Assess whether an automation deployment can produce durable unit economics rather than an impressive equipment count.
Robot counts measure installed equipment, not economic output. The International Federation of Robotics calculated 4,281,585 industrial robots in operation in 2023, while also noting that customer-industry information was unavailable for 17% of that year's installations [1]. Amazon's 2023 filing described fulfillment productivity as a function of volume, staffing, inventory placement, shipment mix and network design [2]. OSHA guidance adds scheduled and unscheduled maintenance, malfunctions and recovery procedures to the operating picture [3]. A useful deployment case must therefore bridge equipment, productive hours, service labor, exceptions, safety and cash—not stop at machines purchased.
Define the productive unit
Start with the output the business buys: accepted picks, welded assemblies, inspected units or completed moves. Count a robot only when it is commissioned for that output. Then measure scheduled hours, available hours, productive cycle time, first-pass yield and downstream acceptance. A fleet utilization average can hide one constrained cell and many idle machines, so retain cell-level distributions and the p10 as well as the mean.
Separate technical uptime from economic utilization. A machine can be healthy while waiting for material, an operator, a tool change or downstream space. IFR's stock and installation series is useful for adoption context [1], but it cannot establish site productivity, payback or even the share of machines doing a particular task. The operating model supplies those missing denominators.
- Output: accepted units per staffed hour.
- Service: planned and unplanned technician hours per operating hour.
- Recovery: median and tail time from fault to stable production.
- Exceptions: manual touches and rework per 1,000 attempted units.
Model the full cost to serve
Annualize equipment, integration, tooling, guarding, floor changes and software over a realistic useful life. Add energy, consumables, spares, vendor support, internal maintenance, supervision, training and the labor that handles exceptions. Include lost contribution during planned service and ramp. Amazon's filing shows why a single automation variable is insufficient: its fulfillment expense is affected by sales volume, product mix, third-party participation, network expansion, productivity and accuracy [2].
Worked example, explicitly hypothetical: a cell costs $1.2 million installed and is expected to run 4,000 scheduled hours. At 80% technical availability, 75% economic utilization and 95% first-pass yield, productive time is 2,280 hours. If accepted throughput is 40 units per productive hour, annual accepted output is 91,200 units. Add $180,000 service and software, $140,000 exception labor, $60,000 energy and consumables, and $240,000 annualized capital. The automation cost is $6.80 per accepted unit before displaced labor or capacity value. At 55% utilization, it rises sharply. The decision turns on utilization and recovery, not the number of robots.
Gate expansion on recovery evidence
Require a pilot to survive representative product mix, shift changes and fault conditions. Report the percentage of incidents resolved by frontline staff, specialist response time and recurring fault classes. NIOSH's 1988 guidance frames corrective robot maintenance as human intervention into work zones and distinguishes availability, maintainability and downtime data [3]. Treat that historical guidance as a floor, then apply current site requirements and standards. Safe recovery controls are not overhead outside the model; they are part of usable capacity.
Scale only after three tests pass: accepted unit cost beats the alternative across a downside utilization case; service staffing grows slower than deployed capacity; and recovery performance remains within the production buffer. If savings depend on perfect uptime or zero exceptions, record the project as technically demonstrated but economically unproven.
Take it into the meeting
- Measure accepted output, not installed machines.
- Price service, exceptions and downtime into every unit-cost case.
- Scale only after recovery performance holds under representative load.
Sources & boundaries
Source statements are attributed; the decision process is Signal Atlas analysis. Examples marked hypothetical are teaching inputs, not observed outcomes.
- The worked economics are hypothetical and exclude taxes and financing structure.
- IFR aggregate data does not measure site-level returns.
- Safety requirements depend on the application and jurisdiction.
- World Robotics 2024 — Industrial Robots: Executive SummaryInternational Federation of Robotics · Source publication: not established · Retrieved 2026-09-19
2023 industrial robot installations and operating stock Customer-industry data gap Aggregate adoption is not a site economics measure
- Amazon.com, Inc. Form 10-K for the year ended December 31, 2023U.S. Securities and Exchange Commission · Source publication: 2024-02-02 · Retrieved 2026-09-19
Fulfillment cost drivers Network complexity and productivity dependencies Useful-life and operating-cost considerations
- Safe Maintenance Guidelines for Robotic WorkstationsNational Institute for Occupational Safety and Health · Source publication: not established · Retrieved 2026-09-19
Maintenance as human intervention in robot work zones Safeguards and procedures during corrective work Availability, maintainability and downtime as distinct measures