Latest / Silicon & Shale Podcast / Episode 15: The week's E&P developments & applied AI

Transcript

Episode 15: The week's E&P developments & applied AI · Silicon & Shale Podcast

• 0:00
  SPEAKER_00: September 14th, 2026.
• 0:02
  SPEAKER_00: Episode 15.
• 0:03
  SPEAKER_00: And I've been chewing on this Venezuela thing all week, Pac, but we've got to stack beyond that.
• 0:08
  SPEAKER_00: GPU capacity is starting to trade like a commodity.
• 0:10
  SPEAKER_00: Oxy's got a chief of AI now, which is a sentence I wouldn't have believed five years ago.
• 0:14
  SPEAKER_00: And there's a whole conversation about why enterprise AI pilots keep dying on
• 0:17
  SPEAKER_00: the vine before they ever hit production.
• 0:19
  SPEAKER_01: That last one's close to my heart.
• 0:21
  SPEAKER_01: Honestly.
• 0:22
  SPEAKER_01: The pilot to production gap keeps getting chalked up to bad models or weak tech.
• 0:26
  SPEAKER_01: But the real failure is almost always organizational.
• 0:28
  SPEAKER_01: Nobody wants to hear that.
• 0:30
  SPEAKER_01: I also want to dig into what's happening with data center pushback in the Midwest
• 0:33
  SPEAKER_01: and what that means for where compute actually lands relative to energy supply.
• 0:37
  SPEAKER_01: Which, if you were listening last week, you know that's a thread we've been pulling on.
• 0:40
  SPEAKER_00: And we'll come back to Chevron in Venezuela, too,
• 0:42
  SPEAKER_00: because there's a real question buried in there about what happens
• 0:44
  SPEAKER_00: when you point modern reservoir tools at data that's barely holding together.
• 0:48
  SPEAKER_00: That's the actual operating problem nobody put on the board slide.
• 0:51
  SPEAKER_01: So let's start there.
• 0:53
  SPEAKER_01: Chevron's back in Venezuela.
• 0:54
  SPEAKER_01: How much are we actually talking?
• 0:56
  SPEAKER_01: Because I've seen numbers floating around that seem aspirational.
• 0:58
  SPEAKER_00: People keep throwing out 7,600,000 barrels a day by 2031, but none of that's confirmed.
• 1:04
  SPEAKER_00: What we actually know is they're in joint ventures trying
• 1:07
  SPEAKER_00: to restore existing production,
• 1:08
  SPEAKER_00: which is a very different sentence than Chevron bets 7 billion on Venezuela.
• 1:12
  SPEAKER_01: Restoring production versus building new capacity.
• 1:15
  SPEAKER_01: Those aren't even in the same zip code from a risk profile.
• 1:17
  SPEAKER_01: And meanwhile, Exxon's just standing there saying the whole country's uninvestable.
• 1:21
  SPEAKER_00: And Exxon doesn't say that lightly.
• 1:22
  SPEAKER_00: They'll go anywhere.
• 1:23
  SPEAKER_00: They've operated in places where the security briefing is longer than the geological survey.
• 1:27
  SPEAKER_00: So when they look at Venezuela and say, no thanks, that should give you pause.
• 1:31
  SPEAKER_01: So what's Chevron seeing that Exxon isn't?
• 1:34
  SPEAKER_01: Or is it just that Chevron's already sunk enough in that walking away feels worse
• 1:37
  SPEAKER_01: than doubling down?
• 1:38
  SPEAKER_00: Little of both, probably.
• 1:39
  SPEAKER_00: They've got legacy assets, relationships with PDVSA that go back decades.
• 1:43
  SPEAKER_00: But here's what nobody's talking about.
• 1:45
  SPEAKER_00: The data infrastructure down there.
• 1:47
  SPEAKER_00: I've talked to guys who've worked those fields.
• 1:49
  SPEAKER_00: The production records are a mess.
• 1:51
  SPEAKER_00: We're talking about intervals where nobody's sure which wells we're even producing.
• 1:54
  SPEAKER_01: So, you're trying to restore production to wells
• 1:56
  SPEAKER_01: where you might not have reliable decline curves, completion records, or...
• 1:59
  SPEAKER_00: Some of it, yeah.
• 2:00
  SPEAKER_00: You're essentially re-exploring assets you theoretically already own.
• 2:03
  SPEAKER_01: That's a modeling nightmare.
• 2:04
  SPEAKER_01: You can't run any kind of forecasting tool against production history that might be fictional.
• 2:08
  SPEAKER_01: And I don't mean somebody fudged the numbers on purpose.
• 2:10
  SPEAKER_01: I mean the reporting systems degraded along with everything else.
• 2:13
  SPEAKER_00: The kind of problem that looks small on a board slide and enormous
• 2:16
  SPEAKER_00: when your reservoir engineer is staring at a screen full of gaps.
• 2:19
  SPEAKER_00: But let's get into Oxy because the Bangert thing connects to all of this.
• 2:23
  SPEAKER_00: Patrick Bangert sitting in a seat called Chief of Artificial Intelligence.
• 2:26
  SPEAKER_00: Not Chief Data Officer, not VP of Analytics.
• 2:28
  SPEAKER_00: Chief of AI at an operator.
• 2:31
  SPEAKER_01: And not a small one.
• 2:32
  SPEAKER_01: Occidental is producing north of a million BOE a day across their portfolio.
• 2:36
  SPEAKER_00: What caught me was the framing.
• 2:36
  SPEAKER_00: Bangert's talking about compute power, data, and scientific innovation as the three legs.
• 2:40
  SPEAKER_00: Not, we bought a platform, or we're partnering with Microsoft.
• 3:01
  SPEAKER_01: He's describing an internal capability build, which,
• 3:03
  SPEAKER_01: if you've watched how most operators approach technology adoption,
• 3:06
  SPEAKER_01: that's a real departure.
• 3:07
  SPEAKER_00: It is, but I'll push back because I've seen this movie.
• 3:11
  SPEAKER_00: Operator creates a shiny new title,
• 3:13
  SPEAKER_00: hires someone impressive,
• 3:14
  SPEAKER_00: announces a transformation initiative,
• 3:16
  SPEAKER_00: and 18 months later that person's gone because they couldn't get
• 3:18
  SPEAKER_00: the field superintendents to change a single workflow.
• 3:21
  SPEAKER_01: That's the pilot purgatory problem, though, not a leadership problem.
• 3:24
  SPEAKER_01: And actually, Kwesi Coleman at Punt AI had a whole thing this week about exactly that.
• 3:29
  SPEAKER_01: Why Enterprise AI pilots fail?
• 3:31
  SPEAKER_01: His argument is the model almost never causes the failure.
• 3:34
  SPEAKER_01: The organization hasn't built the connective tissue between the experiment and
• 3:37
  SPEAKER_01: the operating workflow.
• 3:39
  SPEAKER_00: Which is a polite way of saying nobody told the completions engineer his job was
• 3:42
  SPEAKER_00: about to change.
• 3:43
  SPEAKER_01: It goes deeper than that.
• 3:45
  SPEAKER_01: Coleman's point is that you need expert judgment baked into the scaling process.
• 3:48
  SPEAKER_01: You can't just hand a pilot to IT and say, productionize this.
• 3:52
  SPEAKER_01: The domain experts,
• 3:53
  SPEAKER_01: your reservoir engineers,
• 3:54
  SPEAKER_01: your production techs,
• 3:55
  SPEAKER_01: they have to be in the room when you're deciding what the tool actually does
• 3:58
  SPEAKER_01: at scale.
• 3:59
  SPEAKER_01: Otherwise,
• 3:59
  SPEAKER_01: you build something that works perfectly on test data and falls apart
• 4:02
  SPEAKER_01: the first time it hits a well with an ESP that's been swapped three times.
• 4:06
  SPEAKER_00: The best field technology I've ever seen adopted was stuff where the tool guy or
• 4:09
  SPEAKER_00: the mudlogger or whoever was involved early enough
• 4:11
  SPEAKER_00: that they felt like it was theirs.
• 4:12
  SPEAKER_00: The worst was always something that showed up on a truck with a vendor rep and
• 4:15
  SPEAKER_00: a training manual nobody read.
• 4:17
  SPEAKER_01: So the question with Oxy is whether Bangert's position gives him the authority
• 4:20
  SPEAKER_01: to actually bridge that gap,
• 4:22
  SPEAKER_01: or whether he's just a very senior person with a very nice title
• 4:25
  SPEAKER_01: who still can't make a business unit change its completion design process.
• 4:28
  SPEAKER_00: That's the billion-dollar question.
• 4:30
  SPEAKER_00: And I don't mean that as a figure of speech.
• 4:32
  SPEAKER_00: Oxy's spending real capital on this.
• 4:34
  SPEAKER_00: You don't create a C-level AI role and then give them an intern budget.
• 4:37
  SPEAKER_01: No, and that's what makes it worth watching.
• 4:39
  SPEAKER_01: Because if Oxy actually pulls this off,
• 4:41
  SPEAKER_01: builds internal AI capability that touches drilling optimization,
• 4:44
  SPEAKER_01: reservoir modeling,
• 4:45
  SPEAKER_01: production forecasting,
• 4:47
  SPEAKER_01: that changes the competitive math
• 4:48
  SPEAKER_01: for every mid-cap E&P that's been telling themselves they'll get to AI next year.
• 4:52
  SPEAKER_00: Pac, they've been saying that since 2019.
• 4:54
  SPEAKER_01: Which is why Oxy doing it at the structural level matters.
• 4:57
  SPEAKER_01: Let's move past the conference talk stage.
• 4:59
  SPEAKER_00: Let me stay on this for a second, though, because Oxy's got assets everywhere.
• 5:02
  SPEAKER_00: Permian, DJ Basin, Gulf of Mexico, the whole carbon capture play.
• 5:06
  SPEAKER_00: The data environments across those basins aren't remotely standardized.
• 5:09
  SPEAKER_00: I've seen what happens when you try to run analytics across assets acquired
• 5:12
  SPEAKER_00: from three different operators who each had their own SCADA setup,
• 5:15
  SPEAKER_00: their own naming conventions.
• 5:16
  SPEAKER_00: Some of those production databases disagree with themselves.
• 5:19
  SPEAKER_01: That's the infrastructure problem that makes this role necessary.
• 5:22
  SPEAKER_01: I'll buy that.
• 5:23
  SPEAKER_01: But is the bottleneck really the AI capability or is it still the data plumbing?
• 5:25
  SPEAKER_01: Because every operator I talk to, the story's the same.
• 5:28
  SPEAKER_01: They've got a vendor showing them some incredible demo on clean synthetic data,
• 5:31
  SPEAKER_01: and then the pilot hits real production data and it's... A mess.
• 5:33
  SPEAKER_01: I know.
• 5:50
  SPEAKER_01: When's the last time you saw a vendor lead with,
• 5:51
  SPEAKER_01: we'll spend 18 months cleaning your data before anything interesting happens?
• 5:53
  SPEAKER_01: Never.
• 5:54
  SPEAKER_01: That pitch deck has a 0% close rate.
• 5:55
  SPEAKER_01: And yet that's the actual project.
• 5:56
  SPEAKER_01: Every time.
• 6:15
  SPEAKER_01: Okay, let me shift to the GPU story because this connects to
• 6:18
  SPEAKER_01: the cost problem underneath all of this.
• 6:20
  SPEAKER_01: The energy gang had a whole episode on GPU capacity,
• 6:23
  SPEAKER_01: actual access to compute for training and inference,
• 6:26
  SPEAKER_01: starting to look like a commodity market.
• 6:27
  SPEAKER_01: People are talking about benchmark pricing, hedging tools, future-style contracts.
• 6:32
  SPEAKER_00: Which is fascinating to me because I've watched commodity markets my entire career,
• 6:35
  SPEAKER_00: and they exist for a reason.
• 6:37
  SPEAKER_00: So we've basically recreated the natural gas market but for silicon.
• 7:06
  SPEAKER_00: I know, that one hurt coming out.
• 7:07
  SPEAKER_00: Moving on.
• 7:08
  SPEAKER_01: Kind of?
• 7:10
  SPEAKER_01: Except natural gas has pipelines and storage and decades of contract law.
• 7:13
  SPEAKER_01: GPU capacity doesn't have any of that infrastructure yet.
• 7:16
  SPEAKER_01: Which is why the pricing is so chaotic right now.
• 7:18
  SPEAKER_00: Here's what I want to know.
• 7:19
  SPEAKER_00: Who's the counterparty?
• 7:21
  SPEAKER_00: In oil and gas, you've got producers on one side and refiners or utilities on the other.
• 7:25
  SPEAKER_00: Both have genuine economic exposure to the price.
• 7:28
  SPEAKER_00: In GPU futures, who's actually hedging?
• 7:31
  SPEAKER_00: Is it the data center operators, the AI companies, the cloud providers?
• 7:34
  SPEAKER_01: All of them, potentially.
• 7:36
  SPEAKER_01: A data center developer who's committed to delivering a certain amount of compute to
• 7:38
  SPEAKER_01: a customer in 18 months has real price risk if GPU costs move against them.
• 7:43
  SPEAKER_01: An AI company that needs to train a model on a fixed budget has the same exposure.
• 7:47
  SPEAKER_01: So the hedging need is real.
• 7:48
  SPEAKER_01: The question is whether the market structure can get standardized enough
• 7:50
  SPEAKER_01: for actual price discovery.
• 7:52
  SPEAKER_00: Because right now it's all bilateral, handshake deals and NDAs.
• 7:55
  SPEAKER_01: Which is where oil and gas was before NYMEX standardized crude contracts.
• 7:59
  SPEAKER_01: You had refiners calling producers directly, negotiating spot prices that nobody else could see.
• 8:04
  SPEAKER_00: And we know how long that transition took.
• 8:06
  SPEAKER_00: Decades.
• 8:07
  SPEAKER_01: But everything moves faster now.
• 8:08
  SPEAKER_01: The financial infrastructure for new commodity classes?
• 8:11
  SPEAKER_01: You've got crypto exchanges that stood up derivative markets in months, not decades.
• 8:15
  SPEAKER_01: Whether you like crypto or not, the plumbing exists.
• 8:18
  SPEAKER_00: I don't love the comparison to crypto exchanges, given how many of those blew up.
• 8:22
  SPEAKER_01: The mechanism is the point, not the track record.
• 8:24
  SPEAKER_01: You don't need to build a physical exchange floor anymore.
• 8:27
  SPEAKER_01: You need standardized contracts,
• 8:28
  SPEAKER_01: a clearing mechanism, and enough liquidity that the spread doesn't eat you alive.
• 8:32
  SPEAKER_00: Which means it shows up in price.
• 8:58
  SPEAKER_01: Eventually.
• 8:59
  SPEAKER_01: If this market actually develops, a data center ban in rural Ohio isn't just a NIMBY story.
• 9:04
  SPEAKER_01: It's a supply disruption that moves the forward curve.
• 9:06
  SPEAKER_00: And suddenly,
• 9:07
  SPEAKER_00: every county commissioner in America is accidentally setting supply policy for
• 9:10
  SPEAKER_00: a global compute market.
• 9:12
  SPEAKER_01: The directional logic holds.
• 9:13
  SPEAKER_01: Physical constraints on compute infrastructure create scarcity.
• 9:16
  SPEAKER_01: Scarcity creates price signals.
• 9:18
  SPEAKER_01: Price signals attract capital to wherever the constraints are loosest.
• 9:21
  SPEAKER_01: Which, by the way, is how the Permian became the Permian.
• 9:24
  SPEAKER_01: Friendly regulators,
• 9:25
  SPEAKER_01: available land,
• 9:26
  SPEAKER_01: existing infrastructure—
• 9:27
  SPEAKER_00: So West Texas might end up hosting the GPU farms too.
• 9:30
  SPEAKER_01: Amazon's already building private gas infrastructure out there.
• 9:33
  SPEAKER_01: We covered that a few weeks back.
• 9:35
  SPEAKER_01: If you've got cheap molecules,
• 9:36
  SPEAKER_01: available land,
• 9:37
  SPEAKER_01: and a regulatory environment that says yes, the economics point in one direction.
• 9:41
  SPEAKER_00: Until the molecules going to data centers are
• 9:43
  SPEAKER_00: the same molecules your gas plant needs for processing,
• 9:45
  SPEAKER_00: and now you've got a basis differential problem that didn't exist six months ago,
• 9:49
  SPEAKER_00: but we should save that for the collision zone.
• 9:51
  SPEAKER_01: One more thing on the GPU trading angle.
• 9:53
  SPEAKER_01: Unless GPU futures actually materialize.
• 9:55
  SPEAKER_01: But today?
• 10:13
  SPEAKER_01: Today you're a price taker.
• 10:15
  SPEAKER_01: You signed a cloud contract at a certain rate and you're hoping
• 10:17
  SPEAKER_01: that rate doesn't triple when demand spikes.
• 10:19
  SPEAKER_01: Which for an industry that hedges everything else.
• 10:21
  SPEAKER_01: Crude, gas, diesel, interest rates.
• 10:24
  SPEAKER_01: The idea
• 10:25
  SPEAKER_01: that your AI compute costs are just floating out there unhedged should make your CFO
• 10:28
  SPEAKER_01: uncomfortable.
• 10:29
  SPEAKER_00: I bet most CFOs don't even know those costs exist as a separate line item yet.
• 10:33
  SPEAKER_00: They're buried in IT somewhere between the SAP license and the printer lease.
• 10:37
  SPEAKER_01: Between the SAP license and the printer lease.
• 10:39
  SPEAKER_01: The future of reservoir modeling lives right there.
• 10:41
  SPEAKER_00: And people wonder why pilots don't make it to production.
• 10:44
  SPEAKER_01: So, related to this cost and capability question,
• 10:46
  SPEAKER_01: there's been a bunch of conversation this week about AI agents
• 10:48
  SPEAKER_01: that can actually use computers.
• 10:50
  SPEAKER_01: Not just answer questions or generate text,
• 10:52
  SPEAKER_01: but navigate software, click through interfaces, take actions.
• 10:55
  SPEAKER_00: Meaning what, specifically?
• 10:57
  SPEAKER_00: An AI that can log into my production reporting system and pull a decline curve?
• 11:00
  SPEAKER_01: In theory, yes.
• 11:02
  SPEAKER_01: The leading AI labs have been demonstrating agents with computer use features.
• 11:06
  SPEAKER_01: Tools that can navigate a browser,
• 11:08
  SPEAKER_01: interact with application interfaces, enter data, pull reports.
• 11:11
  SPEAKER_01: You tell the agent what you need and it goes and does it
• 11:13
  SPEAKER_01: across whatever applications are involved.
• 11:15
  SPEAKER_00: The number of people in this industry who can't get their own single sign-on to work
• 11:18
  SPEAKER_00: on a Tuesday morning, and now we're going to hand that workflow to an AI agent.
• 11:22
  SPEAKER_01: That's a real objection, but not the one that keeps me up at night.
• 11:25
  SPEAKER_01: The authentication piece is solvable.
• 11:27
  SPEAKER_01: What's harder is trust boundaries.
• 11:29
  SPEAKER_01: If an agent can click through your SCADA system
• 11:30
  SPEAKER_01: and your production accounting software and your land management platform,
• 11:33
  SPEAKER_01: what stops it from taking an action in one system that's correct in isolation
• 11:36
  SPEAKER_01: but catastrophic in context?
• 11:38
  SPEAKER_01: Like,
• 11:38
  SPEAKER_01: an agent that sees a well underperforming and submits a workover request because
• 11:41
  SPEAKER_01: the numbers support it,
• 11:43
  SPEAKER_01: but doesn't know there's a title dispute on that lease that
• 11:44
  SPEAKER_01: the LAN team's been sitting on for six weeks.
• 11:47
  SPEAKER_00: That's not a hypothetical.
• 11:48
  SPEAKER_00: I've seen humans do clearly that.
• 11:50
  SPEAKER_01: Which is why the agent version is terrifying.
• 11:52
  SPEAKER_01: At least when a human does it,
• 11:53
  SPEAKER_01: someone in the hallway catches them and says, hey, hold off on that one.
• 11:57
  SPEAKER_01: The agent doesn't have a hallway.
• 11:58
  SPEAKER_00: It doesn't have a hallway.
• 12:00
  SPEAKER_00: But here's where I'll push you.
• 12:02
  SPEAKER_00: Last week we were talking about agent security and trust architecture,
• 12:05
  SPEAKER_00: and the answer was basically, we don't have good frameworks yet.
• 12:08
  SPEAKER_00: Has anything actually changed or are we just describing the same problem with a fancier demo?
• 12:13
  SPEAKER_01: Honestly, the demos got better.
• 12:15
  SPEAKER_01: The frameworks didn't.
• 12:16
  SPEAKER_01: And that gap is widening, which is the part that worries me.
• 12:19
  SPEAKER_01: Every month the agents get more capable and every month
• 12:21
  SPEAKER_01: the governance layer is still a whiteboard sketch in someone's architecture meeting.
• 12:24
  SPEAKER_00: Faster cars, no brakes.
• 12:26
  SPEAKER_01: And the brake vendor is still in stealth mode.
• 12:28
  SPEAKER_01: There's a practical angle here for operators specifically though.
• 12:30
  SPEAKER_01: Going back to Coleman's point about why enterprise AI pilots fail,
• 12:34
  SPEAKER_01: companies spin up a pilot,
• 12:35
  SPEAKER_01: it works great in the sandbox,
• 12:36
  SPEAKER_01: and then scaling it requires changing three workflows,
• 12:39
  SPEAKER_01: getting buy-in from two VPs who don't talk to each other,
• 12:41
  SPEAKER_01: and integrating with a system that was last updated during the Obama administration.
• 12:44
  SPEAKER_00: The pilot works.
• 12:45
  SPEAKER_00: The pilot always works.
• 12:47
  SPEAKER_00: It's designed to work.
• 12:48
  SPEAKER_00: And then you try to run it in the actual operating environment and it turns out
• 12:51
  SPEAKER_00: the success criteria for the pilot had nothing to do with the success criteria for
• 12:54
  SPEAKER_00: the business.
• 12:56
  SPEAKER_01: So here's where I want to disagree with you.
• 12:58
  SPEAKER_01: Because I think you're implying that the pilot to production gap is fundamentally
• 13:01
  SPEAKER_01: about organizational dysfunction.
• 13:02
  SPEAKER_01: And I don't think that's the whole story.
• 13:04
  SPEAKER_00: What else is it?
• 13:05
  SPEAKER_01: Most pilots are scoped to prove the technology works.
• 13:08
  SPEAKER_01: They're not scoped to prove
• 13:09
  SPEAKER_01: the technology works inside your existing decision rights,
• 13:11
  SPEAKER_01: your approval chains, your data refresh cadences.
• 13:14
  SPEAKER_01: If you design the pilot from day one to operate under production constraints,
• 13:17
  SPEAKER_01: real data latency,
• 13:18
  SPEAKER_01: real permission structures,
• 13:19
  SPEAKER_01: real exception handling,
• 13:20
  SPEAKER_01: you'd kill half the pilots earlier,
• 13:22
  SPEAKER_01: which is fine, and the ones that survive would actually scale.
• 13:25
  SPEAKER_00: But nobody does that because it's slower,
• 13:27
  SPEAKER_00: it's less impressive,
• 13:28
  SPEAKER_00: and the vendor doesn't want to show the board a demo that includes
• 13:30
  SPEAKER_00: a four-hour data lag and a manual override step.
• 13:33
  SPEAKER_01: Which is why internal teams should own the pilot design, not the vendor.
• 13:37
  SPEAKER_01: The vendor's incentive is to make the demo look good.
• 13:40
  SPEAKER_01: Your incentive is to make the workflow survive contact with your actual operation.
• 13:44
  SPEAKER_01: Those incentives don't align.
• 13:45
  SPEAKER_00: I'll grant you that the design matters,
• 13:47
  SPEAKER_00: but I've watched companies spend a year designing the perfect pilot with all
• 13:50
  SPEAKER_00: the right constraints,
• 13:51
  SPEAKER_00: and it still dies because the asset team that's supposed
• 13:53
  SPEAKER_00: to adopt it doesn't trust it.
• 13:55
  SPEAKER_00: The geologist who's been doing this for 20 years doesn't want an AI telling her
• 13:58
  SPEAKER_00: which wells to look at.
• 13:59
  SPEAKER_00: Not because the AI's wrong,
• 14:01
  SPEAKER_00: because nobody proved to her that it's right in a way she finds credible.
• 14:04
  SPEAKER_01: And that's Coleman's point, too.
• 14:05
  SPEAKER_01: Expert judgment has to be in the loop,
• 14:07
  SPEAKER_01: not as a rubber stamp, but as a genuine feedback mechanism that improves the system.
• 14:11
  SPEAKER_01: The pilots that work are the ones where the domain expert feels like a co-designer,
• 14:15
  SPEAKER_01: not a test subject.
• 14:16
  SPEAKER_00: So the practical conclusion,
• 14:17
  SPEAKER_00: you need the technology design right and you need the human adoption design right,
• 14:20
  SPEAKER_00: and neither one compensates for the other.
• 14:22
  SPEAKER_01: You can't engineer your way around organizational trust,
• 14:25
  SPEAKER_01: and you can't culture change your way around bad data architecture.
• 14:28
  SPEAKER_01: You need both,
• 14:28
  SPEAKER_01: which is why the Oxy model of having someone who owns the full stack,
• 14:31
  SPEAKER_01: technical and organizational, might be the right structure.
• 14:34
  SPEAKER_00: Might be.
• 14:35
  SPEAKER_00: If Bengard actually has the authority to tell a business unit
• 14:37
  SPEAKER_00: to change its workflow, at that point these roles usually stall out.
• 14:41
  SPEAKER_01: One more thing on the computer use agents before we move on.
• 14:44
  SPEAKER_01: I caught myself this week sketching out how you could chain one of these agents
• 14:47
  SPEAKER_01: through a full AFE approval workflow.
• 14:49
  SPEAKER_01: Have it pull the cost estimates,
• 14:50
  SPEAKER_01: check them against analogs, root for signatures, the whole thing.
• 14:53
  SPEAKER_01: And I got about 40 minutes into the design before I realized I'd automated a process
• 14:56
  SPEAKER_01: that currently takes two days and replaced it with something
• 14:59
  SPEAKER_01: that would take 15 minutes,
• 15:00
  SPEAKER_01: plus a three-week implementation to handle the 11 exception paths that
• 15:03
  SPEAKER_01: the current process handles by someone walking down the hall and asking a question.
• 15:07
  SPEAKER_00: So you saved a day and a half and created a month of work.
• 15:10
  SPEAKER_01: The math only works if you ignore the exceptions,
• 15:12
  SPEAKER_01: which is what demos do, which is why I should know better by now.
• 15:15
  SPEAKER_01: And yet here I am.
• 15:17
  SPEAKER_01: 40 minutes deep in a Miro board before the obvious hits me.
• 15:20
  SPEAKER_00: At least you caught it at the whiteboard stage and not after you'd built it.
• 15:23
  SPEAKER_01: Small mercies.
• 15:24
  SPEAKER_01: But the broader point stands.
• 15:25
  SPEAKER_01: These computer use agents are going to be transformative for repetitive,
• 15:28
  SPEAKER_01: well-defined workflows where the exception rate is low.
• 15:31
  SPEAKER_01: The mistake is pointing them at the messy stuff first because the pain is highest there.
• 15:34
  SPEAKER_01: High pain doesn't mean high tractability.
• 15:36
  SPEAKER_00: That should be stamped on every AI vendor's business card in Houston.
• 15:40
  SPEAKER_00: So here's where I keep landing this week, Venezuela.
• 15:42
  SPEAKER_01: We're back.
• 15:43
  SPEAKER_00: Not the geopolitics.
• 15:44
  SPEAKER_00: We've been over that.
• 15:46
  SPEAKER_00: I'm thinking about the data problem,
• 15:47
  SPEAKER_00: because Chevron's sitting on joint ventures where the production records are,
• 15:50
  SPEAKER_00: at best, inconsistent.
• 15:51
  SPEAKER_00: Decades of PDVSA operational chaos, regime changes, staff turnover.
• 15:56
  SPEAKER_00: And they're trying to restore existing production, not build greenfield.
• 15:59
  SPEAKER_00: Which means they need to understand what those wells did,
• 16:01
  SPEAKER_00: what the reservoirs look like now, and what the facilities can actually handle.
• 16:04
  SPEAKER_01: And you're saying that's an AI problem?
• 16:07
  SPEAKER_00: I'm saying it's a data problem that AI could help with if,
• 16:09
  SPEAKER_00: and this is a big if, you could get the inputs into any kind of usable shape.
• 16:13
  SPEAKER_00: Because what I've heard from guys who've worked those fields is
• 16:15
  SPEAKER_00: that you've got paper records in some places,
• 16:17
  SPEAKER_00: digital logs that contradict each other,
• 16:19
  SPEAKER_00: and production data that just stops for years at a time.
• 16:22
  SPEAKER_00: You don't fill that with a spreadsheet.
• 16:23
  SPEAKER_01: No, that's a reconstruction problem.
• 16:25
  SPEAKER_01: And some of the newer multimodal approaches get interesting here.
• 16:28
  SPEAKER_01: You're not just running structured data through a model.
• 16:30
  SPEAKER_01: You're pulling from scanned documents,
• 16:32
  SPEAKER_01: handwritten field notes, maybe satellite imagery to infer surface activity.
• 16:36
  SPEAKER_01: The question is whether you can stitch those together into something
• 16:38
  SPEAKER_01: an engineer would trust enough to make a completion decision on.
• 16:40
  SPEAKER_00: And that loops right back to the trust problem.
• 16:42
  SPEAKER_00: You've got a reservoir engineer in Houston who's never set foot in Zulia State,
• 16:46
  SPEAKER_00: looking at AI reconstructed production curves from data that's been
• 16:48
  SPEAKER_00: through three different national oil company regimes.
• 16:51
  SPEAKER_00: How confident is she?
• 16:53
  SPEAKER_01: Not very.
• 16:54
  SPEAKER_01: Not without validation.
• 16:55
  SPEAKER_01: But what's the alternative?
• 16:57
  SPEAKER_01: You send a team down there,
• 16:58
  SPEAKER_01: they spend six months trying to piece together the same records manually,
• 17:00
  SPEAKER_01: and you're still guessing.
• 17:02
  SPEAKER_01: At least with a model, you can flag confidence intervals.
• 17:04
  SPEAKER_01: You can say,
• 17:05
  SPEAKER_01: here's what we think this well did between 2014 and 2019,
• 17:08
  SPEAKER_01: and here's how uncertain we are about it.
• 17:11
  SPEAKER_01: That's more honest than a single-point estimate from someone eyeballing a half-complete log.
• 17:14
  SPEAKER_00: Maybe, but confidence intervals only help if
• 17:16
  SPEAKER_00: the person reading them understands what they mean operationally.
• 17:19
  SPEAKER_00: I've seen plenty of dashboards where the uncertainty band is right there on
• 17:22
  SPEAKER_00: the screen, and the decision-maker treats the midpoint as gospel anyway.
• 17:26
  SPEAKER_01: That's a training and workflow problem.
• 17:28
  SPEAKER_01: The tool can surface the uncertainty.
• 17:30
  SPEAKER_01: But if the decision process downstream doesn't account for it, you're right.
• 17:33
  SPEAKER_01: It doesn't matter.
• 17:34
  SPEAKER_00: Which makes it a people problem on top of everything else.
• 17:37
  SPEAKER_01: But think about who actually needs this.
• 17:39
  SPEAKER_01: Goes beyond Chevron in Venezuela.
• 17:41
  SPEAKER_01: Any operator doing A&D work on distressed assets or coming back into fields
• 17:44
  SPEAKER_01: that were shut in during a down cycle or inheriting production from an acquisition
• 17:47
  SPEAKER_01: where the seller's data room was a disaster, they've all got the same issue.
• 17:51
  SPEAKER_01: Incomplete historical data and decisions that need to happen faster than
• 17:54
  SPEAKER_01: a team can manually reconstruct the record.
• 17:56
  SPEAKER_00: That's a bigger market than people realize.
• 17:59
  SPEAKER_00: Every basin has orphaned data,
• 18:01
  SPEAKER_00: wells that changed hands four times, each time losing a layer of context.
• 18:04
  SPEAKER_00: And the guys buying those assets are doing their own decline curve analysis
• 18:07
  SPEAKER_00: on whatever's left,
• 18:08
  SPEAKER_00: which sometimes means they're building a type curve from 11 wells
• 18:11
  SPEAKER_00: when there should be 40 in the dataset.
• 18:13
  SPEAKER_01: And then you get burned.
• 18:14
  SPEAKER_01: Because if your type curve is built on survivors,
• 18:16
  SPEAKER_01: the wells that kept producing,
• 18:18
  SPEAKER_01: kept reporting, you're going to overestimate what the full inventory looks like.
• 18:22
  SPEAKER_01: The wells that went offline,
• 18:23
  SPEAKER_01: the ones with the messy records,
• 18:25
  SPEAKER_01: Those are the ones dragging the average down, and you don't even see them.
• 18:28
  SPEAKER_00: Survivorship bias in well performance, not new.
• 18:31
  SPEAKER_00: But you're saying AI could help correct for it by finding and incorporating the data
• 18:34
  SPEAKER_00: from the missing wells.
• 18:35
  SPEAKER_01: Or at least flagging that the data's missing.
• 18:38
  SPEAKER_01: Half the battle.
• 18:39
  SPEAKER_01: Someone pulls a data set,
• 18:40
  SPEAKER_01: runs their analysis,
• 18:41
  SPEAKER_01: and nobody asks,
• 18:42
  SPEAKER_01: hey, why are there only 11 wells here when the lease map shows 40 locations?
• 18:46
  SPEAKER_01: An automated system could at least raise the question.
• 18:48
  SPEAKER_00: If it knows to look at the lease map.
• 18:50
  SPEAKER_01: Which brings us back to the multimodal piece.
• 18:53
  SPEAKER_01: You need something that can cross-reference spatial data,
• 18:55
  SPEAKER_01: production records, regulatory filings, maybe even old scout tickets.
• 18:59
  SPEAKER_01: And the approval constraint is real.
• 19:01
  SPEAKER_01: Who signs off on the reconstructed data?
• 19:03
  SPEAKER_01: Because if you're using it in a reserve report, that's SEC territory.
• 19:07
  SPEAKER_01: You need a qualified reserves evaluator who's willing to put their name on inputs
• 19:10
  SPEAKER_01: that came out of a model.
• 19:11
  SPEAKER_00: I don't know a single reserves evaluator who'd sign that today.
• 19:14
  SPEAKER_01: Not today.
• 19:15
  SPEAKER_01: But think about what they sign now.
• 19:17
  SPEAKER_01: They already rely on engineering software that interpolates between data points.
• 19:20
  SPEAKER_01: They already accept analogs from offset wells.
• 19:23
  SPEAKER_01: The question is whether an AI-assisted reconstruction is fundamentally different
• 19:26
  SPEAKER_01: from those existing approximations or just a more sophisticated version
• 19:29
  SPEAKER_01: of what they're already doing.
• 19:31
  SPEAKER_00: Different because the engineer chose the analog.
• 19:33
  SPEAKER_00: She picked the offset well.
• 19:34
  SPEAKER_00: She made the judgment call about why it's comparable.
• 19:37
  SPEAKER_00: When the model does it,
• 19:38
  SPEAKER_00: the reasoning's opaque,
• 19:39
  SPEAKER_00: and you can't put the model thought these wells were similar in an SEC filing.
• 19:43
  SPEAKER_01: You could if the model showed its work.
• 19:45
  SPEAKER_01: The starting point is better, I'll give you that.
• 19:46
  SPEAKER_01: But there's a failure mode you're not addressing.
• 19:48
  SPEAKER_01: What happens when the model reconstructs a production history that looks plausible but is wrong?
• 19:53
  SPEAKER_01: Not obviously wrong, subtly wrong.
• 20:13
  SPEAKER_00: That's the scariest failure mode in any predictive system.
• 20:15
  SPEAKER_00: Plausible but wrong.
• 20:16
  SPEAKER_00: And the human in the loop isn't just a nice-to-have.
• 20:18
  SPEAKER_00: It's the regulatory requirement.
• 20:31
  SPEAKER_01: The useful outcome here isn't AI produces a reserve report.
• 20:34
  SPEAKER_01: The useful outcome is AI produces a draft reconstruction with uncertainty flags,
• 20:38
  SPEAKER_01: an engineer validates it against every independent data source available,
• 20:41
  SPEAKER_01: and the final product is the engineer's work product with better inputs
• 20:43
  SPEAKER_01: than she would have had otherwise.
• 20:45
  SPEAKER_00: Which means the ROI isn't in replacing the engineer, it's in...
• 20:48
  SPEAKER_01: Time to first decision.
• 20:49
  SPEAKER_01: Instead of six months reconstructing the record manually, you're at a validated draft in weeks.
• 20:53
  SPEAKER_01: And on a deal timeline, that matters.
• 20:56
  SPEAKER_01: If you're competing for a distressed asset package,
• 20:58
  SPEAKER_01: the bidder who understands the inventory faster has a real advantage.
• 21:01
  SPEAKER_00: Assuming the model doesn't confidently reconstruct a well history that's actually
• 21:04
  SPEAKER_00: from a different formation because somebody fat-fingered the API number in 2016.
• 21:08
  SPEAKER_01: Which brings us right back to the data quality issue.
• 21:11
  SPEAKER_01: Remember what I said about the AFE workflow?
• 21:13
  SPEAKER_01: 40 minutes on the whiteboard before the exception list hit me?
• 21:16
  SPEAKER_01: The number of ways a dataset can be quietly wrong in an acquired asset is
• 21:19
  SPEAKER_01: the same kind of problem.
• 21:21
  SPEAKER_01: Transposed digits, mislabeled formations, production allocated to the wrong lease.
• 21:24
  SPEAKER_00: At least you're consistent.
• 21:26
  SPEAKER_00: 40 minutes seems to be your threshold before reality intervenes.
• 21:29
  SPEAKER_01: That I can get behind.
• 21:30
  SPEAKER_01: Providence on every input.
• 21:31
  SPEAKER_01: Because right now she's trusting a spreadsheet that somebody emailed her
• 21:35
  SPEAKER_01: and she has no idea who touched it last or what assumptions went into
• 21:39
  SPEAKER_01: the decline curve someone else built in 2019.
• 21:56
  SPEAKER_01: So the tool isn't asking her to trust more.
• 21:59
  SPEAKER_01: It's giving her more information about when not to trust, which is arguably more valuable.
• 22:03
  SPEAKER_00: If it works.
• 22:04
  SPEAKER_00: And if the operator actually builds the governance around it instead
• 22:06
  SPEAKER_00: of just turning it loose and hoping for the best.
• 22:09
  SPEAKER_01: Which is the story of every enterprise AI deployment, right?
• 22:12
  SPEAKER_01: The technology is usually ahead of the organization's ability to use it responsibly.
• 22:16
  SPEAKER_00: So what's the field report this week?
• 22:17
  SPEAKER_00: Because I feel like we've been circling one idea for the last 20 minutes without naming it.
• 22:21
  SPEAKER_01: The reconstructed data problem.
• 22:23
  SPEAKER_01: The thread through everything.
• 22:24
  SPEAKER_01: Venezuela, Acquired Assets, Reserve Evaluations.
• 22:28
  SPEAKER_01: The common bottleneck is always,
• 22:29
  SPEAKER_01: we don't trust the historical record and nobody has time to rebuild it manually
• 22:32
  SPEAKER_01: before the decision deadline.
• 22:34
  SPEAKER_00: Every acquisition I've worked,
• 22:35
  SPEAKER_00: there's a moment where the engineering team looks at the production data and goes,
• 22:38
  SPEAKER_00: this can't be right.
• 22:40
  SPEAKER_00: Might be a well that's supposedly produced for three years
• 22:42
  SPEAKER_00: but has no reported volumes for 18 months in the middle.
• 22:44
  SPEAKER_00: Might be formation tops that don't match the logs and somebody's got
• 22:47
  SPEAKER_00: to go figure out what actually happened before you can run economics on it.
• 22:50
  SPEAKER_01: Which right now means an engineer or a tech opening 17 tabs,
• 22:54
  SPEAKER_01: state regulatory database,
• 22:56
  SPEAKER_01: the operator's internal system,
• 22:57
  SPEAKER_01: maybe a third-party data vendor,
• 22:59
  SPEAKER_01: and manually cross-referencing until they can piece together a plausible history.
• 23:03
  SPEAKER_01: On one well.
• 23:04
  SPEAKER_01: And the package has 400 wells.
• 23:05
  SPEAKER_00: And that's when people start making assumptions they shouldn't.
• 23:08
  SPEAKER_00: Because you've got a bid deadline in three weeks,
• 23:10
  SPEAKER_00: and you can either do careful forensic work on 400 wells,
• 23:13
  SPEAKER_00: or you can do it on 40 and hope the rest behave like the type curves.
• 23:16
  SPEAKER_01: Here's the build.
• 23:18
  SPEAKER_01: What should exist and mostly doesn't?
• 23:20
  SPEAKER_01: An automated data reconciliation layer that pulls from every available source,
• 23:24
  SPEAKER_01: state records, SCADA,
• 23:25
  SPEAKER_01: internal databases,
• 23:27
  SPEAKER_01: third-party feeds,
• 23:28
  SPEAKER_01: and for each well produces a unified timeline with conflict flags.
• 23:31
  SPEAKER_01: Not a clean answer.
• 23:32
  SPEAKER_01: A map of where the sources agree, where they disagree, and how big the disagreement is.
• 23:37
  SPEAKER_00: Hold on.
• 23:38
  SPEAKER_00: I like the concept,
• 23:39
  SPEAKER_00: but you're describing something that needs to handle,
• 23:41
  SPEAKER_00: what, dozens of different data formats?
• 23:43
  SPEAKER_00: State reporting varies by jurisdiction,
• 23:45
  SPEAKER_00: operator systems are all different,
• 23:47
  SPEAKER_00: and half the time the API numbers don't even match across sources
• 23:49
  SPEAKER_00: because somebody transposed two digits in 2016.
• 23:52
  SPEAKER_01: And that's why the system has value.
• 23:54
  SPEAKER_01: The fuzzy matching on well identifiers,
• 23:55
  SPEAKER_01: the format normalization,
• 23:57
  SPEAKER_01: the logic for resolving conflicts between sources with different reporting cadences.
• 24:00
  SPEAKER_01: That is when it earns its keep.
• 24:02
  SPEAKER_01: A human doing that across 400 wells is slow and inconsistent.
• 24:05
  SPEAKER_01: A well-designed system does it the same way every time and shows its work.
• 24:08
  SPEAKER_00: Shows its work is easy to say.
• 24:10
  SPEAKER_00: In practice, the provenance trail has to be genuinely useful to an engineer under time pressure.
• 24:15
  SPEAKER_00: If I've got to click through nine layers of metadata to figure out why
• 24:17
  SPEAKER_00: the system thinks so well was shut in for six months, I'm not using it.
• 24:20
  SPEAKER_00: I'm going back to my spreadsheet.
• 24:22
  SPEAKER_01: Fair.
• 24:23
  SPEAKER_01: And I should be honest,
• 24:23
  SPEAKER_01: I've built tools where the transparency layer looked great in the design doc
• 24:26
  SPEAKER_01: and was completely unusable in practice.
• 24:29
  SPEAKER_01: Six nested menus to get to the source data.
• 24:31
  SPEAKER_01: Nobody clicked past the second one.
• 24:33
  SPEAKER_01: So the output needs to be more like a diff view.
• 24:36
  SPEAKER_01: Here's what source A says.
• 24:37
  SPEAKER_01: Here's source B.
• 24:38
  SPEAKER_01: Here's where they diverge.
• 24:40
  SPEAKER_01: Here's our best reconciliation and the confidence level.
• 24:42
  SPEAKER_01: Two clicks max.
• 24:43
  SPEAKER_01: If the engineer sees a low confidence flag, she drills in.
• 24:47
  SPEAKER_01: If it's high confidence, she moves on.
• 24:48
  SPEAKER_00: And who's the user here?
• 24:50
  SPEAKER_00: Because the engineer running economics has different needs than
• 24:52
  SPEAKER_00: the landman verifying title or the geologist checking formation assignments.
• 24:56
  SPEAKER_01: Initially?
• 24:58
  SPEAKER_01: The A and D team.
• 24:59
  SPEAKER_01: The people evaluating acquisition packages on tight timelines.
• 25:02
  SPEAKER_01: They're the ones who feel the pain most acutely
• 25:04
  SPEAKER_01: because they're making capital allocation decisions on data they know is incomplete.
• 25:08
  SPEAKER_01: But the same underlying reconciliation would feed into reserves work,
• 25:10
  SPEAKER_01: development planning, even regulatory compliance.
• 25:13
  SPEAKER_00: The human approval boundary.
• 25:14
  SPEAKER_00: Where does the system stop and the person start?
• 25:16
  SPEAKER_01: The system never makes a final determination on what actually happened downhole.
• 25:20
  SPEAKER_01: It presents the reconciled view, flags the conflict, and the engineer signs off.
• 25:24
  SPEAKER_01: I've seen something like that tried with a traffic light dashboard before.
• 25:26
  SPEAKER_01: The problem was everything was yellow, no useful discrimination.
• 25:45
  SPEAKER_01: Which means whoever builds it needs to understand how reserves engineers think
• 25:49
  SPEAKER_01: about materiality, not just the data science.
• 26:09
  SPEAKER_01: And that connects to what we were talking about with the modeling gap.
• 26:11
  SPEAKER_01: Remember last week the disconnect
• 26:13
  SPEAKER_01: between how forecasting tools handle logistics disruptions versus price signals?
• 26:17
  SPEAKER_01: Same structural problem.
• 26:18
  SPEAKER_01: The tool needs to understand what matters to the decision maker,
• 26:21
  SPEAKER_01: not just what's technically measurable.
• 26:23
  SPEAKER_01: A 1% production discrepancy on a stripper while nobody cares about.
• 26:27
  SPEAKER_01: A 10% discrepancy on your best horizontal.
• 26:29
  SPEAKER_01: Different conversation entirely.
• 26:31
  SPEAKER_00: And in a Venezuela-type situation where the production records might have years
• 26:34
  SPEAKER_00: of gaps, you're not even in traffic light territory.
• 26:37
  SPEAKER_00: You're in reconstruct from first principles using whatever indirect evidence exists.
• 26:41
  SPEAKER_01: Which is the extreme case, but it proves the value.
• 26:44
  SPEAKER_01: If the system can handle the hard version,
• 26:46
  SPEAKER_01: incomplete records,
• 26:47
  SPEAKER_01: multiple operators over decades,
• 26:49
  SPEAKER_01: political disruptions that interrupted reporting,
• 26:51
  SPEAKER_01: then a normal Permian acquisition package is manageable by comparison.
• 26:54
  SPEAKER_00: I wouldn't call anything in the Permian easy, but I take your point.
• 26:57
  SPEAKER_01: The evidence base is there.
• 26:59
  SPEAKER_01: Every operator we've talked to who does acquisitions says the same thing.
• 27:02
  SPEAKER_01: Data reconciliation is the bottleneck.
• 27:04
  SPEAKER_01: It's manual, it's error-prone, and it directly affects bid quality.
• 27:08
  SPEAKER_01: The system that solves this doesn't replace the engineer.
• 27:11
  SPEAKER_01: It gives her a three-week head start.
• 27:13
  SPEAKER_00: And the bid that understands the inventory faster wins, or at least doesn't overpay.
• 27:17
  SPEAKER_01: Knowing what not to bid on is half the value.
• 27:19
  SPEAKER_00: That's actually fair.
• 27:21
  SPEAKER_01: So, if you're listening to this,
• 27:22
  SPEAKER_01: thinking,
• 27:23
  SPEAKER_01: that's my exact problem,
• 27:24
  SPEAKER_01: we just lost two weeks reconciling production data on our last acquisition.
• 27:27
  SPEAKER_01: Reach out.
• 27:28
  SPEAKER_01: Siliconandshale at gmail.com.
• 27:31
  SPEAKER_01: We build this stuff.
• 27:32
  SPEAKER_00: Whether it's Venezuela or a Permian bolt-on, the question's the same.
• 27:34
  SPEAKER_00: Does your system understand what matters, or does it just understand what's measurable?
• 27:37
  SPEAKER_00: Two very different things.
• 27:53
  SPEAKER_01: And if your company sells to operators,
• 27:55
  SPEAKER_01: landmen,
• 27:55
  SPEAKER_01: engineers,
• 27:56
  SPEAKER_01: anyone in the energy space,
• 27:57
  SPEAKER_01: and you want to reach the people who actually make buying decisions,
• 28:00
  SPEAKER_01: we're taking sponsors.
• 28:01
  SPEAKER_01: Siliconandshale at geemail.com.
• 28:04
  SPEAKER_01: We'll use your product before we talk about it.
• 28:06
  SPEAKER_00: That's not a marketing line.
• 28:07
  SPEAKER_00: We're not going to talk about something we haven't touched.
• 28:09
  SPEAKER_01: I've got a call with an asset team in 20 minutes,
• 28:11
  SPEAKER_01: and I still haven't looked at their decline curves.
• 28:13
  SPEAKER_00: Go.
• 28:14
  SPEAKER_00: I'll be here reconciling production data by hand like it's 2014.