There is a line every research project has to cross before it is a product, and it is not the line the team thinks it is. It is not peak accuracy, and it is not a feature list. It is the moment an outside buyer agrees to put their own scan through the tool and put their name on the number that comes out. We had a model that read curves off scanned paper well-logs, and for a long stretch it was a demo: impressive to watch, run by us, on data we had chosen. What moved it over the line was three signed letters of intent, and the fact that three buyers who agreed on almost nothing agreed on exactly what the tool had to do before they would trust it. This is the account of how those three signatures defined pilot readiness and set the scan-to-LAS contract we then built to meet.
Readiness is a buyer's word, not an engineer's
Inside the team, "ready" had always meant an internal thing: the recall had climbed, the loss had settled, the validation notebook ran clean. That definition is comfortable because it is entirely under our control, and it is close to useless, because it never asks the question that decides whether a demo becomes a pilot: would someone who did not build this run their own well through it and act on the result. A model can be excellent by every internal measure and still be a demo, because a demo is defined by who is willing to depend on it, and until the letters that was nobody but us.
A letter of intent changes that definition from the outside. It is not a contract and it is not a sale, but it is a buyer describing, in writing, the conditions under which they would commit, which makes it a readiness specification written by the only people whose opinion of readiness counts. We had three of them, and rather than treat them as sales artifacts, we treated them as the acceptance criteria for the pilot. The question stopped being whether the model was good enough for us and became what these three independent buyers said the tool must do before they would run it, a far more answerable question and a much harder bar.
Three buyers, five segments, one archive problem
The three letters came from buyers who could not have been more different, and the difference is the point, because it is what made their agreement mean something. The first was an exploration group that wanted to read its own filing cabinet of scanned logs. The second was a private-equity holder of producing assets, which wanted to digitise across a portfolio it did not itself operate. The third was a national oil company, sitting on one of the largest paper archives in existence and wanting the same curve-reading engine behind its own firewall. Between them the three letters spoke for the breadth of the five customer segments the plan addressed, from the smallest independent to the largest state operator, and that breadth is what let us read their overlap as a market signal rather than a coincidence.
They disagreed on almost everything about the product wrapper. But every one of them, without coordinating, named the same three things the tool had to do before a pilot could start. It had to prove itself against real scans, not synthetic ones. It had to be something their own people could drive without us in the room. And whatever it produced had to load in the petrophysical software already on their desks. Those three demands, arrived at independently by an explorer, a financial holder, and a national oil company, were the readiness bar; everything else was preference.
What the overlap actually required
Each of the three shared demands mapped to a concrete piece of engineering on the critical path to a pilot.
Prove it against real scans meant a validation step we could show a buyer. We ran the model against a small set of real field scans and put every one in front of a human reviewer, overlaid on the prediction and checked against the operator's own log, so the output that left the building had been read by a person rather than trusted on faith. It was a small set, and we never pretended otherwise, but a handful of real scans a reviewer signed off is a categorically different claim than a strong metric on data we curated, and it is the claim a buyer's letter actually asked for.
Something their own people could drive meant the model had to sit behind a product, not a notebook. That became a four-step self-serve dashboard: upload a scan, run the segmentation, validate the overlay, download the result. It was the difference between a tool only its authors could operate and one an interpreter at the buyer's own desk could run, which is the workflow-reach point Koroteev and Tekic make about upstream tools, that value is gated less by sophistication than by whether the tool reaches the person meant to use it [1]. A per-seat product at 1200 USD per user per year only exists if a seat holder can serve themselves.
Load in the software on their desk was the strictest demand, and the one an engineering team is most likely to underweight. The buyers did not want an array or a proprietary export. They wanted a file their existing petrophysical software would open without complaint, which in this industry means a Log ASCII Standard file, the plain-text format for digital logs since 1992 [3]. That set a hard contract on the output: the digitiser was done not when it produced a correct curve but when it produced a spec-compliant scan-to-LAS deliverable. We had a reference for what right looked like, because the public Texas Railroad Commission dataset we trained against ships thousands of digital LAS files alongside its scanned rasters, so the format the buyers demanded was the format our ground truth already spoke [2].
Where the argument peaks
The instrument below is the readiness bar made legible. Each of the three letters is a column; each requirement the letters named is a row; a cell lights when that buyer's letter asked for that capability. The convergence lever sets how many of the three letters must independently name a requirement before the pilot commits to it, and the readiness verdict climbs a ladder as the shared set thickens.
Drag the lever to three, the strictest setting, and only the rows every buyer signed for survive: the scan-to-LAS export contract, the real-scan validation sign-off, and the four-step dashboard. Those three, and only those three, are what an explorer, a private-equity holder, and a national oil company all demanded without talking to each other, and they are exactly the three that carry the verdict from research demo to pilot-ready. Ease the lever and the single-buyer demands light up too, the on-prem wrap and the white-label theme the national oil company wanted, but easing the lever is the mistake: those are real for one account and irrelevant to a pilot's readiness. The verdict is honest only at the strict setting, and there the shared three are the whole story.
Why the shared requirements, and not the loudest one, defined ready
The pull in the other direction was strong. The national oil company was the largest buyer and its letter was the most demanding, and there is a permanent temptation to let the most impressive signature define what ready means. Had we done that, pilot readiness would have meant on-prem deployment and a white-labelled interface, and we would have spent the run building infrastructure for a single account while the tool stayed a demo for everyone else. The three letters, read together, corrected that. Readiness was not the union of what every buyer wanted, which is unbuildable, nor the wish list of the biggest buyer, which is bespoke. It was the intersection, the narrow set every independent buyer arrived at on their own, buildable precisely because it is narrow.
That framing matched the market the plan addressed. The oil and gas transactions market was sized at roughly 134B USD, of which the serviceable software slice was about 6.7B USD, and a per-seat product at 1200 USD per user per year earns inside that slice by widening seat count, not by deepening a single account. A requirement only the largest buyer wants is scoped to a sliver of an already-thin market, while a requirement all three named unlocks seats across the segments. Defining readiness by the overlap was the same discipline as pricing per seat: both reward what recurs across buyers and discount what lives inside one.
What the three signatures left us with
The lasting effect of those three letters was a change in what we meant by ready. Before them, ready was an internal number and the pilot never quite arrived. After them, ready was a checklist the buyers had written themselves: a small set of real scans a reviewer had signed off, a four-step dashboard a stranger could drive, and a scan-to-LAS file the operator's own software opened on the first line. When those cleared, the demo was a pilot, not because we decided it was, but because the people who signed the letters had told us in advance that it would be. A research team learns to trust its own metrics. The more useful lesson from this engagement is that the definition of ready belongs to the buyer, and when three very different buyers write down the same definition without conferring, that overlap is not a suggestion. It is the specification, and it is the one worth building to.
Limitations
The per-letter requirement map in the instrument is an illustrative reconstruction, consistent with what the three letters asked for but schematised into a clean grid for the argument; the letters themselves, the buyer identities, the client, and all personnel are anonymised under operator confidentiality. The counts are the engagement's own records: three letters of intent, five customer segments, a small real validation set, a four-step self-serve dashboard, and the 1200 USD per seat per year price, alongside the roughly 134B USD total and 6.7B USD serviceable market framing from the plan. The validation set is a deliberately small, early sample and is presented as a pilot-readiness signal, not as a statistical accuracy claim; the broad accuracy of the model is treated in separate write-ups. The market figures are planning estimates, not audited results, and the readiness verdict is a product-adoption judgment about when a demo becomes a pilot, not a measure of model quality.
References
-
Koroteev, D. and Tekic, Z. (2021). Artificial intelligence in oil and gas upstream: Trends, challenges, and scenarios for the future. Energy and AI. https://www.sciencedirect.com/journal/energy-and-ai
-
Texas Railroad Commission (Texas RRC). Public well-log dataset, scanned raster (TIF) and digital (LAS) records of Texas wells. https://www.rrc.texas.gov/
-
Canadian Well Logging Society (1992). LAS Version 2.0: A Digital Standard for Logs, Log ASCII Standard. https://www.cwls.org/products/
