Every supervised model we built in this engagement ran on labelled image logs, and the labelling was the part that kept stalling. A mid-project count made the shortfall plain: to train a model we were confident in, we had proposed 25 wells be annotated, and by the time the work was underway only 8 had actually been delivered. The reflex, when annotation is the thing holding up the model, is to write it up as a cost and escalate it: more contractor hours, a bigger line item, a faster invoice. We went the other way. In a leadership meeting on 24 May 2022, at the operator's offices in Muscat, we recommended that the client hire young Omani interns to do the data collection and annotation, and that everyone in the room read that spend not as a cost to be minimised but as the intake of a pipeline.
The distinction matters because the two framings buy completely different things. Annotation-as-cost buys labelled pixels and nothing else; when the contract ends, the capability leaves with the contractor. Annotation-as-pipeline buys labelled pixels and a cohort of people who now understand both the interpretation and the model that consumes it, and who stay after we go. On a producing carbonate field where the deepest interpretation knowledge sits with a handful of senior geologists approaching retirement, that second thing is worth more than the labels.
The labelling task is a training ground in disguise
Annotating a borehole image log is not clerical work. To place a sinusoid, a person has to read the unrolled electrical image, decide whether a wave is a bed or a fracture, fit its depth, dip, and azimuth, and know when a feature is real versus a tool artefact. That is exactly the skill a junior interpreter needs, practised on hundreds of intervals with immediate feedback. An intern who annotates a well has, without anyone calling it a course, spent a week doing the core motion of the job the senior interpreters do.
Put a model in the loop and the training compounds. The annotator sees where the model agrees, where it fails, and why, which is a faster route to understanding a detection pipeline than any lecture on transformers. The person comes out fluent in two things at once: the geology and the tooling that now reads it. This is why we framed the ask as capability building rather than staffing. The same hours that produce the labels produce the people.
Enrolment was already underway
This was not a slide-only idea. By July 2022 we had 15 young Omanis enrolled in a data-science track, filtered and tracked through our EdTech reporting tool, then in beta, as part of a wider cohort of 55 professionals trained across the region. The 15 were the local core of an Omanization effort that ran alongside the model work: the same people who would take annotation intervals were the ones being brought up the data-science curve. The pipeline had a real intake, not a hypothetical one, and the meeting notes tied it directly to speeding both the supervised and the unsupervised programs.
The same meeting recorded a second, quieter argument: wider client-team participation drives adoption before a project closes. Our engagement was anchored to an August 2023 close. A model that only two contractors can run does not survive that date; a model that a growing group of client staff have annotated for, argued with, and corrected does. Adoption is not a launch event at the end. It is the accumulated participation of the people who were in the loop the whole way, and annotation is the cheapest, highest-frequency way to put them there.
The funnel, and the retirement it has to out-run
The mechanism, drawn as it actually runs. The intern cohort is the top of a funnel. A fraction advance to junior interpreter; a fraction of those become staff who can work fluently with the AI pipeline. That is the talent half. The retention half runs in parallel: as senior interpreters retire, tacit knowledge leaves the organisation at a steady rate, and the people the pipeline produces are what offset that drain. Widen the intake and you widen the AI-literate base at the bottom, which is the inflow that decides whether the operator's knowledge stock rises or falls over the engagement window.
Drag the cohort lever and the argument is visible in one line. Below a certain intake, the retention inflow cannot keep pace with retirement and the knowledge stock slides down the same path it would with no pipeline at all. Above it, the orange line separates from the drain line and the operator ends the engagement knowing more than it started with, in-house, without us. The crossover is the whole recommendation compressed to a point: annotation staffing is not a cost curve to flatten, it is a lever you push up until the pipeline out-runs the retirements.
Why this belongs to the vendor, not just the client
There is a self-interested reading of all this, and it is worth being honest about it. We were blocked on data. Recommending that the client resource annotation solved our blocker. But the framing we chose changed what the client got for solving it. We could have asked for more contractor hours and left with our labels. Instead we asked them to build a cohort, and the cohort is theirs to keep. A vendor who turns a supply problem into a capability handed to the client is worth re-hiring; a vendor who turns it into an invoice is worth replacing.
The move generalises. Any AI vendor blocked on a client's proprietary data is sitting on the same opportunity: the labelling that unblocks you is also the fastest apprenticeship your client's junior staff will ever get. The mid-engagement workshop where we made the operational version of this ask, and the phased operating model behind standing up in-house capability, are told in their own companion pieces; the point here is narrower and older than either. Before any of that, in a single meeting in May 2022, the recommendation was to read annotation headcount as a talent-and-retention pipeline. The numbers we had, 15 enrolled, 55 trained, 25 wells asked for and 8 delivered, are small. The reframing is not.
Limitations
The pipeline funnel in the exhibit is a model, not an HR audit. The stage conversion fractions (how many interns reach junior interpreter, how many juniors become AI-literate staff) and the per-quarter retirement-drain rate are illustrative inputs chosen to show the mechanism; they are not measured attrition or promotion outcomes from the operator. What is sourced is the recommendation itself, the enrolment figures (15 Omanis in July 2022, a wider cohort of 55), the annotation demand (25 wells proposed versus 8 delivered), the adoption-before-close argument, and the August 2023 anchor. We did not track the individual interns' career outcomes past the engagement, so the retention payoff is argued from the design of the pipeline rather than verified from a multi-year follow-up. The knowledge-stock index is a relative illustration, not a calibrated measure of tacit expertise.
References
[1] T. Fountaine, B. McCarthy, and T. Saleh, "Building the AI-Powered Organization," Harvard Business Review, July-August 2019. https://hbr.org/2019/07/building-the-ai-powered-organization
[2] McKinsey & Company, "Scaling AI like a tech native: The CEO's role," 2021. https://www.mckinsey.com/capabilities/quantumblack/our-insights
