When the Robot Is Cheaper Than the Hand – and When It Isn’t

Real plant data from Finland shows what AI construction waste sorting delivers: 99 percent metal recovery, 60 percent plastic, and a payback point that disappears at Hungarian wage levels

Sorting construction and demolition waste remains one of the least automated operations in the sector. A case study conducted at a Finnish recycling facility and published in Waste Management has now, for the first time, set machine learning-based robotic sorting against conventional shredder-and-manual-picking technology – with fraction-by-fraction recovery and purity figures alongside a full cost model. Over seven years the gap comes to EUR 8.7 million in the robots’ favour, with payback early in the third year. Re-run the model’s parameters at Hungarian wage levels, however, and the entire advantage evaporates. Below we set out why – and when that might change.

What was studied, and how

The paper by Zeinab Farshadfar, Siavash H. Khajavi, Tomasz Mucha and Kari Tanskanen (Aalto University, Department of Industrial Engineering and Management) appeared online in January 2025 in volume 194 of Waste Management, under a CC BY open access licence.

The research draws on two Finnish companies. One is a materials recovery facility, referred to by the authors as “Recycling Facility” for confidentiality; its construction and demolition line is fully automated, with ZenRobotics robots performing the final sort. The other is ZenRobotics itself (commercially, Terex), the Finnish robot manufacturer.

Data collection comprised eight semi-structured interviews of 60–90 minutes each and two site visits, between August 2023 and May 2024. Interviewees included the facility’s business and production managers, the robot manufacturer’s head of sales and general manager, three academic experts and a construction project manager. This was supplemented by 25 industry documents and 77 minutes of publicly available video.

One methodological detail deserves flagging at the outset, because it shapes how the whole comparison should be read: the automated side rests on real plant data, the conventional side does not. The authors state plainly that they could not obtain direct operational data from a facility using conventional sorting; the 70 tonnes-per-hour throughput comes from an equipment manufacturer’s public video (Van Dyk Recycling Solutions). Top of their list of future research directions is precisely an empirical survey of conventional facilities. This does not invalidate the finding, but it does mean real operating data is being compared partly against reference values – worth keeping in mind for any domestic reading.

AI construction waste sorting: Where the two technologies diverge

The authors mapped both processes step by step, then reduced them to nine generic functional stages. This matters because the cost comparison covers only those stages where the two methods genuinely differ, so shared pre-treatment and logistics do not distort the result. There are four such differences:

1. Shredding. The conventional line begins with a shredder to break down larger pieces – because for human sorters, size and weight are the binding constraints. The automated line simply omits this step: the robotic arms handle objects of up to 40 kilograms and 1.5 metres directly.

2. Size separation. Vibrating and trommel screens on the conventional line, assisted by human labour. The same equipment on the automated line, without human intervention.

3. Optical and NIR recognition. Absent from the conventional line. On the automated line, optical and near-infrared sensors identify materials in the smaller fractions by spectral analysis.

4. Final sorting. Manual quality control and picking on the conventional line – labour-intensive and prone to human error. On the automated line, 12 robotic arms across two stations, six for larger and six for smaller fractions.

One telling detail about the limits of conventional technology: manual picking is typically confined to two bins per operator, meaning one worker can separate two material types at a time.

Circularity: the fraction-by-fraction numbers

This is the study’s most practically useful contribution. The authors report actual performance from an interview with the facility’s production manager – not laboratory values, but operating experience.

Fraction	Recovery rate (MLAS)	Purity
Metal (ferrous)	up to 99%	up to 90%
Aluminium	–	up to 70%
Wood	up to 90%	up to 99% (C-grade wood)
Inert (mineral)	up to 80%	up to 90%
Plastic	up to 60%	up to 90%

For metals, magnetic separation alone achieves around 95 percent; the robots push this to 99. Wood losses arise mainly where material is too contaminated for the system to recognise. For inert fractions, size, weight and recognition are the limiting factors.

Plastic is the weak point, and the story behind it is instructive. The 60 percent figure is not simply a lower number – the facility redirected its robots to pick wood instead of plastic, because plastic recovery had previously been too low. In other words, a working plant allocates AI capacity where it pays, and the harder fraction temporarily drops out of the system. That kind of operator decision is entirely absent from the laboratory literature.

The study does not quantify conventional recovery rates – no data was available. It records only, on the basis of expert interviews, that conventional methods struggle to meet the EU’s 70 percent recovery target for construction and demolition waste, citing reliance on energy recovery, inconsistent sorting accuracy and contamination in output streams.

Where the robots fall short

To its credit, the study also reports the limitations, in the operators’ own terms.

Grippers sometimes fail to hold material reliably: an item is dropped prematurely or lands in the wrong container. This is less of an issue with human labour, which adapts better to varied shapes and sizes.

Occluded objects defeat the robots. A human looks around, moves things aside, digs out the buried piece – the robot is confined to what its sensors can see and cannot perform such fine manipulation.

The authors’ summary is balanced: humans are more reliable at manipulating the environment and gripping smaller objects within arm’s reach – but they tire, lose concentration, have limited reach and cannot lift heavy items. Robots, by contrast, can be operated in conditions of dust and heat unsuitable for people.

The cost model: where payback falls

The model costs the four divergent process stages: personnel (wages plus 50 percent for benefits and contributions), machinery (price plus 5 percent annual maintenance) and facility costs.

The key input parameters:

Parameter	Conventional (CS)	Automated (MLAS)
Throughput	70 t/hour	30 t/hour
Manual labour per shift	24 (20 + 4)	2
Annual wage per head	EUR 50,000	EUR 60,000
Machinery	1 shredder, EUR 50,000	2 optical/NIR units at EUR 100,000 + 12 robot units at EUR 300,000
Machinery park value	EUR 50,000	EUR 3,800,000
Machinery lifespan	10 years	10 years (robot) / 15 years (sensor)
One-off training cost	EUR 25,000/head (worker)	EUR 50,000/head (supervisor)

Across three shifts this means 72 people on the conventional line and six on the automated one. The model assumes a higher individual wage on the automated side – EUR 60,000 against EUR 50,000 – because more skilled technical staff are needed. Automation therefore does not mean cheaper labour, but fewer and more expensive people.

To handle the throughput gap, the authors divide the conventional line’s total cost by a capacity ratio of 2.3.

The result, year by year:

	Year 1	Subsequent years	7-year cumulative
Automated (MLAS)	EUR 9,408,666	EUR 558,666	EUR 12,760,666
Conventional (CS)	EUR 7,380,543	EUR 2,347,934	EUR 21,468,152

The structure is clear: the automated system’s first year is more expensive, because of the EUR 3.8 million machinery park – but from year two its running cost is roughly a quarter of the conventional line’s. On the conventional line, running costs are essentially all labour: 72 people × EUR 50,000 × the 1.5 contribution multiplier reproduces the EUR 2.35 million annual figure once divided by 2.3 almost exactly.

The payback point. The paper states that cumulative costs turn in the automated system’s favour by the end of the second year. Recalculating from the published figures, the crossover in fact occurs somewhat later, early in the third year (at roughly 2.1 years): at the end of year two the automated system’s cumulative cost is still around EUR 239,000 higher. This does not affect the practical conclusion – payback still lands around the two-year mark – but it is worth recording for accuracy.

The authors also provide a return-on-investment calculation: break-even requires EUR 9.41 million of revenue in year one, and EUR 10.35 million for a 10 percent ROI; over five years, EUR 12.81 million of revenue would deliver 10 percent on total costs of EUR 11.64 million.

The wage threshold: EUR 20,980

In September 2025 the same research group published a follow-up (Applied Sciences, 15/19, 10550) extending the same case data with the time value of money (NPV at a 4.5 percent discount rate) – an acknowledged gap in the original – and conducting a sensitivity analysis.

The results quantify how differently the two technologies respond to the same cost change, over seven years:

Variable	Conventional (CS)	Automated (CVAS)
+1 employee	+EUR 558,209	+EUR 201,670
+EUR 1,000 annual wage per head	+EUR 289,150	+EUR 55,400
+EUR 1,000 training cost per head	+EUR 31,304	+EUR 6,000
Discount rate 0% → 80%	−52.1% NPV	−25.2% NPV

Conventional sorting is an order of magnitude more sensitive to every labour-related cost; the automated system to machinery prices and maintenance. Two tipping points emerge: above an average machinery cost of EUR 512,000 per unit, conventional technology becomes cheaper – and below an annual personnel wage of EUR 20,980, likewise.

In international context (ILO Global Wage Report 2024/25): Finland EUR 48,084, the United States EUR 55,776, China EUR 15,763, India EUR 2,657. Of the 79 countries surveyed in 2023, roughly 62 percent had average annual wages below the threshold.

Re-running the model for Hungary: the advantage disappears

The study does not examine Hungary, but the conventional side of the model can be reconstructed from the published parameters, allowing an indicative Hungarian run. The calculation below is our own, at roughly HUF 390 to the euro; it follows the model’s logic (gross wage, plus the study’s own 50 percent contribution multiplier) and is not a substitute for plant-level costing.

2026 wage levels, annual gross, converted to euro:

	Monthly gross	Annual gross	≈ EUR
Minimum wage	HUF 322,800	HUF 3.87m	~EUR 9,900
Guaranteed wage minimum	HUF 373,200	HUF 4.48m	~EUR 11,500
Median earnings (Apr 2026)	HUF 616,000	HUF 7.39m	~EUR 19,000
National average (May 2026)	HUF 764,100	HUF 9.17m	~EUR 23,500

At this exchange rate the EUR 20,980 threshold corresponds to a gross monthly wage of roughly HUF 682,000 – above the national median, below the national average. Manual sorting, however, is typically staffed at around the guaranteed wage minimum, which is 55 percent of the threshold.

Running the model’s conventional side at Hungarian wages:

In the model, the conventional line’s annual running cost is essentially the wage bill for 72 people: 72 × EUR 50,000 × 1.5 = EUR 5.4 million, or EUR 2.35 million after dividing by 2.3. The same calculation at a guaranteed-wage-minimum level of EUR 11,500 gives 72 × EUR 11,500 × 1.5 = EUR 1.24 million, or roughly EUR 540,000 after normalisation.

The automated line’s annual running cost in the model is EUR 559,000.

The two figures are effectively identical. At Hungarian sorting-line wage levels, the roughly EUR 1.79 million annual operating cost gap that recovers the EUR 3.8 million robot investment within two years in Finland falls to zero. There is nothing left to recover – payback does not occur at all within the seven-year modelling horizon.

This is the key domestic message, and it is sharper than the bare wage threshold suggests: it is not that the technology pays back more slowly in Hungary, but that at current wage levels, by the logic of this model, it does not pay back.

When might that change? The guaranteed wage minimum rose 7 percent in 2026 and the minimum wage 11 percent. If sorting wages continue growing at around 9 percent a year – hardly an extreme assumption given Hungarian wage dynamics – today’s roughly EUR 11,500 reaches the EUR 20,980 threshold in about seven years, so in the early 2030s. The calculation assumes a constant exchange rate; forint weakness pushes the date out, strength brings it forward.

Two further factors move the threshold downward: falling technology prices (the EUR 512,000 machinery threshold and the wage threshold are two sides of the same equation), and investment subsidies. The authors explicitly recommend that policymakers lower the investment barrier through subsidies or tax relief, and tighten purity and recovery requirements – the latter because such standards demand performance conventional technology struggles to deliver.

Methodological limitations

The authors list several limitations themselves, and they matter particularly for a Hungarian reading.

The conventional side is not real plant data. Noted above, but worth repeating: they had no access to operational data from a conventional facility.

A single case, a single country. One recycling facility and one technology supplier.

The revenue side is missing. The model costs only. The higher market value of cleaner fractions, avoided landfill costs, the value of meeting recovery targets – all omitted. The authors flag this as future work. Against rising landfill levies and secondary raw material prices this item would tilt the balance meaningfully towards the robotic option, and would alter the Hungarian calculation above.

No financing costs. The original model omits interest on the investment for simplicity; the follow-up’s NPV framework partly compensates.

Some parameters are internally inconsistent. The study states that workers’ pay is 20 percent of supervisors’ (a 1:5 ratio), while the supporting reference describes a 2–3× difference. Nor can the annual cost figures be fully reconstructed from the published tables – the detailed derivation sits in the supplementary material. The conventional side’s numbers do reconcile; the automated side’s do not entirely.

One Hungarian specificity. Construction and demolition waste here is largely a market-based stream outside the concession system, with project-driven, volatile volumes. A 12-arm robot park carries utilisation risk that the Finnish model’s assumptions – optimised for continuous, high-volume operation – do not reflect.

The takeaway

The study’s most valuable contribution is not that the robots won. It is that it made measurable the conditions under which they win – and that its fraction-level recovery and purity figures supply benchmarks the literature has lacked. The AI waste management literature of recent years has reported almost nothing but classification accuracy, under laboratory conditions, without economic consequence. A grasping success rate above 90 percent says nothing on its own about whether the investment pays.

For Hungarian operators the practical message is twofold. In the short term: at current domestic sorting-line wages the investment does not pay back on this model’s logic, and anyone arguing for a domestic project on Finnish payback figures is working with the wrong parameters. In the medium term, however, wage dynamics, falling technology prices and tightening recovery requirements all point the same way, with the tipping point in the early 2030s – within the lifetime of a plant upgrade being planned today.

The fraction-level data is usable in the meantime: 99 percent metal recovery, 90 percent wood, 80 percent inert and 90–99 percent purity are performance levels against which Hungary’s manually sorted lines can reasonably be measured. And the 60 percent plastic figure is a reminder that the technology does not solve everything.

---

Sources:

Farshadfar, Z., Khajavi, S. H., Mucha, T., Tanskanen, K. (2025): Machine learning-based automated waste sorting in the construction industry: A comparative competitiveness case study. Waste Management, 194, 77–87. https://doi.org/10.1016/j.wasman.2025.01.008 (open access, CC BY 4.0)

Liu, X., Farshadfar, Z., Khajavi, S. H. (2025): Computer Vision-Enabled Construction Waste Sorting: A Sensitivity Analysis. Applied Sciences, 15(19), 10550. https://doi.org/10.3390/app151910550

Hungarian Central Statistical Office (KSH), Earnings, April and May 2026; Government Decree 394/2024 (XII. 12.) on statutory minimum wages

ILO: Global Wage Report 2024/25