Amazon's AI power project exposes a buyer's blind spot: measure AI by cost per completed job, not model price or call volume.
Amazon is backing a massive gas power plant in West Texas to support an AI data center, according to Ars Technica. Reports describe a 7.65 GW facility in Pecos County with a permit allowing more than 30 million metric tons of greenhouse gas emissions annually, although actual output may be lower.
The important business story is not an argument about Amazon's climate commitments. It is that AI infrastructure now has a bill large enough to see.
That should change how we buy operational AI.
Most buyers still compare vendors using model pricing, monthly license fees, and impressive activity numbers. They ask how many calls an agent can answer, how quickly it can generate a quote, or how many invoices it can process. Those metrics are easy to demonstrate and easy to misunderstand.
The better question is simple: what does the system cost per completed job, and what waste does it remove along the way?
The model price is the easy number
AI vendors usually make pricing legible. You get a per-seat fee, a per-minute voice charge, a usage tier, or a token rate. Those numbers matter, but they are rarely the full cost of deploying automation inside a real operating business.
The expensive work often sits around the model:
- Connecting the system to scheduling, telephony, CRM, accounting, and payment tools
- Cleaning and maintaining customer, vendor, service, and pricing data
- Reviewing uncertain outputs and correcting bad records
- Designing workflows for exceptions, cancellations, reschedules, and incomplete job details
- Training staff to supervise the system and handle escalations
- Paying for duplicate tools when the new platform cannot replace the old ones
- Managing usage spikes when storms, heat waves, or seasonal demand hit
- Fixing the operational damage when an automated action is technically successful but commercially wrong
Research on enterprise AI costs consistently points in the same direction. Data readiness, integration, governance, and ongoing support can outweigh the model line item. For a local service business, the same principle applies at a smaller scale. A $500 monthly subscription can become an expensive system if it creates extra review work, duplicate dispatches, or more customer callbacks.
A cheap model does not make an inexpensive operation.
Count the work the system creates
The common ROI mistake is counting completed actions instead of completed outcomes.
An AI receptionist can answer 1,000 calls. That does not mean it created 1,000 valuable jobs. It may have handled robocalls, repeated existing requests, or captured incomplete information that someone still had to reconstruct manually.
A dispatch system can assign 500 jobs. That does not mean it improved utilization. If technicians arrive without access instructions, drive across town for low-value work, or need a second visit because the initial diagnosis was weak, the automation may have increased cost while improving a dashboard metric.
A quoting tool can produce hundreds of estimates. That does not mean revenue increased. The relevant question is how many quotes became profitable, completed jobs.
For operational AI, the denominator should be the completed job. Track:
- Cost per completed job
- Administrative minutes per completed job
- Calls required to book and confirm a job
- Dispatch changes per completed job
- Miles or drive time caused by poor routing or incomplete information
- Repeat visits caused by missing job details
- Quote-to-book rate by source and service type
- Invoice exceptions and payment delays after the job is complete
This reframes automation around the physical business. A local service company does not sell AI interactions. It sells hauling, repairs, installations, maintenance, and other work that has to happen in the field.
Ask vendors for an operating bill
The next AI vendor evaluation should look more like a job-costing exercise than a software demo.
Ask the vendor to model your current workflow using your actual volumes. Do not accept a generic example based on idealized calls or clean data. Bring a representative sample of service requests, including emergencies, cancellations, repeat customers, unusual access requirements, and jobs that require vendor coordination.
Then require answers to these questions:
- What new costs appear after deployment?
Include implementation, integration, telephony, usage, monitoring, human review, data cleanup, and support. If the vendor cannot separate recurring costs from one-time costs, the estimate is not ready for approval.
- Which manual steps disappear, and which move somewhere else?
Automation often shifts work from dispatch to accounting, from reception to field technicians, or from internal staff to customers. A task is not eliminated because it changed owners.
- What happens when the system is uncertain?
This is not just a reliability question. It is a labor and cost question. Does uncertainty create a review queue? Who owns it? How long does review take? What is the cost of delaying the job?
- What is the expected effect on completed-job economics?
Ask for a before-and-after model covering booking rate, travel, dispatch time, repeat visits, administrative effort, and collections. The vendor should be willing to state which assumptions drive the result.
- How will we prove the result after 90 days?
Define the baseline before launch. If you do not know your current average response time, quote-to-book rate, dispatch edits, or administrative minutes per job, you cannot honestly claim improvement later.
ROI needs an audit trail
This extends the argument in The DOGE Report's Lesson: Show Your AI Work. A savings claim is not evidence. The evidence is a traceable change in operating results.
For example, suppose a company says its AI system saved 80 staff hours in a month. We should ask what those hours represent. Were they removed from the workflow, or did employees spend the time checking AI output? Did the business complete more jobs, reduce overtime, or improve collections? If not, the number may describe activity rather than savings.
A useful review ties every claimed benefit to an operational measure:
- Fewer missed calls should produce more qualified bookings
- Faster quoting should produce more booked work or less idle capacity
- Better dispatch should reduce travel, delays, or repeat visits
- Cleaner job data should reduce invoice exceptions and collection time
- Automated follow-up should improve payment speed without increasing customer complaints
If the chain breaks, the ROI claim needs revision.
The practical buying standard
Do not ask whether an AI vendor is affordable in isolation. Ask whether the full workflow is cheaper, faster, and more predictable with the system than without it.
Start with one service line or geography. Record a baseline for 30 days. Deploy the automation with explicit cost categories and exception tracking. Review results by completed job, not by AI activity. Include the work your staff still performs around the system. Then expand only when the economics survive contact with real operations.
Reeve is built around this operating view, connecting calls, dispatch, quotes, invoices, and QuickBooks so the business can measure the workflow rather than a single AI feature.
Amazon's power project makes the physical cost of AI visible. Your vendor should make the operational cost visible too. Ask for the bill, define the denominator, and measure what reaches completion.
