Characterising the subsurface for a permit, whether for a well, a storage site or a geothermal project, takes a specialist team weeks. Very little of that is computation. One specialist deconvolves waveforms into receiver functions, another inverts those into a velocity model, a third writes the interpretation, and a reviewer signs it off. Each hand-off adds queue time, context loss and rework, and the queue is the part that costs weeks.
That is a scheduling problem wearing a geophysics costume, which is why the fix is organisational rather than geophysical. Every processing step here is long established. What changes is that each one is wrapped as an agent, the agents are connected without human hand-offs, and two safeguards are added so that speed does not quietly become exposure.
Four stages, and one person
The pipeline runs left to right. Raw seismic traces enter. A receiver-function agent deconvolves them. An inversion agent turns those into a velocity and subsurface model. A synthesis agent assembles the model into a drafted, permit-oriented report. Four stages, no hand-offs, and one human touchpoint at the end.
The two safeguards sit on either side of that person. An automated critic inspects intermediate outputs, the residuals, the misfit, the physical plausibility, and loops a flagged run back to the stage that produced it. It catches errors before a human ever sees them. The human review gate then holds final sign-off: work that clears the automated checks with high confidence can release, and the rest is escalated for judgment.
Both are switchable, which is the useful part, because switching them is how you find out what the headline number is actually counting.
Of 100 reports, 22 reach a person and 78 release on the automated checks alone. Turning the critic off moves 12 more reports in every hundred onto the human queue, which is 55 percent more human work for the same hundred reports.
The pipeline automates 4 stages in order: waveforms, receiver functions, inversion, synthesis. The gate is the only human touchpoint in it.
| configuration | auto-approved | to a human | released unreviewed | autonomy reads |
|---|---|---|---|---|
| critic on, gate on (shown) | 78 | 22 | 0 | 78% |
| critic off, gate on | 66 | 34 | 0 | 66% |
| critic on, gate off | 100 | 0 | 22 | 100% |
| critic off, gate off | 100 | 0 | 34 | 100% |
The number that improves when you remove the reviewer
Set the exhibit to both safeguards on. 78 reports in 100 release on the automated checks, and 22 stand up in amber as the share that reaches a person. That configuration reads 78% autonomous.
Now turn the human gate off. The amber ridge does not shrink because the work improved. It turns red and lies flat, because the reports that would have gone to a person are released instead, and the readout climbs to 100%. Nothing about any report changed. No model was retrained, no misfit fell, no residual improved. The only thing that moved is that the reviewer is gone.
This is worth saying plainly to anyone who is asked to manage against an autonomy target. Autonomy as a percentage is a ratio whose denominator includes the safeguard, so it can always be improved by deleting the safeguard. The configuration that reads highest on this exhibit is the one with no accountability in it at all, which makes the reading useless on its own and actively misleading as a target. What the number needs beside it is the count of reports released without the review they would otherwise have had, which is the red population, and which every version of this pipeline can report and most dashboards do not.
Stages automated
Human touchpoints
Reports reaching a person
Same, with the critic off
What the critic actually buys
The second control is quieter and more interesting. Turn the critic off and leave the gate on. The autonomy reading falls from 78 to 66, and the escalated population grows from 22 to 34.
12 more reports in every 100 now reach a person. Measured against the human queue rather than against the headline, that is 55% more review work for the same 100 reports and the same four processing stages. The critic did not make the pipeline more accurate in any way the reader can see on this exhibit. It made the errors surface early enough to be re-computed rather than escalated, and its return shows up as work a person no longer has to do.
That distinction matters for how the thing is justified. An automated critic sold as an accuracy improvement invites a comparison it will usually lose, because the honest answer is that a flagged run is re-computed and ends up where a careful specialist would have put it anyway. Sold as a load reduction, it is measurable, and the measurement is the size of the queue in front of the one person in the loop.
What these numbers are, and are not
Every share above is illustrative. They come from our own research note and they were chosen to show the mechanism, not drawn from a live deployment. Real rates depend on data quality, on geological complexity, and above all on how conservatively the gate is tuned: a cautious operator sets the auto-approve bar high, accepts a larger human share, and should.
Two further limits are worth stating rather than discovering later. The critic is only as good as the checks it runs, so edge cases it does not test still need human eyes, and a pipeline that reports a falling escalation rate may be reporting a critic that has stopped looking rather than a process that has improved. And permit-ready describes a report that meets format and defensibility standards. It does not mean any regulator has pre-approved anything.
What survives all of that is the structure. Removing hand-offs is where the weeks go, the critic decides how much of the remainder lands on a person, and the gate decides whether anyone is accountable for what ships. Those are three separate decisions, and only the first one is about speed.
Key takeaways
- Weeks of turnaround in subsurface reporting are queue time between specialists, not computation, so the saving comes from removing hand-offs rather than from faster models.
- Autonomy expressed as a percentage can be improved by deleting the reviewer: the configuration with no human gate reads 100% while no report has changed.
- Report the count of documents released without the review they would otherwise have had, beside any autonomy figure.
- An automated critic pays in human load rather than in visible accuracy: turning it off moves 12 more reports in every 100 onto the review queue, which is 55% more work for the same output.
- A falling escalation rate is ambiguous evidence. It can mean the pipeline improved or that the critic stopped catching things.
- Permit-ready means the report meets format and defensibility standards, not that a regulator has pre-approved it.
References
- Receiver function (seismology), for the deconvolution step the first agent performs.
- Seismic inversion, for the stage that turns receiver functions into a subsurface model.
- Multi-agent system, for the orchestration pattern the pipeline uses.
- Pipeline structure adapted from the LitMod3D input and output flowchart (Fig 4.2) in a 2018 PhD thesis on integrated geophysical and petrological modelling.
- An illustrative scenario we published on moving retrains out of a queue, a composite rather than a client engagement: Agentic MLOps, from six-week retrains to overnight.
