For mid-tier oil & gas operators · the cost-side discipline
Personal use of AI is heavily subsidized. ChatGPT free tier, Claude free tier, GitHub Copilot at $10 to $20 a month, Microsoft 365 Copilot at $30 per user. Those costs do not scale. The mental model that follows from those tools is that AI is basically free or nearly free. That mental model breaks the moment you try to do anything operationally meaningful with it.
Start with the end in mind. Score the business value in dollars. Account for the cost honestly. The AI initiative survives the first board review. Skip the discipline and you join the 95 percent with no measurable ROI on $30 to $40 billion of enterprise GenAI spend through 2025 (MIT NANDA). This page is the cost-side counterpart to the three objectives that count. Both sides matter.
The personal-use subsidy reality
Every major AI vendor (OpenAI, Anthropic, Microsoft, Google) runs consumer products at a substantial loss. The strategy is customer acquisition for enterprise conversion. The consequence is that the mental model individual users develop (“AI is basically free”) is wrong at any operationally meaningful scale. Here is how the subsidies show up.
OpenAI has reported multi-billion-dollar annual losses through 2024 and 2025 against revenue, with publicly disclosed projections of meaningfully larger near-term losses before profitability. The free tier is part of how that loss profile was built. The unit economics that work for an individual asking ten questions a day do not survive an enterprise deployment running thousands of agentic calls per hour.
Anthropic has raised at successively higher valuations to fund inference compute and growth, and publicly discusses heavy compute investment ahead of revenue. The pattern across the major AI labs is the same: consumer tiers are run cheap to drive enterprise conversion. Enterprise SKUs and API spend are where the unit economics begin to work, and at scale they are real money.
Trade press through 2023 and 2024 reported (e.g., Wall Street Journal coverage in October 2023) that Microsoft was losing money per user on early Copilot subscriptions at usage levels productive developers actually hit. Pricing has since moved up and tiers have shifted. The takeaway: the $10 personal-developer SKU was not a sustainable economic model on its own, and enterprise versions price meaningfully higher.
The published enterprise SKU. $30 per user per month for 2,000 employees is $720,000 per year. The math that worked for a developer at $10 a month does not survive an enterprise CFO review without a clear productivity-uplift case. Adoption inside enterprises has been measured rather than uniform.
The mental model that “AI is basically free” is the most expensive assumption an operator can carry into a real deployment.
Where enterprise costs actually live
The vendor quote covers the SaaS license. The TCO covers what actually shows up in your AP system over three years. The gap between the two is where most AI initiatives blow their board-approved budget.
A multi-step agentic workflow consumes 10 to 100 times the tokens of a single chat exchange. Production deployment runs 24 hours a day across thousands of decisions. A ranked-work agent that scores 5,000 wells daily on cash flow, risk, and regulatory exposure is generating millions of tokens per week of inference. At enterprise API rates, that becomes a real line item.
Personal AI runs on shared infrastructure with best-effort latency. Operational AI runs against an SLA. 24/7 availability, sub-second response on the work-loop critical path, and on-prem or BYOK deployment for OT security all cost money. The dedicated infrastructure is several multiples of the per-token spot price.
Industrial data volumes are large. SCADA at 1-second granularity across 50,000 to 500,000 tags is terabytes of data. Pulling that into AI workflows, normalizing it, storing the historical context the model needs, and moving it across cloud regions or on-prem infrastructure all carry cost. Most consumer AI uses are stateless; operational AI is data-heavy.
A working closed-loop AI deployment connects to SCADA, ERP, CMMS, GIS, and historian, plus the operator-specific quirks of each. The integration is not free. Whether you pay it as vendor implementation cost, internal IT spend, or systems-integrator fees, it is real money. Vendors who quote "deploy in a week" with no integration cost are usually skipping the work or hiding it in a later year.
Drift detection, retraining, QA-layer monitoring, audit-trail storage, and version management all require ongoing engineering attention. The model that worked in month three quietly degrades by month nine without active maintenance. The cost of maintenance is real and is rarely in the initial vendor quote.
Operations, Engineering, Safety, and Maintenance loops each consume their own inference, their own integration, and their own maintenance attention. Vendors that price per-loop or per-module compound the cost across loops. Vendors that ship one architecture across all four loops on one data layer have a structurally different cost curve.
Why oil and gas amplifies the cost
The cost curve that works for a developer using GitHub Copilot does not transfer to an upstream operator running closed-loop AI on industrial data. The structure of the work is different, and the cost reflects the structure.
A typical mid-tier upstream operator pulls SCADA at 1-second to 10-second granularity across 50,000 to 500,000 tags. The data layer that supports a closed-loop AI deployment carries terabytes of historical context plus continuous streaming. That is several orders of magnitude more data than a typical SaaS customer relationship.
A daily ranked work plan in the truck cab by 6 AM means inference runs overnight against the latest data, every night, on time. Anomaly detection 48 to 72 hours before failure means continuous inference against streaming data. The AI runs 24/7 on dedicated capacity, not on-demand on shared infrastructure. The cost profile is closer to industrial software than to consumer AI.
Single-tenant deployment, BYOK encryption, IEC 62443 zone-and-conduit compliance, and read-only network paths all require dedicated infrastructure rather than shared public-cloud resources. Each of those requirements adds cost relative to the multi-tenant SaaS model that consumer AI runs on.
Three-year manifest retention under Texas 16 TAC Chapter 4. Multi-year compliance retention for safety-related decisions. SB 253 emissions audit trail. Every AI decision has to be drillable to source, indefinitely, for any operator with regulatory exposure. Storage of the audit trail at compliance-grade durability is real money.
The cost discipline
The QA discipline keeps the AI honest. The cost discipline keeps the AI economical. Both pair with the value discipline of the three objectives. All three are required for an initiative that survives the first board review.
Year-1 cost, Year-3 cost, and worst-case Year-3 cost (where token spend, integration overrun, or maintenance load runs higher than initial quote). The board case starts with the TCO, not with the vendor quote. The vendor quote is one input among several.
For agentic workflows, the inference volume scales with deployment surface (loops, wells, decisions per day). Forecast token spend at 1x current scale, 5x current scale, and at full deployment. Vendors who cannot give you a forecast at those three scales do not understand their own cost curve.
The vendor who is willing to be measured on cost-per-outcome (cost per ranked work plan, cost per anomaly detected, cost per defensible audit trail) is a vendor who has thought about scaling. The vendor who quotes per-user or per-token without per-outcome accountability has not.
The same 4-week measurement window that protects you from dashboard theater also protects you from cost runaway. If by Week 4 the cost trajectory is heading where the vendor said it would and the metric is moving, expand. If the cost is exceeding forecast and the metric is not moving, walk away. The discipline is symmetric on cost and value.
AI as displacement means a SaaS contract gets retired and total tech spend goes down. AI as addition means a new line item on top of an already-fat SaaS budget. The cost discipline of "what gets retired" is what separates a winning AI initiative from a vendor-additive one. Objective 1 of the three is also the cleanest cost-discipline test.
Vendors that ship one architecture across multiple loops on one data layer have a structurally different cost curve from vendors who price per-module. WorkSync's Bundle pricing protects against multi-loop cost compounding because the data layer integrates once and the loops are incremental. Per-module vendors compound cost as scope expands.
The three-question cost diagnostic
We ask these of ourselves on every deployment. We ask them of other vendors on every comparison. The vendor whose answers are vague has not thought hard enough about their own cost curve to be trusted with yours.
Most pilot quotes are Year-1, single-loop, sub-scale. The board case is Year-3 at full deployment across all the loops you intend to run. If your vendor cannot give you a defensible Year-3 number with a confidence range, you do not have a vendor, you have a sales pitch.
Some vendors pass usage-based cost increases through to the operator. Some absorb them. Some have fixed-price tiers with usage caps. The contractual answer is the cost-risk answer. If you cannot answer this question after reading the contract, the answer is "you eat it."
If the answer is "none" or "we will figure that out later," the AI is additive and your total tech spend is going up. AI as displacement is the cleanest cost discipline. Name the contract you are retiring before you sign the one you are starting.
AI is not free. Start with the end in mind. Score the value in dollars (three objectives). Account for the cost honestly (this page). Deploy on your existing stack (four-loop framework). Measure on the metric in 4 weeks. Walk away if either the value or the cost is heading the wrong way.
Why the cost discipline matters for us
WorkSync's mission is to leverage technology to drive efficiency and safety in field operations. We see a world where there are zero fatalities and engineers never spend time on data entry again.
Two AI products on one Data Hub. WellOPS for field operations. FlowSync for engineering. The Data Hub is the read-only integration backbone that connects to your existing SCADA, ERP, CMMS, GIS, and historian.
Where to go next
WHY you buy. Replace a SaaS contract, speed up a costly decision, automate a process. Anything else is dashboard theater.
Four loops, one QA discipline. The HOW behind the objectives, structured to keep cost curves predictable.
Per-module, per-asset. Below VP signing authority on the entry tier. Bundle pricing protects against multi-loop cost compounding.
Because the major AI vendors (OpenAI, Anthropic, Microsoft, Google) are running consumer products at a substantial loss to drive enterprise conversion later. The free tier is customer-acquisition cost. The $10 to $20 per developer per month Copilot subscription has been flagged in Microsoft earnings calls as having negative per-user margin at productive usage levels. The economics are real, the subsidy is real, and the subsidy ends the moment you cross into enterprise SKUs.
Six places. Token spend at agentic scale (multi-step workflows running 24/7 burn 10 to 100 times the tokens of single-chat exchanges). Inference compute at OT-grade availability (dedicated infrastructure with SLAs is several multiples of spot pricing). Data movement and storage (industrial data volumes are large). Integration and custom engineering (real engineering work the vendor or your IT team has to do). Ongoing model maintenance (drift detection, retraining, QA monitoring). Multi-loop compounding (each loop has its own inference, integration, and maintenance load if the architecture is not unified).
Four amplifiers. Industrial data volume (SCADA at 1-second granularity across 50,000 to 500,000 tags is terabytes). Real-time inference requirements (24/7 availability against SLAs, not on-demand). OT security requirements (single-tenant, BYOK, IEC 62443, dedicated infrastructure). Audit-trail retention (three-year minimum under Texas Chapter 4, multi-year under various other regimes, compliance-grade durability). Each amplifier compounds with the others.
Per-module, per-asset, with Bundle pricing that protects against multi-loop cost compounding. Below VP signing authority on the entry tier. Land FREE with Data Hub for the integration phase. Better tier ~$95K/year for a 1,000+ well deployment. The pricing is structured for the cost discipline this page describes: predictable, single-architecture, single-tenant where required, and explicitly aimed at AI as displacement (something gets retired) rather than AI as addition.
Score the business value in dollars before deployment, not after. Forecast Year-1 + Year-3 cost honestly. Identify what you will retire to offset the new spend. Set a 4-week measurement window with cost AND value targets. Know what walking away looks like before you start. The AI initiative that survives a board review is the one where the dollar value justifies the cost at scale, not the one where the pilot cost was low and the production cost will be figured out later.
The Three Objectives page is the value side of the dollar-scoring discipline. This page is the cost side. Both feed into the same calculation: business value (Objective 1, 2, or 3 measured in dollars) minus total cost (this page) equals AI ROI. Vendors that pretend either side does not exist are vendors who fail at the board review. Read /ai-initiative-three-objectives-oil-gas for the value side; this page covers the cost side.
No. We are saying use AI deliberately. The 95 percent failure rate (MIT NANDA 2025) and the 94 percent no-significant-value rate (McKinsey State of AI 2025) come from initiatives that skipped the cost-and-value discipline. The 5 percent that succeed start with one clear objective in dollars, account for the cost honestly, and measure on the metric in 4 weeks. AI is not free. AI is also not optional for mid-tier operators who want to compound cash flow against their larger competitors. The discipline is what separates the two outcomes.
Start with the end in mind · score value AND cost · 4-week walk-away
Below VP signing authority on the entry tier. Bundle pricing that does not compound across loops. Single-architecture deployment so you do not pay separately for Operations, Engineering, Safety, and Maintenance integration. AI is not free. The discipline that makes it economical is the same discipline that makes it work.
24-hour reply · 4-week scope + pricing · the cost discipline pairs with the value discipline