Two to Three Times More Rigs per Engineer Is a Claim About Alert Precision

ADNOC's release of 4 August 2026 says the company, with SLB, has deployed its artificial intelligence enabled Real-Time Operations Center across a fleet of more than 120 onshore and offshore rigs. The platform, it says, cuts engineering effort by 30% to 40% and lets engineers support "two to three times more rigs" while keeping effective oversight [1]. The same release says the RTOC can identify potential issues before they escalate, reducing incident response times by 4 to 12 hours and avoiding one to two days of rig downtime [1]. SLB's newsroom update carries the same figures on the same day and marks them as borrowed, prefacing them with "According to ADNOC" [2].

Three more items sit around it in the same month. On 1 September YPF published a note on renewing its agreement with Corva, which the note calls one of the central technology platforms of the Real Time Intelligence Center for drilling and workover in Vaca Muerta, describing a new stage aimed at "operaciones progresivamente autonomas" by way of advisors, predictive analytics, smart alerts, automation, digital copilots and agents [3]. On 25 August Halliburton announced that bp had awarded it an integrated drilling services contract for the first appraisal campaign in the Bumerangue field, offshore deepwater Brazil, in which LOGIX automation and remote operations "will be deployed" [5]. On 13 August Aramco Digital and SANAD signed for a cloud ERP platform, infrastructure modernisation and managed technology services, framed by both chief executives as the digital foundation for Industrial AI that has not been built yet [6].

Read those four together and the shape of the evidence is clear. Only one of them carries figures, it is the oldest of the four, and its central figure is a ratio with no denominator. ADNOC does not say how many rigs one engineer covered before, so "two to three times more rigs" is a multiple of a number nobody has published. And not one of the four states the quantity that decides whether a span of control that size is real: how many alerts an engineer has to work through per rig-hour, counting the ones that turn out to be nothing.

What each release says, and who is saying it

The provenance matters here more than usual, because two of the four legs are a service company speaking about an operator.

ADNOC's is the strongest of the four and it is first-party. The figures are ADNOC's own claims: no baseline, no measurement method, no before-and-after period, and no model, technique or metric named anywhere on the page. SLB's DrillOps is named as what enables the platform [1][2], and SLB's president of Digital describes it as turning "real-time drilling data into operational intelligence" [1]. The platform is hosted in ADNOC's own cloud environment in the United Arab Emirates [1][2]. Nothing on either page is an accuracy number, a detection rate or an alert count.

YPF's note is first-party and carries no figures at all. It says the RTIC allowed a move from traditional monitoring toward active management of operations, where specialists and field crews can see performance in real time, "anticipar eventos de riesgo" and reduce non-productive time [3]. Smart alerts and digital agents are named as the direction, not as something already running. The nearest thing to a headcount on any YPF page is older: the RTIC inauguration note of 13 December 2024 says the room runs 24 hours a day, seven days a week, with 88 professionals across seven operating units, five of them on drilling and covering the 20 rigs then working in Vaca Muerta, and two on completions covering up to 8 frac sets at once [4]. That is a staffing statement from December 2024 and it is used here as nothing else.

The bp leg is Halliburton's account, not bp's. Halliburton's release quotes its own senior vice president for Latin America and nobody from bp [5]. What it describes is an integrated drilling services package with LOGIX automation and remote operations; the word AI appears once, in a closing sentence about data, AI and advanced drilling technologies delivering real-time insights [5]. Reading that as bp adopting AI would be reading a vendor's marketing line as an operator's commitment, so it is used here only as a remote-operations data point.

The Aramco Digital item is a readiness statement. The services are cloud ERP, infrastructure modernisation and managed technology services, and the Industrial AI in it is future tense in both quotes: Aramco Digital's chief executive says "Industrial AI begins with secure, connected, and trusted digital foundations", and SANAD's says the work prepares the organisation "to benefit from emerging digital and Industrial AI technologies" [6]. SANAD is described by Nabors in its second-quarter 2026 results as "the SANAD land drilling joint venture" in the Kingdom of Saudi Arabia, and as "our SANAD joint venture"; that document does not name the other partner, and neither does this piece [7].

The queue, not the dashboard

An operations centre does not run out of screens. It runs out of engineer-minutes. Every alert that reaches the desk has to be looked at by somebody before anyone knows whether it was a precursor or noise, because deciding that is what triage is. So the false alerts do not sit in a separate bucket. They sit in the same first in first out queue as the real ones, they consume the same attention, and they push the real ones back.

That is the constraint span of control runs into. Adding a rig to an engineer's plate does not add a screen, it adds an arrival stream. And what an alert is worth is not what the model scored it, but whether a human reached it while there was still time to act.

Our own record puts a number on that time. For a Gulf national oil company we built a real-time advisory agent on WITSML rig feeds; it ran at sub-30-second latency from sensor to alert, routed to the driller console and to the remote operations centre, and it caught 82% of stuck-pipe events between two and eight minutes before the driller would have seen the first console alarm [8]. That two to eight minute band is the window. Reach the alert inside it and the driller can reduce weight on bit, increase circulation or pull back into casing; reach it after and the reaction time the model bought has already been spent. The same engagement recorded what the other side of the ledger costs: early thresholds produced so many nuisance alerts that drillers ignored the system within three days, and it took six weeks of calibration, two of them on site, to reach 85% of incidents caught with fewer than two false positives per well [8].

The arithmetic

Take an engineer covering $n$ rigs, each sending $\lambda$ alerts an hour counting the false ones, with $s$ minutes of attention per alert. Offered load over service capacity is

$$\rho \;=\; \frac{n\,\lambda\,s}{60}$$

FormulaLaTeX

Copy LaTeX

For a single-server first in first out queue with Poisson arrivals and exponential service, the waiting time in queue has a standard distribution, and the share of alerts reached before an escalation window of $T$ minutes closes is

$$P(W_q \le T) \;=\; 1 - \rho\,e^{-\frac{T}{s}\left(1-\rho\right)}, \qquad \rho < 1$$

FormulaLaTeX

Copy LaTeX

with the steady-state share at $\rho \ge 1$ equal to zero, because the queue does not drain and the backlog grows for as long as the shift lasts. True precursors and false alerts wait the same distribution, so this is also the share of true precursors reached in time, which is the only share that matters.

Two structural facts fall out of that expression and neither depends on any number in it. First, $n$ and $\lambda$ appear only inside $\rho$, and only as their product. Multiply the span by $m$ and the alert rate has to be divided by exactly $m$ to hold the same share. Not improved, not tightened: divided by $m$. Second, the share is flat over most of the range and then collapses across a narrow band as $\rho$ approaches one, so the largest span that holds a target share is an edge rather than a slope.

The bench

The exhibit below computes that surface rather than drawing it. Left to right is the span of control, 1 to 6 rigs per engineer. Back to front is the alert rate, 0.1 to 10 an hour per rig by decade, counting the false ones. Height is the share of true precursors an engineer reaches before the window closes. The amber sheet is 95% of them, so the curve where the surface cuts that sheet is the set of spans and alert rates that hold 95%. The white curtain is your span and the labels follow it; the bright line is your alert rate.

The escalation window opens at 8 minutes, the top of our recorded 2 to 8 minute band, which is the most generous end of our own evidence [8]. The alert rate of 6 an hour per rig and the 2 minutes of attention per alert are stated assumptions, and the plate says so. We recorded false positives per well, not per rig-hour, and no well duration was published with them, so a per-hour rate cannot be derived from our record without inventing one. That slider is where your own rate goes.

rigs per engineer, 1 to 6: 1.00alerts per rig-hour counting false ones, 0.1 to 10: 6.00escalation window, 2 to 15 minutes: 8.0triage minutes per alert, 0.5 to 10: 2.00

What the bench shows

At the opening settings one engineer on one rig reaches 99.2% of precursors inside the window, at 20.0% utilisation. Drag the span to 2 and it reads 96.4%. Drag it to 3, which is the top of ADNOC's multiple if the baseline was one rig, and it reads 87.9%: roughly one precursor in eight is now reached too late. ADNOC does not say what the baseline was. At 4 rigs it is 64.1%. At 5 the utilisation readout reaches 100.0%, the hero reads 0.0%, and the queue never catches up.

The readout that carries the argument is the one on the left. At 6 alerts an hour per rig, the span that holds 95% is 2.25 rigs. That number sits inside the two-to-three band, which is exactly why the band is the interesting one: an operation can be comfortably inside it at the bottom and past the edge at the top, with nothing in the release to tell you which.

Now use the alert-rate slider to buy the span back. The alert budget readout says 13.51 an hour at one rig and 4.50 at three: a factor of three in the span, a factor of three in the budget, because the two enter only as their product. So the honest form of a three-rigs-per-engineer target is not a better model. It is an alert rate one third of what one rig could carry, and that is a precision requirement at fixed recall that somebody has to own.

Then drag the escalation window from 8 minutes down to 2, the bottom of our own recorded band. At three rigs the alert budget falls from 4.50 an hour to 1.20, which is one alert per rig every 50 minutes, counting the false ones. The span that holds 95% at 6 alerts an hour falls from 2.25 rigs to 0.60, which is less than one rig per engineer. The window is our measurement, not a modelling choice, and the two ends of our own recorded band are the difference between an alert budget of 4.50 an hour and one of 1.20.

That is the finding. ADNOC published a response-time improvement and a downtime avoidance, both of which are outcomes of reaching things in time, and a span-of-control multiple, which is what reaching things in time has to be traded against [1]. It did not publish the alert rate that connects the two, and neither did the other three. On this model a span of three rigs holds 95% only at an alert rate at or below 4.50 an hour per rig at the top of our recorded window, and at or below 1.20 at the bottom of it. Those two ceilings are our inference from our own queue and our own recorded window, and never anything ADNOC said.

Where the model is weakest, and what survives

A drilling superintendent will push back on two of the assumptions, and both objections are right.

Real desks triage by severity, not first in first out. A priority discipline does move true precursors up the queue, and it would raise the surface everywhere. But it can only sort by a score assigned before a human has looked, so its benefit is bounded by how well that score separates precursors from noise, which is the same precision question in a different place. A priority queue does not remove the alert budget; it changes the exchange rate.

Exponential service times are a convenience. Triage is probably tighter than exponential around its mean, which would help. What does not depend on that assumption is the shape: any single-server queue has a region where the wait is short against the window and a collapse as utilisation approaches one, and in any model where the rigs superpose into one stream the span and the per-rig rate reach the desk only as the total arrival rate. Those two features are properties of the queue, not of the exponential tail, and they are the two the argument rests on.

One more caveat, and it is ours. The 2 to 8 minute band and the sub-30-second alert latency are recorded on land rigs on one fleet [8]. A deepwater appraisal campaign of the kind Halliburton describes for bp is a different operation with different timescales [5], and the window slider is there because the number should be the reader's own.

Three questions before signing an operations-centre contract

How many alerts reach the desk per rig-hour, counting the ones that resolve to nothing? Not the model's precision on a test set: the arrival rate at the human. On this model it is the number that sets the span, and none of the four releases above contains it.

How long is the window on the events you actually care about, measured from your own alert to the point where the intervention stops working? Ours is 2 to 8 minutes on stuck pipe [8]. The bench shows that moving the window from the top of that band to the bottom cuts the alert budget at three rigs from 4.50 an hour to 1.20.

And what is the span of control today, in rigs per engineer? A multiple is not a number. ADNOC's two to three times is a multiple of an unpublished baseline [1], and YPF's own December 2024 page is the only one of the four that ever put a headcount and a rig count on the same line [4]. Ask for both, then read the alert rate against them.

Key takeaways

1. 1ADNOC's 4 August 2026 release says the Real-Time Operations Center runs across more than 120 rigs and lets engineers support two to three times more of them, with incident response 4 to 12 hours faster and one to two days of rig downtime avoided. Those are ADNOC's own claims with no baseline and no method, and SLB's parallel update attributes them to ADNOC rather than measuring them.
2. 2Two of the four legs are not the operator speaking. Halliburton, not bp, announced the Bumerangue integrated contract and the LOGIX automation and remote operations on it, and quotes no one from bp. Aramco Digital and SANAD signed for cloud ERP, infrastructure and managed services, with Industrial AI in the future tense in both chief executives' quotes.
3. 3Span of control is set by the triage queue, because a false alert is indistinguishable from a precursor until a human has looked at it and so occupies the same queue. In a single-server queue the span and the alerts per rig-hour enter only as their product, so tripling the span needs the alert rate divided by exactly three.
4. 4At an assumed 6 alerts per rig-hour, 2 minutes of triage and the top of our recorded 2 to 8 minute lead time, one engineer reaches 99.2% of precursors at one rig, 96.4% at two, 87.9% at three and 64.1% at four, and the queue stops draining at five rigs. The span that holds 95% is 2.25 rigs, inside the two-to-three band.
5. 5The alert budget at 95% is 13.51 an hour at one rig and 4.50 at three. At the bottom of our recorded window rather than the top it is 1.20 an hour at three rigs, one alert per rig every 50 minutes, and the span that holds 95% at 6 alerts an hour falls to 0.60 rigs.
6. 6The procurement question is therefore alerts per rig-hour at the desk, counting false ones, and the length of the window on your own events. None of the four releases states either, and a span-of-control multiple without a baseline states neither.

References

[1] ADNOC. ADNOC and SLB Deploy AI Platform Across Over 120 Drilling Rigs to Strengthen Upstream Performance. 4 August 2026. https://adnoc.ae/en/news-and-media/press-releases/2026/adnoc-and-slb-deploy-ai-platform-across-over-120-drilling-rigs-to-strengthen-upstream-performance

[2] SLB. ADNOC Deploys SLB Technology Across its Rig Fleet Through AI-Enabled Operations Center. 4 August 2026. https://www.slb.com/newsroom/updates/2026/2026-0804-slb-adnoc-rtoc

[3] YPF. YPF y Corva renuevan un acuerdo clave para el RTIC de Perforacion y Workover. YPF Hoy, 1 September 2026. https://novedades.ypf.com/ypf-corva-renovacaciondelacuerdo.html The body of this page is rendered client side and was read in a browser.

[4] YPF. YPF revoluciona la industria energetica con su Centro de Operaciones en Tiempo Real. YPF Hoy, 13 December 2024. https://novedades.ypf.com/ypf-revoluciono-la-industria-energetica-centro-de-operaciones-tiempo-real.html Also client side rendered, read in a browser.

[5] Halliburton. bp Awards Halliburton Integrated Contract for Bumerangue Field Appraisal in Brazil. 25 August 2026. https://www.halliburton.com/en/about-us/press-release/bp-awards-halliburton-integrated-contract-for-bumerangue-field-appraisal-in-brazil A Halliburton release. No bp representative is quoted in it.

[6] Aramco Digital. Aramco Digital and SANAD Sign Digital Transformation Agreement. 13 August 2026, dated from the listing at https://aramcodigital.com/news. https://aramcodigital.com/news/news-article-2

[7] Nabors Industries Ltd. Nabors Announces Second Quarter 2026 Results, exhibit 99.1. US Securities and Exchange Commission, 2026. https://www.sec.gov/Archives/edgar/data/1163739/000110465926087564/tm2621313d2_ex99-1.htm

[8] EarthScan. Cutting drilling non-productive time 38% with a real-time advisory agent. https://earthscan.io/case-studies/cutting-drilling-npt-real-time-advisory-agent