[Manufacturing](https://bfcs.ai/industry/manufacturing/) · [Faster](https://bfcs.ai/faster/)

# Can AI Increase Line Throughput?

A production line runs only as fast as its slowest step, and that step can shift. Machine data can find it, and AI scheduling promises to keep it fed. The detection evidence is real but small, the headline throughput gains are company-reported, and the payoff depends on a plant that already measures itself.

Updated **October 2026** · 7 min read · Evidence-graded · 8 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Moderate confidence | Finds the constraint; does not fix it · Peer-reviewed tests on three automotive lines found bottlenecks from existing machine logs and split each bottleneck's active time into producing, set-up and down states. The fix still falls to production and maintenance teams. |
| Faster | Conditional · Low to moderate confidence | Where the line is instrumented and the bottleneck is acted on · We found no independent study that measured a throughput gain on a real line after AI detection or scheduling. The largest figure we found, a 73% efficiency gain at a World Economic Forum Lighthouse site, comes from a Foxconn release republished by a trade site. It is company-reported and not independently verified. |
| Cheaper | Conditional · Low confidence | Pays only with the complements in place · A US Census-based study of more than 30,000 plants, read from its working-paper abstract, links predictive analytics to higher productivity only alongside IT capital, educated workers or flow-efficient production. These are cross-plant estimates, not costed line results. |
| Safer | Conditional · Governance-gated | Pushing the constraint harder needs limits set by people · We found no study measuring injury or quality outcomes from AI scheduling. Maintenance windows, isolation and staffing limits have to be fixed constraints, owned by a named planner, not costs the model can trade away. |

## The problem, in numbers

On one official measure, Australian manufacturing became less productive in 2024-25. The Australian Bureau of Statistics estimates that the sector's multifactor productivity, the output it gets from its combined labour and capital, fell 2.9% in 2024-25 on an hours-worked basis, while the market sector as a whole fell 0.5%. (Source 2)

AI use is not yet widespread among manufacturers. The ABS business survey for 2024-25 estimates that 9% of manufacturing businesses used AI, against 12% of all businesses, 18% in mining and 38% in information, media and telecommunications. (Source 1) The ABS chart note leaves the base of the industry shares ambiguous, so we read them as a ranking rather than a precise count. In our reading, AI scheduling depends on data habits that are also uncommon. Across all industries, the survey found 10% of businesses actively collected or analysed data to make decisions and 6% used predictive analysis to plan, down from 24% and 18% in 2021-22. (Source 1)

### Business use of AI by industry, 2024-25 (Source 1)

Information, media and telecommunications · 38%

Mining · 18%

All businesses · 12%

Manufacturing · 9%

Construction · 6%

Survey estimates as published by the ABS for selected industries. Manufacturing sits below the all-business figure; see the source note on the base of the industry shares.

### The evidence, in four numbers

9% · of Australian manufacturing businesses reported using AI · 2024-25 survey estimate; 12% of all businesses; base of industry shares not stated (ABS)

\-2.9% · manufacturing multifactor productivity growth · 2024-25, hours-worked basis; market sector fell 0.5% (ABS estimate)

14 · bottleneck detection methods found in a systematic review · run by walking the floor, simulation or data science (Skoogh et al., 2023)

73% · production efficiency gain claimed for one Lighthouse factory · company-reported in a Foxconn release, republished; no baseline or method

## What improves, and why

A line's throughput is constrained by one or more bottlenecks, the machines or stations that limit how fast everything else can run. (Source 3) The practical difficulty is that real factories have complex dynamics, and a large plant may have several hundred machines running at once, as one review of the field puts it. (Source 4)

On the lines in one study, a manufacturing execution system (MES) logged when each machine was producing, setting up, down, blocked or starved. Researchers at Chalmers University of Technology built an algorithm that turns such logs into an "active period" for each machine, picks out the machine or group that is busy most of the time, and tests whether the difference is statistically significant. (Source 3) It then splits the bottleneck's active time by state, so a team can see whether the fix is a faster cycle or better uptime. On a car body assembly line, the bottleneck station was producing for 55.62% of its active time and down for 44.38%, so downtime was a large target alongside cycle time; the authors leave that choice to production and maintenance teams. (Source 3)

Scheduling is the second lever. In our reading, a good schedule keeps the bottleneck fed and spares it avoidable changeovers. Reinforcement learning and other AI methods aim to produce schedules that adapt as jobs arrive and machines fail, where traditional approaches often fall short when conditions change quickly. (Source 5)

> A faster machine that is not the bottleneck only makes a longer queue. The value of AI on a line is in knowing which machine to fix this week.

## What the evidence actually says

Detection works on real data, at small scale. The Chalmers algorithm was tested on three automotive lines: a 12-machine engine machining line over 62 production runs, a five-station assembly line over 40 runs, and a 26-machine component line with five gantries over 31 runs. (Source 3) Validation meant that production and maintenance experts judged the results met their needs. The paper reports no throughput result after any bottleneck was addressed. (Source 3) In our reading, it shows the constraint can be found from data a plant already holds, not how much output fixing it adds.

The wider literature is broad but uneven. A 2023 systematic review identified 14 bottleneck detection methods, run through three modes: walking the floor, discrete event simulation and data science. (Source 4) The Chalmers authors noted in 2018 that most detection approaches had been validated only in simulation. (Source 3) For AI scheduling, a 2025 review of reinforcement learning in dynamic job shops lists scalability, interpretability, data availability and the lack of standard performance metrics as unresolved. (Source 5)

The best large-scale evidence concerns prediction in general, not throughput. Working with the US Census Bureau, researchers surveyed more than 30,000 manufacturing plants and reported productivity significantly higher among those using predictive analytics, with up to US$918,000 higher sales than non-adopters using similar inputs. We could read only the working-paper abstract, not the published journal version. (Source 6) The gains appeared only alongside complements: significant IT capital, educated workers, or workplaces designed for flow-efficient production. (Source 6) OECD economists, summarising their own paper on an OECD blog, compile selected firm-level studies that put productivity gains from pre-generative AI at up to 10%, comparable to earlier digital technologies. (Source 7) Both are statistical associations across plants or firms, although the Census authors argue their estimates suggest a causal link. Neither measures output on a line.

The biggest throughput-style claims come from companies. A Foxconn release, republished by a trade site, says its Ingrasys factory in Taoyuan, named a World Economic Forum Lighthouse in December 2023, was credited with a 73% production efficiency improvement after applying AI to forecasting, warehouse and production scheduling, product design, quality and testing. (Source 8) That is company-reported data, not independently verified, without a published baseline or method, and it bundles scheduling with other changes.

## Where it falls short

-   **History comes first.** The Chalmers method needs enough time-stamped machine data to represent steady-state running, and its insight stops at the states the MES records. (Source 3)
    
-   **Utilisation is not the whole story.** The same authors note their method finds bottlenecks by utilisation only, while a machine can also constrain a line through quality. (Source 3)
    
-   **Few businesses report using their data.** In 2024-25, an estimated 10% of Australian businesses across all industries actively analysed data for decisions, and 16% named insufficient staff skills as a limit on their use of ICT. (Source 1)
    

## Where governance decides the outcome

A scheduler that drives the constraint harder changes how machines and people are loaded, so the decisions around it need owners. In our reading, three controls matter. A named planner owns the schedule, and the system stays advisory until it has been compared with the existing plan on the same orders. Maintenance windows, isolation procedures and staffing limits are written into the model as fixed constraints, not costs to be optimised away. And every claimed gain is measured against a baseline recorded before go-live, so a quiet month is not credited to the software. We found no study measuring injury or quality outcomes from AI scheduling, so the safety case rests on these controls, not on evidence.

## How to think about adoption

Start with the data you already have. If the line's MES or controllers log machine states with time stamps, a bottleneck analysis can be run before any AI is bought, and it can show whether the constraint is cycle time, downtime or set-ups. Record throughput, overall equipment effectiveness and the bottleneck's share of active time as a baseline. Fix the constraint the analysis names, then run it again, because the bottleneck may move. The Census evidence ties the payoff from prediction to complements, flow-efficient production among them. (Source 6) In our reading, high-mix job shops may still gain from smarter scheduling, but we found less evidence for them. Run any AI schedule beside the planner's before it takes over.

## How we read this evidence

BFCS.ai does not fabricate figures. Detecting a bottleneck and raising throughput are different claims; the peer-reviewed evidence here supports the first. The productivity studies are associations across plants and firms, and the largest throughput-style figure is company-reported data, republished. We label each kind and grade confidence by the weakest link.

## Sources

1.  Australian Bureau of Statistics, "Characteristics of Australian Business, 2024-25 financial year," released 25 June 2026 (survey of employing businesses; exclusions apply). [abs.gov.au](https://www.abs.gov.au/statistics/industry/technology-and-innovation/characteristics-australian-business/latest-release)
    
    **What we took from it:** 12% of businesses used AI in 2024-25. Chart "Business use of Artificial intelligence (AI), by industry, 2024-25": Manufacturing 9%, Mining 18%, Construction 6%, Information, Media and Telecommunications 38%. The chart note reads "Proportions are of all businesses that reported using AI in each output category," which leaves the base ambiguous; we report the shares as published. 10% of businesses actively collected or analysed data to make informed decisions (24% in 2021-22) and 6% used predictive analysis to plan business activities (18% in 2021-22); 16% named insufficient staff skills and capabilities as a factor that prevented or limited their use of ICT (the base is not stated); 55% of manufacturing businesses were innovation-active.
    
2.  Australian Bureau of Statistics, "Estimates of Industry Multifactor Productivity, 2024-25 financial year," released 6 February 2026. [abs.gov.au](https://www.abs.gov.au/statistics/industry/industry-overview/estimates-industry-multifactor-productivity/latest-release)
    
    **What we took from it:** Manufacturing multifactor productivity growth in 2024-25 was -2.9% on an hours-worked basis (table); market sector multifactor productivity fell 0.5% and market sector labour productivity fell 0.2%.
    
3.  Subramaniyan, M., Skoogh, A., Salomonsson, H., Bangalore, P., Gopalakrishnan, M. & Sheikh Muhammad, A., "Data-driven algorithm for throughput bottleneck analysis of production systems," Production & Manufacturing Research 6(1), 2018, pp. 225-246 (peer-reviewed, open access; version of record read on the Chalmers repository). [research.chalmers.se](https://research.chalmers.se/en/publication/504761)
    
    **What we took from it:** Active period algorithm on manufacturing execution system data with t-tests for statistical significance, tested on three automotive lines: a 12-machine engine machining line (62 production runs), a five-station car body assembly line (40 runs) and a 26-machine component line with five gantries (31 runs); validation was against the requirements of production and maintenance experts. On the assembly line the bottleneck station S2 was producing for 55.62% and down for 44.38% of its active period. The authors note most existing approaches were "limited to validation in a discrete event simulation-based environment"; the algorithm needs sufficient time-stamped history representative of steady-state behaviour, is limited to the data types in the MES, and detects bottlenecks from a utilisation perspective only. No post-intervention throughput result is reported.
    
4.  Reviews of bottleneck analysis, both peer-reviewed and read as abstracts on the Chalmers research portal. [research.chalmers.se](https://research.chalmers.se/en/publication/538741)
    
    Subramaniyan, M., Skoogh, A., Bokrantz, J., Sheikh, M. A., Thürer, M. & Chang, Q., "Artificial intelligence for throughput bottleneck analysis: State-of-the-art and future directions," Journal of Manufacturing Systems 60, 2021, pp. 734-751: [research.chalmers.se](https://research.chalmers.se/en/publication/525855)
    
    **What we took from it:** Skoogh, A., Thürer, M., Subramaniyan, M., Matta, A. & Roser, C., "Throughput bottleneck detection in manufacturing: a systematic review of the literature on methods and operationalization modes," Production & Manufacturing Research 11(1), 2023, article 2283031: identifies 14 bottleneck detection methods, grouped by whether they use queue states, process states or both, and three operationalisation modes (gemba walk, discrete event simulation and data science).
    
    factories may have "several hundred machines operating at any given time"; earlier work relied mainly on analytical models and discrete event simulation; research is sorted into identify, diagnose, predict and prescribe.
    
5.  Ngwu, C., Liu, Y. & Wu, R., "Reinforcement learning in dynamic job shop scheduling: a comprehensive review of AI-driven approaches in modern manufacturing," Journal of Intelligent Manufacturing 37(3), 2026, pp. 1093-1108, published online 8 March 2025 (peer-reviewed; abstract read on the University of Glasgow repository). [eprints.gla.ac.uk](https://eprints.gla.ac.uk/349209)
    
    **What we took from it:** Traditional approaches "often fall short when faced with rapid changes and high computational complexity"; the abstract names "sudden job arrivals, equipment failures, and fluctuating demands"; "challenges remain in algorithm scalability, interpretability, data availability, and standardization of performance metrics."
    
6.  Brynjolfsson, E., Jin, W. & McElheran, K., "The power of prediction: predictive analytics, workplace complements, and business performance," Business Economics, published 22 September 2021, DOI 10.1057/s11369-021-00224-5 (peer-reviewed). [digitaleconomy.stanford.edu](https://digitaleconomy.stanford.edu/publication/the-power-of-prediction-predictive-analytics-workplace-complements-and-business-performance)
    
    and [eurekalert.org](https://www.eurekalert.org/news-releases/936776)
    
    **What we took from it:** We could not reach the journal version; figures are from the working-paper abstract (29 June 2021) on the Stanford Digital Economy Lab page and from the University of Toronto Rotman School release of 2 December 2021. With the US Census Bureau the authors surveyed over 30,000 manufacturing establishments (survey years 2010 and 2015); productivity was significantly higher among plants using predictive analytics, with up to $918,000 (US dollars) higher sales than non-adopters using similar inputs (the release gives about $500,000 to $1 million); instrumental variables estimates and timing of gains "suggest a causal relationship"; the payoff occurs only with complements: significant IT capital, educated workers, or workplaces designed for flow-efficient production.
    
7.  Filippucci, F., Gal, P., Jona-Lasinio, C., Leandro, A. & Nicoletti, G., "Artificial intelligence: Promises and perils for productivity and broad-based economic growth," OECD ECOSCOPE blog, 16 April 2024, summarising OECD Artificial Intelligence Papers No. 15. [oecdecoscope.blog](https://oecdecoscope.blog/2024/04/16/artificial-intelligence-promises-and-perils-for-productivity-and-broad-based-economic-growth/)
    
    **What we took from it:** Figure 1, an authors' compilation from micro-level studies: firm-level productivity gains from pre-generative AI are comparable to previous digital technologies, up to 10% (selected estimates; we did not read the underlying OECD paper); AI adoption is currently low, under 5% of US firms.
    
8.  Company data (company-reported and republished, not independently verified): [iconnect007.com](https://iconnect007.com/article/139065/foxconns-ingrasys-honored-as-worlds-first-ai-server-lighthouse-factory/139062/ein)
    
    **What we took from it:** Foxconn, "Foxconn's Ingrasys Honored as World's First AI Server Lighthouse Factory," Foxconn-bylined release republished by I-Connect007, 26 December 2023. The Taoyuan NanChing factory was named a World Economic Forum Lighthouse on 14 December 2023; the release presents, as the WEF's evaluation, a 73% production efficiency improvement, a 97% product defect rate reduction, 21% shorter delivery time and 39% lower per-unit manufacturing cost, from introducing AI in order forecasting, warehouse and production scheduling, product design, quality and assembly testing. No baseline, period or method is given. The World Economic Forum pages could not be read (automated access refused).
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Cut Scrap and Waste?](https://bfcs.ai/articles/can-ai-cut-scrap-and-waste/): More in Manufacturing · MANUFACTURING · CHEAPER · 7 min · Read →
- [Can AI Predict Maintenance Needs?](https://bfcs.ai/articles/can-ai-predict-maintenance-needs/): More in Manufacturing · MANUFACTURING · BETTER · 7 min · Read →
- [Can AI Cut Clinical Documentation Time?](https://bfcs.ai/articles/can-ai-cut-clinical-documentation-time/): Also Faster · HEALTHCARE · FASTER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-increase-line-throughput/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml