

The 65 buyers who reported human intervention falling over the past 12 months, out of 85 · Buyers ranked up to three, so the shares add to more than 100%

Data for What Contributed to Falling Intervention?
CategoryIn top three (%)
We narrowed the range of tasks / conditions the system attempts88%
We route harder or unusual cases to people before the system tries them75%
We slowed things down to make the system more reliable71%
We added more human supervision around the edges63%
Other3%
0%20%40%60%80%100%We narrowed the range of tasks / conditionsthe system attemptsWe narrowed the range of tasks / conditions the system attempts — In top three (%): 88%88%We route harder or unusual cases to peoplebefore the system tries themWe route harder or unusual cases to people before the system tries them — In top three (%): 75%75%We slowed things down to make the systemmore reliableWe slowed things down to make the system more reliable — In top three (%): 71%71%We added more human supervision around theedgesWe added more human supervision around the edges — In top three (%): 63%63%OtherOther — In top three (%): 3%3%

The 65 buyers who reported human intervention falling over the past 12 months, out of 85 · Buyers ranked up to three, so the shares add to more than 100%

See the full benchmark
Confidence the Reduction Is Real

89%

of the robotics buyers whose intervention rate had fallen are confident the fall reflects a genuine improvement in the system’s autonomy, rather than a narrower set of tasks or more human handling

The same 65 buyers who reported intervention falling · single-select, so the shares add to 100%

Data for Confidence the Reduction Is Real
Category% of the 65
Very confident23%
Somewhat confident66%
Not very confident11%
Not at all confident0%
0%20%40%60%80%Very confidentVery confident — % of the 65: 23%23%Somewhat confidentSomewhat confident — % of the 65: 66%66%Not very confidentNot very confident — % of the 65: 11%11%Not at all confidentNot at all confident — % of the 65: 0%0%

The same 65 buyers who reported intervention falling · single-select, so the shares add to 100%

See the full benchmark
Whether the Operating Envelope Widened or Narrowed

69%

of robotics buyers say the range of tasks and conditions the system handles has widened over the past 12 months

All qualified buyers, n=85 · single-select, so the shares add to 100%

Data for Whether the Operating Envelope Widened or Narrowed
Category% of buyers
It handles a materially broader range of tasks / conditions12%
It handles a slightly broader range58%
About the same24%
It has narrowed slightly4%
It has narrowed materially0%
Too early to tell4%
0%20%40%60%It handles a materially broader range oftasks / conditionsIt handles a materially broader range of tasks / conditions — % of buyers: 12%12%It handles a slightly broader rangeIt handles a slightly broader range — % of buyers: 58%58%About the sameAbout the same — % of buyers: 24%24%It has narrowed slightlyIt has narrowed slightly — % of buyers: 4%4%It has narrowed materiallyIt has narrowed materially — % of buyers: 0%0%Too early to tellToo early to tell — % of buyers: 4%4%

All qualified buyers, n=85 · single-select, so the shares add to 100%

See the full benchmark
Demonstrations as an Indicator of Production Readiness

69%

of robotics buyers agree that demonstrations are not a reliable indicator of production readiness

All qualified buyers, n=85 · single-select, so the shares add to 100%

Data for Demonstrations as an Indicator of Production Readiness
Strongly agreeSomewhat agreeNeither agree nor disagreeSomewhat disagreeStrongly disagree
All buyers27%42%16%13%1%
All buyers27%42%16%13%0%100%
Strongly agreeSomewhat agreeNeither agree nor disagreeSomewhat disagreeStrongly disagree

All qualified buyers, n=85 · single-select, so the shares add to 100%

See the full benchmark