Developing excellent workflows and crafting the perfect prompts to achieve the results you want are excellent ways to spend time. You’re training yourself to think about the outcomes and the correct sequence to follow to actually achieve the goal, and so the so-called “vibe coding” benefits tremendously from any programming languages that you know, no matter if it’s Visual Basic, C++, R, Python, or even old format-unfriendly coding languages such as Fortran.
Podcast link: https://youtu.be/CYOhWZPyl1k?si=HLQN5m6MxAPpnSLd
Platforms like Spotfire have made it easy to integrate the LLM so that you can manipulate geospatial layers and then layer in more complex tasks such as well planning, supply chain, construction staging, and more.
The use cases can span vast subjects and domains, and the potential for successfully building complex processes that provide automation, process integration, and quality assurance is quite simply transformative.
However, there’s a snag.
That snag is data.
The second snag is the deployment of support functions that actually do things with your data.
It’s impossible to achieve the full potential of the AI models – whether generative AI with Large Language Models, small language models, and foundational models – or whether they are using amped-up neural networks in machine learning or predictive, in the case of deep learning (building in uncertainty and popularly incorporating Bayesian algorithms and stuff like Random Forest). Why is it impossible – the answer is – and let’s go back to it
DATA CHAOS.
The big behind-the-scenes problem – the minotaur in the labyrinth ready to devour you just as you think you’ve successfully navigated the corridors, halls, and blind alleys – is the fact that if your data is not cleaned, organized, harmonized, and structured in ways that the algorithms and large language models can handle them, you’ll fail. And, you could fail catastrophically.
Petrabytes developed a “smart ingestion” functionality that allowed for rapid ingestion of up to petabytes of data, with the ability to sort and assign the data so the data is perfectly conditioned for higher-level applications such as seismic interpretation. I’m oversimplifying things, but I’m not exaggerating when I say that Petrabytes saved months of time and countless nervous breakdowns.
APP MADNESS
Here’s where it really breaks down. Agentic AI is great, except when the apps are deployed in the wrong order, or are actually better suited for other domains or subjects. Further, how compatible are the apps? Has anyone really checked? And, in what sequence do you do the data QC and how many times? Someone or something needs to be able to coordinate it – be the air traffic controller who uses technology but also big-picture domain insights.
ENTER AMAZON QUICK
Amazon Quick has been around for a while as "Quick Suite," promising to be an integrator and accelerator. It could function that way, but the early iterations were complicated and the commercial model took some teasing out.
Quick Suite went live on October 9, 2025 as an evolution of QuickSight, and AWS shortened the name to Amazon Quick in 2026. It bundles six components — Quick Sight, Quick Flows, Quick Automate, Quick Index, Quick Research, and Apps in Quick — behind a single chat interface, with agents grounded in "spaces" that combine documents, dashboards, and connected data. It reaches internal repositories, wikis and intranets, AWS services like S3 and Redshift, and — through MCP — well over a thousand applications. As of August 2026 the agentic capabilities are also available in AWS GovCloud under FedRAMP authorization, which matters for anyone working state data, federal minerals, or regulated environmental workflows.
 |
| Amazon QUICK |
The interesting shift isn't the feature list. It's the posture. Quick doesn't ask you to finish cleaning before you start. It is built to go into the mess — on the working assumption that the more chaotic the pile, the more diamonds are buried in it. The integration is the product.
It sounds so easy that it frightens people who have thirty-five years of hard-won judgment and a suspicion that they are being told their expertise was the bottleneck.
It wasn't. But the expertise has to relocate.
EXAMPLE ONE: THE WAREHOUSE
Start with the most physical version of data chaos: a metal building full of banker's boxes. Paper logs, some scanned at 200 dpi in 2003 onto a server nobody maintains, some still on the original sepia. Handwritten annotations. Photocopies of photocopies.
This has always been treated as archival sentiment rather than an asset. By one recent estimate, roughly 60% of the world's producing wells have log data trapped behind raster scans that nobody has had the time, money, or patience to digitize by hand — and those aren't dead wells. They're producing, or adjacent to wells under reactivation review, or they hold the stratigraphic context for an infill decision being made this quarter.
What's changed is that vision-language models now read logs the way a petrophysicist reads them: holistically, header and tracks and curves together. A 2025 ADIPEC paper described a pipeline combining computer-vision segmentation to localize tracks, headers, and curve line styles with visual language models tuned to extract log names, units, curve limits, and depth scales — robust across varied log styles and low-quality or handwritten scans, and, deployed in the cloud, completing in minutes what used to take weeks, feeding straight into OSDU.
Now chain it:1. Ingest and condition. Smart ingestion classifies documents, extracts metadata, resolves mnemonics, reconciles depth references, and flags what it can't resolve rather than guessing.
2. Normalize into the standard. LAS out, OSDU records in, lineage preserved, with the original raster still linked so a human can check the machine's reading against the sepia.
3. Quality-control the QC. A second agent audits the first — cross-checking digitized curves against offset wells, catching depth shifts and unit inversions. This is the loop most pilots skip.
4. Feed the big model. Once the curves exist as data, a seismic foundation model has well control. One such model, pretrained across 192 global surveys, showed robust transfer to downstream interpretation tasks including geobody segmentation under limited labeled data; the NCS-model published in March 2026 was trained on Norway's public data repository; and in August 2026 TGS was awarded a multi-year contract by an international oil company to deploy its Seismic Foundation Model — trained on TGS's global multi-client library — for structural and stratigraphic interpretation. Foundation models have moved from papers to purchase orders.
5. Interpret, then argue with it. The geoscientist reviews, disputes, and re-prompts.
The warehouse stops being a liability on a facilities budget and becomes training data with regional specificity nobody else has. That is the actual competitive asset in the room — not the model, which everyone can rent.
EXAMPLE TWO: THE ORCHESTRATED PROSPECT REVIEW
Second example, harder. You want a screening-level evaluation across a multi-county position: seismic volumes, several hundred wells of varying vintage, production histories, land and lease status, regulatory filings, and a set of economic thresholds.
Under the old model, this is a six-week effort across four people and five applications, with the integration done in PowerPoint at the end — which is to say, done in a medium that cannot be tested.
Under an orchestrated model, an agent framework runs it as a directed sequence: pull and condition, interpret structure and stratigraphy against the foundation model, tie to well control, populate the property model, push to simulation, run the economics against your thresholds, and return a ranked list with an uncertainty band and a full audit trail of which step touched which record.
The value isn't the speed, although the speed is startling. It's that the reasoning chain is inspectable. You can ask why prospect nine ranked above prospect two and get an answer that points at specific data rather than a recollection of a meeting.
Which is why the controller has to be a geoscientist.
THE RELOCATION OF EXPERTISE
Here is the part that should encourage anyone worried about the Dark Ages accusation.
The 2026 GETI report found only about 45% of oil and gas professionals currently use AI in their work — a sharp rise, but still trailing other industries, while Deloitte's 2026 outlook puts AI and generative AI at under 20% of total IT spending among US oil and gas companies, projected to pass 50% by 2029. The spend is coming. The judgment is not.
What agentic systems cannot supply is the thing thirty years in a basin gives you: the instinct that a number is wrong before you can say why. The knowledge that a particular operator's 1981 logs always ran two feet shallow. The recognition that a beautiful automated fault interpretation is following a processing artifact.
Air traffic controllers don't fly the planes. They also don't get replaced by better planes.
And dreams
There's a reason the useful metaphor here is the labyrinth and not the assembly line.
The dream comes first. Someone has to imagine the workflow before anyone can build it — has to picture the warehouse emptied into a model, the six-week review compressed to an afternoon, the orphan well that turns out to be a storage asset. That imaginative act is not decoration on top of the engineering. It is the specification. Everything downstream — the ingestion, the harmonization, the sequencing, the QC loops — is just the discipline required to make the dream survive contact with the data.
Which is the real inversion in all of this. We used to think the data was the solid part and the vision was the soft part. It's the reverse. The data is a rumor until someone conditions it. The dream is the only thing in the building with structural integrity.
Feed the minotaur good data, and it turns out to be a very fast colleague.
Link to a video: https://youtu.be/CYOhWZPyl1k?si=HLQN5m6MxAPpnSLd
