the age of the industrial engineer: research report

01 · abstract and method

abstract

Every engineering era has been named for the discipline facing the hardest build problem of its day. This study tests whether that problem has moved. Industrial engineering, 117 years old, designs the system that produces things rather than the thing, and has never been an era's headline discipline. The argument here is that it was blocked by a constraint it could not touch: it could not build the machines its factories ran on. Published measurements from 2024 to 2026 show large software organizations reporting a quarter to more than 80% of new code model-authored, while delivery research over the same period finds throughput rising and stability falling together. That pairing is an industrial-engineering problem statement. One modeled AI-native company was then examined against its own record across five IE tool families. The verdict is qualified, and the era claim itself is an inference no instrument here measures.

what was read

Three evidence packs compiled for this study, all committed: a pack on IE history, definitions and employment data including its own verification pass; a pack on engineering eras and AI's effect on software labor; and an internal pack of the modeled company's operating artifacts, each fact carrying its tree, path and read date. External figures were located by search and, where the primary page was reachable, fetched before citing. Where a primary page returned 403, was paywalled, or was a secondary report of an unfetched paper, the figure carries [l] and is never presented as verified.

the observation boundary

The self-test in section 05 observes one modeled company: one human operator, two shipped consumer products plus a company layer, agents doing nearly all execution. The window is summer 2026, roughly 2026-05-18 through 2026-09-09; the cost instrumentation covers a 30-day slice, 2026-08-10 through 2026-09-08, across 224,572 model turns. A six-day unattended stretch inside that window is excluded from every before-and-after reading, on the source's own reasoning that it was a process stall rather than a saving.

Nothing here supports a claim about companies in general. Where a figure has no denominator, that is said rather than smoothed.

observer bias, and the control for it

The authors operate the company graded in section 05, hold the hypothesis under test, and the operator is himself an industrial engineer by training. That last fact is a selection effect on the result, not a coincidence beside it, and it is treated as a confound in section 06 rather than as credentialing.

The control is artifacts over impressions. Every internal figure was produced by reading a system of record over a stated window with a stated denominator, and is re-runnable. Anything not artifact-tied is labeled [estimated]. Unflattering readings travel with the flattering ones, and in this study the unflattering half is the more interesting one: a cost lever modeled to save nearly 40% saved nothing, and its corrected version made things worse before it made them better.

Confidentiality applies throughout section 05. Capabilities and measured outcomes appear; mechanisms, tool names, vendor names, program names and absolute spend do not. Cost is expressed as ratios, percentages and indices.

02 · what industrial engineering is

IISE, the discipline's professional body, founded 1948, describes its members this way: "Industrial and systems engineers make things better in any industry, from automobile manufacturing and aerospace, to healthcare, forestry, finance, leisure, and education" [v]. The BLS handbook is narrower: "Industrial engineers design, develop, and test integrated systems for managing industrial production processes" [v]. The classic longer definition, the one repeated across university and accreditation pages, describes IE as concerned with "the design, improvement and installation of integrated systems of people, materials, information, equipment and energy" [l]. Note the position of equipment in that list. It is one input among five, never the object of design.

IISE's own body of knowledge names 14 areas [v]: work design and measurement; operations research and analysis; engineering economic analysis; facilities engineering and energy management; quality and reliability engineering; ergonomics and human factors; operations engineering and management; supply chain management; engineering management; safety; information engineering; design and manufacturing engineering; product design and development; and system design and engineering.

Five tool families carry most of the weight for the argument here: optimization and operations research, statistics and process control, queueing and flow, simulation, and human factors. That compression is the authors' own, and the pack flags it as inference rather than as IISE's structure, since queueing and discrete-event simulation sit inside the broader operations areas rather than being separately named.

employment, pay and growth

All three rows fetched from the BLS Occupational Outlook Handbook [v]. Every row's projection carries the cycle label the pages themselves showed at fetch time, 2025 to 2035, not the 2024 to 2034 cycle a reader may expect. The software-developer row is the standalone "Software Developers" occupation, not BLS's combined "Software Developers, Quality Assurance Analysts, and Testers" group.

The OOH page (2025-35 cycle, fetched 2026-09-09) carries three separate reads: the combined group, "Software Developers, Quality Assurance Analysts, and Testers," at 1,905,400 jobs, $134,040 median pay, +10% growth; software developers alone at 1,717,800 jobs, $135,980 median pay, +10% growth; and quality assurance analysts and testers alone at 187,600 jobs, $104,300 median pay, +6% growth. The comparison table above uses the developers-alone reading throughout, since that is the closer occupational match to industrial engineers, and the combined-group total should not be read as software-developer headcount. Software developer employment is stated as a single figure, 1.72 million, not a range.

The counter-narrative point is worth stating against the thesis rather than for it. Industrial engineers are paid least of the three and are outnumbered by software developers roughly 5 to 1. The only column IE leads is growth rate, and that rate is off a much smaller base.

03 · the engineer of the era

the timeline

No comparable 19th-century engineering-degree count was found, and the pack recommends dropping rather than guessing one.

IE's own timeline

Hugo Diemer coined the term "industrial engineering" in 1900 and founded the world's first program at Penn State in 1908; the department was formally established in 1909 and graduated its first two students in 1910 [v]. Frederick Taylor published The Principles of Scientific Management in 1911 [v]. Frank and Lillian Gilbreth published Motion Study the same year, following it with Fatigue Study in 1916 and Applied Motion Study in 1917, and used motion-picture cameras to decompose work into elemental motions [l]. Walter Shewhart proposed the first control chart on 16 May 1924 [v], though the sources disagree on his employer at that date: one places him at Bell Labs, another at Western Electric's Hawthorne plant with the move to Bell Labs following in 1925.

Operational research emerged in Britain from about 1939, originally to integrate radar into RAF tactics, with P.M.S. Blackett forming analysis groups at Anti-Aircraft Command, Coastal Command and the Admiralty [v, partial]. Deming lectured Japanese industrial leaders at Hakone in 1950, to an audience estimated to represent 75% of Japan's industrial capital base at the time [v]. Taiichi Ohno built the Toyota Production System across roughly two decades from the late 1940s [l]. John Little proved L = λW in 1961, confirmed here through Little's own 50th-anniversary retrospective (Little, "Little's Law as Viewed on Its 50th Anniversary," Operations Research 59(3), pp. 536-549, 2011) rather than the original paper, which was behind a 403 on three attempts [v]. Goldratt published The Goal in 1984 [l, secondary only; the publisher's own pages name anniversary editions and never state the original year]. MIT's motor-vehicle program named the whole approach "lean" in 1990 [v].

Georgia Tech's program carries the argument in its own name. It is officially the H. Milton Stewart School of Industrial and Systems Engineering, abbreviated ISyE, and its undergraduate program has ranked first in the nation by U.S. News for 25 consecutive years (2026 Best Colleges, released 2025-09-23), its graduate program first for 35 years [v]. ISyE describes itself as training students to "design the systems behind the products and services that touch everyday life," with 1,387 undergraduates, 387 on-campus and 5,350 online master's students, and 193 doctoral students [v]. The "and systems" in the name is the discipline's own self-description, not this study's framing.

the method-not-machine thread

Read that list for what each item actually is. A control chart is a method for reading a machine's output. Little's Law is an identity that holds regardless of what the queue contains. Ohno is consistently described as having built TPS by rearranging machines that already existed, and Toyota's own internal description frames it as a resource-conserving framework rather than a how-to; Liker's chapter title on the subject is "Building in Quality Is a Principle, Not a Technology" [l]. Operational research was invented by physicists and biologists analyzing the deployment of equipment other people built.

One honest gap: no citable primary quote was found for the formulation "engineer of the system rather than the component." The thread above is argued from what the canon consists of, not from a founder's line.

the constraint claim, as a hypothesis

This is a hypothesis, and five counter-hypotheses were written down before the evidence was gathered and given equal weight: (a) IE was absorbed into "operations" once already and will be again; (b) software's near-zero reproduction cost breaks the factory analogy; (c) "the era of X engineer" is narrative rather than a measurable quantity; (d) one company, one operator and two products is not a factory floor; (e) software engineers simply become the systems engineers, with no discipline shift. Counter-hypothesis (a) has live support: inside INFORMS there is an unresolved debate about whether "operations research" and "industrial engineering" should be consolidated under newer names [l], and the field's own defensive literature includes a paper arguing IE is being extended rather than replaced [l], which is evidence the anxiety is real enough to need rebutting. Counter-hypothesis (c) is not resolved by this study and is restated as a limit in section 06.

04 · the constraint moved

Published measurements only. Each carries its date and its verification tier.

DORA's own explanation is the load-bearing sentence in the table: "AI accelerates software development, but that acceleration can expose weaknesses downstream. Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability" [v]. Read alongside the Stack Overflow trust decline and the METR trial's perception gap, the pattern is output up, variation up, trust down, and the control system missing. Those four terms are the first four chapters of a quality-engineering textbook.

The Stanford and BLS rows disagree in a way that matters and is not a contradiction. Aggregate software employment is still projected to grow, while the entry rung of AI-exposed occupations is measured shrinking relative to its own trend. The squeeze is at exactly the point where the job was writing the code.

robustness at the bottom of the range

The thesis should survive the least favorable defensible reading, so take the floor. Suppose the honest AI-authored share across large software organizations is 25%, Google's October 2024 figure (since risen to 75% per Google's own Cloud Next 2026 update), rather than 80%.

The thesis still holds, for a specific reason. DORA measured the stability penalty across roughly 5,000 professionals at 90% adoption, and at that population the per-company authored share was certainly nearer 25% than 80%. The control-system problem appeared at the low end of the range, not the high end. What the floor costs the argument is the strong version: at 25%, integration and review become a serious cost center inside conventional teams, but they do not become the whole job, and a company cannot be run by one person. The era claim weakens from "the discipline that runs the factory becomes the headline" to "a control discipline becomes mandatory equipment." At 80%, the strong version is available. The report claims the first and treats the second as conditional on the higher figures being sustained.

05 · the self-test

One modeled company, one operator, two products plus a company layer, summer 2026. Every figure carries its prior and the population it was measured on. No absolute spend appears; costs are indexed.

instance 1 · an ordered objective function

The company checks every non-trivial decision against one ranked formula: maximize users, then revenue without costing growth, also velocity measured as shipped customer value per week, subject to a fixed cash constraint. The check is written into the artifact carrying the decision as a one-line tag naming each term's direction and whether the decision is one-sided or a genuine fork. One-sided decisions are taken and reported; forks go to the operator.

This is the weakest instance in the set. Eleven instances of a rule adopted one day before the reading is an adoption count, not evidence the discipline holds.

instance 2 · flow, work in progress and backlog drains

Two rules operate here. A backlog is a stock and gets a one-time drain to target; a recurring cadence exists to hold the target and never to close a multi-week gap. The tell that the wrong instrument is in use is backlog divided by nightly rate landing in weeks.

The last row is the Little's Law point and the reason the drain is not the fix. Clearing a board without changing the arrival process buys roughly four weeks before the queue returns to its prior level.

instance 3 · statistical process control

Model output is validated against labelled ground truth rather than against whether it parsed. Deterministic rules live in code; the model extracts and judges, the code decides. A judging stage runs at temperature zero on a different model from the producer.

The 79% row is the finding an IE would care about most, and it argues against the company's own instincts: four fifths of a queue built for an agent lane did not need the agent lane.

instance 4 · simulation before the change ships

Five fixed personas run any user-facing change before code exists. The panel is fixed at five deliberately, on the reasoning that more draws from one model are correlated and manufacture false consensus. A clean 5 of 5 with no enacted friction is read as a yellow flag and re-run adversarially.

instance 5 · cost per unit, and the J-curve

The company modeled a saving from one configuration change, then measured it rather than assuming it from the diff. The first attempt did nothing. The corrected version made the number worse before the setting was retuned from the data.

The J-curve is the shape worth carrying out of this section. A lever that looks free gets worse before it gets better, because the naive model omitted a real cost driver, the equivalent of a changeover cost inside a lot-size decision. The response was to correct the model rather than the data.

the human-touch ratio

Under ad-hoc prompting, the company measured roughly 22 agent actions per human touch. Under an orchestrated program covering the same class of work, roughly 200 agent actions per human touch, which is roughly 9x. Populations differ and this is the weakest comparison in the section: the 200:1 figure comes from one program of 42 stages, 3,191 tool calls, 15 interruptions and 1 confirmation; the 22:1 figure is a median across 128 separate windows. Treat it as an order of magnitude, not a measurement.

06 · what this comparison cannot tell you

No control company. A conventional twin, same objectives, same window, does not exist. Every before-and-after in section 05 compares the company to its own past, which changed for many reasons at once.

The operator is an industrial engineer. The company applies IE tools because the person running it was trained in them. That makes section 05 an existence proof that the tools work in this setting, and no evidence at all about whether they are the tools a differently trained operator would have reached for.

The instruments moved during the window. The cohort study underpinning the cost-per-deliverable row published a correction pass that revised or reversed three of its own headline figures. Only corrected numbers appear here, but a corpus that needed correcting once may need it again.

Simultaneous changes. Board hygiene, routing policy, orchestration and the objective-function rule all changed inside four months. Attribution across them is contaminated, and one adjacent series the sources refuse to attribute at all is CI failure rate, which ran 73.8%, then 44.3%, then 57.5% across three weeks.

Young products. Throughput comparisons run against a young baseline on small boards, where a single process change moves a percentage that would not move at scale.

honest limits

Software has near-zero reproduction cost, so the factory analogy bends exactly where a physical plant would talk about finished-goods inventory. A defect in generated prose is fuzzier than a defective part, and much of IE's canon assumes a measurable specification. n = 1, one operator, two products, four months. Aggregate software employment is still projected to grow 10%, and the labor evidence for the thesis is confined to the entry rung. Industrial engineering was absorbed into "operations" once already, and the field's own professional body is currently debating whether its name still serves. And the simplest counter-explanation remains available: software engineers may just become the systems engineers, absorbing the control work without any discipline changing hands.

07 · conclusion

falsification tests

Four observations would refute or seriously damage the thesis, and each is checkable.

• IE employment growth reverting below software's in the 2027 BLS projection cycle. The current 12% against 10% is the only quantitative edge IE holds. If the next cycle inverts it, the demand claim loses its one number.
• DORA 2026 finding delivery stability recovering at higher AI adoption without corresponding investment in control systems. That would mean the models absorb their own variation and no control discipline is required.
• AI-authored share plateauing under 40% at the named companies through 2027. Anthropic's stated expectation is crossing 90% by end of 2026; a plateau under 40% would put the whole range at the floor described in section 04 and reduce the claim to its weak version.
• METR's time-horizon doubling reverting past 12 months, or a replication of the July 2025 trial finding no slowdown effect. The first would mean the authoring constraint is not clearing; the second would mean the perception gap that motivates the control argument is not real.

A fifth, internal: the self-test's cost per deliverable failing to stay below the prompt-driven cohort over a longer window than one program.

Review date: 2027-03. By then the 2027 BLS cycle, the 2026 DORA report and Anthropic's end-2026 figure will all have landed, which settles tests 1, 2 and 3 together.

08 · sources

fetched and verified this study

• IISE, About: https://www.iise.org/details.aspx?id=282
• IISE, Industrial and Systems Engineering Body of Knowledge: https://www.iise.org/Details.aspx?id=43631
• BLS OOH, Industrial Engineers: https://www.bls.gov/ooh/architecture-and-engineering/industrial-engineers.htm
• BLS OOH, Software Developers (fetched twice, 2025-35 cycle: standalone and combined-with-QA reads both recorded): https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm
• BLS OOH, Computer Hardware Engineers: https://www.bls.gov/ooh/architecture-and-engineering/computer-hardware-engineers.htm
• NCES Digest, Table 325.35, computer and information sciences degrees: https://nces.ed.gov/programs/digest/d23/tables/dt23_325.35.asp
• NASA, The First Step: Langley's Contributions to Apollo: https://www.nasa.gov/history/the-first-step-langleys-contributions-to-apollo/
• Penn State IME, department history: https://www.ime.psu.edu/department/history.aspx
• Taylor, The Principles of Scientific Management, 1911: https://www.gutenberg.org/ebooks/6435
• Lean Enterprise Institute, The Machine That Changed the World: https://www.lean.org/store/book/the-machine-that-changed-the-world/
• Shewhart control chart centenary: https://www.leanblog.org/2024/05/celebrating-100-years-of-shewharts-control-charts-a-century-of-quality-management/
• Shewhart, 16 May 1924 memorandum, and the Hawthorne attribution: https://pmc.ncbi.nlm.nih.gov/articles/PMC2464836/
• Deming Institute, 1950 speech to Japanese business leaders: https://deming.org/speech-by-dr-deming-to-japanese-business-leaders-in-1950/
• INFORMS, The Origins of OR: https://www.informs.org/Explore/History-of-O.R.-Excellence/Bibliographies/The-Origins-of-OR
• Little, "Little's Law as Viewed on Its 50th Anniversary," Operations Research 59(3), pp. 536-549, 2011: https://people.cs.umass.edu/~emery/classes/cmpsci691st/readings/OS/Littles-Law-50-Years-Later.pdf
• Georgia Tech ISyE, undergraduate ranking, 25th consecutive year: https://www.isye.gatech.edu/news/isye-earns-no-1-undergraduate-ranking-25th-consecutive-year
• Georgia Tech, ISyE graduate program ranking: https://www.gatech.edu/news/2025/04/08/h-milton-stewart-school-industrial-and-systems-engineering-graduate-program-ranked
• Georgia Tech ISyE, program overview and enrollment: https://www.isye.gatech.edu/
• Anthropic Economic Index, September 2025 report: https://www.anthropic.com/research/anthropic-economic-index-september-2025-report
• METR, early-2025 AI and experienced developer productivity: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• METR, measuring AI ability to complete long tasks: https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ and https://metr.org/time-horizons/
• Google Cloud, announcing the 2025 DORA report: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report
• Alphabet Q3 2024 earnings message, Pichai: https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2024/
• Google Cloud Next 2026, Pichai: https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai/
• Anthropic, "When AI builds itself" (Anthropic Institute, recursive self-improvement paper): https://www.anthropic.com/institute/recursive-self-improvement
• ASEE, "Profiles of Engineering and Engineering Technology, 2022" (403 on direct WebFetch; fetched via curl -A "Mozilla/5.0"): https://ira.asee.org/by-the-numbers/engineering-undergraduate-students/
• METR, Feb 2026 update to the developer-productivity experiment: https://metr.org/blog/2026-02-24-uplift-update/
• Stanford Digital Economy Lab, "Canaries in the Coal Mine" (publication page and Aug 2026 update): https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ and https://digitaleconomy.stanford.edu/news/canariesaug26/

corroborated via search results

• SEBoK, systems engineering and industrial engineering: https://sebokwiki.org/wiki/Systems_Engineering_and_Industrial_Engineering
• ABET, criteria for accrediting engineering programs: https://www.abet.org/accreditation/accreditation-criteria/criteria-for-accrediting-engineering-programs-2025-2026/
• Gilbreths, motion study: https://uen.pressbooks.pub/ompeople/chapter/frank-and-lillian-gilbreths/ and https://web.mit.edu/allanmc/www/TheGilbreths.pdf
• Henry Gantt and the Gantt chart: https://en.wikipedia.org/wiki/Henry_Gantt and https://www.apm.org.uk/blog/a-brief-history-of-gantt-charts/
• Taiichi Ohno and TPS: https://en.wikipedia.org/wiki/Taiichi_Ohno and https://www.qad.com/blog/2018/03/taiichi-ohno-toyota-production-system
• Toyota Production System: https://en.wikipedia.org/wiki/Toyota_Production_System
• Goldratt, The Goal, 1984: https://en.wikipedia.org/wiki/The_Goal_(novel) and https://northriverpress.com/the-goal/
• Little's Law: https://en.wikipedia.org/wiki/Little's_law
• ASCE: https://www.asce.org/about-asce and https://en.wikipedia.org/wiki/American_Society_of_Civil_Engineers
• ASME engineering history: https://www.asme.org/about-asme/engineering-history
• Formation of IEEE from AIEE and IRE: https://ethw.org/Formation_of_IEEE_by_the_Merger_of_AIEE_and_IRE
• Apollo workforce, secondary: https://www.ketk.com/news/top-stories/apollo-11-moon-landing-had-400000-people-working-behind-scenes/
• Andreessen, "Why Software Is Eating the World," 2011: https://a16z.com/why-software-is-eating-the-world/
• Nadella on AI-written code at Microsoft: https://techcrunch.com/2025/04/29/microsoft-ceo-says-up-to-30-of-the-companys-code-was-written-by-ai/ and https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html
• Stack Overflow Developer Survey 2025: https://survey.stackoverflow.co/2025/ai and https://stackoverflow.blog/2025/12/29/developers-remain-willing-but-reluctant-to-use-ai-the-2025-developer-survey-results-are-here/
• Grand challenges in industrial and systems engineering: https://www.tandfonline.com/doi/full/10.1080/00207543.2024.2432463
• Rediscovering scientific management, industrial engineering to industrial data science: https://polipapers.upv.es/index.php/IJPME/article/view/16617
• INFORMS naming debate: https://pubsonline.informs.org/do/10.1287/orms.2023.04.06/full/
• ISE Magazine, September 2025 AI feature: https://www.iise.org/isemagazine/details.aspx?id=55976
• Reinertsen, The Principles of Product Development Flow, 2009: https://xp123.com/the-principles-of-product-development-flow-reinertsen/
• Anderson, Kanban, 2010: https://kanbanbooks.com/discovering-kanban-david-anderson-article/
• Forsgren, Humble and Kim, Accelerate, 2018: https://itrevolution.com/product/accelerate/
• Kim, Behr and Spafford, The Phoenix Project, 2013, and its debt to The Goal: https://itrevolution.com/articles/beyond-phoenix-project-making-book-goldratt-deming/