ZF Friedrichshafen AG¶

Company Overview¶

Business Overview¶

ZF Friedrichshafen AG ("Zahnradfabrik Friedrichshafen", founded 1915) is a German Tier-1 automotive supplier headquartered in Friedrichshafen. It is not listed: the Zeppelin Foundation owns 93.8% and the Dr. Jürgen und Irmgard Ulderup Foundation 6.2% — an ownership structure that removes stock-market pressure but also limits access to equity capital, which matters for the debt story below. [Evidence] ¹⁴

Key financials and corporate state:

2024: sales €41.4 billion (down 11% nominal, ~3% organic); adjusted EBIT €1.5 billion (margin 3.6%, down from 5.1%); net debt €10.5 billion; adjusted free cash flow €305 million; R&D €3.6 billion (8.6% of sales); headcount down 4% to 161,631, including ~4,000 fewer full-time equivalents in Germany. [Evidence] ⁵
H1 2025: sales €19.7 billion; adjusted EBIT €874 million (margin 4.4%, up from 3.5%); then-CEO Holger Klein: ZF is advancing "the most comprehensive restructuring plan in its history." [Evidence] ⁶
Nine-month 2025: sales €28.9 billion (−8% nominal, −0.6% organic); adjusted EBIT margin 3.7%; net debt ~€10.6 billion, leverage down from 3.27× to 3.15×; adjusted free cash flow +€683 million (a €1.6 billion year-on-year swing). [Evidence] ⁷
Leadership churn: CEO Holger Klein left 30 September 2025; Mathias Miedreich (board member since January 2025, previously head of the Electrified Powertrain Technology division) became CEO 1 October 2025. Board member Peter Laier left the same day over "differing strategic perspectives." [Evidence] ⁸
Portfolio surgery: the Passive Safety division was spun off as ZF LIFETEC (completed by mid-2025), and in 2025 ZF agreed to sell its passenger-car ADAS business to Harman (Samsung) for €1.5 billion. [Evidence] ⁶¹⁴

Automotive Business¶

ZF supplies passenger-car, commercial-vehicle and industrial customers worldwide from 162 production locations in 29 countries, organised into divisions including Electrified Powertrain Technology, Chassis Solutions / Car Chassis Technology, Commercial Vehicle Solutions, Industrial Technology, Electronics & ADAS and Aftermarket. [Evidence] ¹⁴

Engineering Business (the part this study cares about)¶

ZF designs and manufactures its own components — so it is a user of CAD, PLM, simulation and moulding technology, not a seller. Two verified anchor points:

Moulded plastics at ZF span at least (a) chassis components — the 2025 injection-moulded IPA stabilizer link with fibre-reinforced plastic rod and all-plastic housing (Zhangjiagang, China)³ — and (b) power-electronics encapsulation — overmoulding circuit boards with reactive thermoset epoxy in the Electronics Systems unit for an e-drive program with "a major European automaker."¹ [Evidence]
Toolchain (narrow evidence): ZF uses Siemens Teamcenter Product Cost Management for parametric product- and tool-cost calculation (die-cast, cast, forged parts; carbon-footprint add-on), with quantified outcomes — bottom-up calculation time cut ~95% to 2–3 minutes, 2,400+ calculations and 40 technology cost models in two years, payback typically under one year. The full CAD/PLM authoring stack is not publicly documented (see Appendix C). [Evidence — narrow] ¹³

Role in Automotive Moulded Parts Workflow¶

Read as: where does ZF apply AI in its own moulded-parts workflow?

Workflow Stage	AI present?	Notes
Requirements	Partial (generic)	AI-assisted requirements analysis claimed by Corporate R&D — not plastics-specific¹⁰
Industrial Design	No evidence	—
Concept Design	No evidence	—
CAD Modelling	No evidence	Toolchain itself undocumented publicly
Engineering Review	No evidence	—
Simulation	Generic claim only	"Testing and validation via AI-powered simulations" (unspecified domain)¹⁰
DFM	No evidence	Nothing public on AI or even rule-automation for mouldability
Tool Design	No evidence	Tool costing is parametric/automated (Teamcenter PCM) — classical, not AI¹³
Mould Flow	No evidence	No public trace of ML surrogates or AI-assisted moulding simulation
Prototype	No evidence	—
Validation	Yes (non-plastics)	ZF Annotate — AI validation service for ADAS sensor data⁹
Manufacturing Engineering	No evidence	—
Production / Process	Yes — the one real case	sensXPERT in-mould ML on encapsulation moulding¹
Production Release	No evidence	—

The pattern is stark: in the moulded-parts chain, ZF's only verified AI sits at the press, during the shot — everything upstream of the press is, on public evidence, conventional engineering. [Synthesis]

AI Strategy¶

Public AI Vision¶

ZF's stated position: "AI is of strategic importance for ZF because it helps us to redesign and optimize our products and development processes and develop more efficiently" (Torsten Gollewski, EVP Corporate R&D). AI is framed as an "enabler, guide and pacemaker" across the portfolio, organised around an internal AI Tech Center in Friedrichshafen (head: Dr. Manuel Götz) inside Corporate R&D, with a named "Head of AI, Digital Engineering, and Validation" role (Dr. Stefan Sicklinger). [Evidence + Marketing] ⁹¹⁰¹¹

The 2019 foundation: Saarbrücken¶

On 12 March 2019 ZF announced a Technology Center for Artificial Intelligence and Cybersecurity in Saarbrücken — about 100 new hires (two-thirds AI, one-third cybersecurity), a shareholding in DFKI¹⁵ (the German Research Center for Artificial Intelligence) and a strategic partnership with CISPA¹⁶ (Helmholtz Center for Information Security). The scoped applications explicitly included "machine learning in product development and production" alongside automated driving, predictive maintenance and intelligent transmission control — so ML upstream of the factory floor has been on the official agenda since 2019. [Evidence] ⁴

Engineering AI Strategy — what it actually is¶

Assembled from the dated evidence, ZF's real AI splits four ways. [Synthesis]

Product-embedded ML — TempAI virtual sensors shipping inside e-motors (§5).¹¹
Development-process AI — requirements analysis, "AI-powered simulations," ZF Annotate validation; concrete only for ADAS/e-drive domains.¹⁰⁹
Bought-in process AI — sensXPERT on the moulding line: the one moulded-parts case, and it was purchased, not built.¹
Business-process AI — ZF-AInA, an AI negotiation assistant for procurement powered by the external vendor Pactum AI¹⁷ (again: bought, not built).¹²

Note what is absent from all four: any AI touching part geometry, mouldability, tooling or moulding simulation. [Evidence of absence]

Timeline of AI Evolution¶

Date	Event
Mar 12, 2019	Saarbrücken AI & Cybersecurity Center announced; DFKI shareholding; ML in product development & production scoped⁴
2020	sensXPERT collaboration begins (Electronics Systems / Synergy Group Electronics)¹
© 2023 (PDF Feb 2024)	sensXPERT/NETZSCH white paper: −4% avg curing time, up to −9%; decision to fit all moulds on the new line¹
Aug 26, 2024	"AI at ZF" story: AI Tech Center, requirements AI, AI-powered simulation claims¹⁰
Jul 31, 2025	sensXPERT ceases business operations²
Aug 13, 2025	IPA stabilizer link SOP announced (Zhangjiagang) — no AI content³
Sep 11, 2025	CEO change announced: Klein out, Miedreich in (effective Oct 1, 2025)⁸
May 12, 2026	TempAI virtual-sensor story: ML temperature model in series e-motors¹¹

Products Relevant to Engineering¶

Read as technologies in ZF's own engineering/production, since ZF sells components, not software.

sensXPERT Digital Mold (bought in; vendor: NETZSCH Process Intelligence GmbH). Dielectric sensors mounted inside the mould, plus temperature and pressure sensors, feeding machine-learning models that characterize the resin's behaviour during the shot and predict the moment the target degree of cure (DoC) is reached — so the press can eject as soon as the material is ready instead of waiting out a fixed safety-margin cure time. Sold "Equipment-as-a-Service." [Evidence] ¹ Vendor-level marketing claims ("up to 50% scrap reduction, 23% energy savings, 30% cycle-time reduction") are the vendor's cross-customer figures, not ZF's. [Marketing] ¹

TempAI (built in-house). A self-learning model that predicts rotor and stator temperatures inside electric motors from indirect signals (oil temperatures, cooling volume, rotor speed), trained on test-bench data (~30 calibrated measurement points). In series production in ZF's new e-motor generation; claimed benefits: predictions within a few degrees, 6–18% efficiency gain, motor development calibration work cut "from months to days," fewer rare-earth magnets needed. This is ZF's most concrete learning-from-data engineering AI. [Evidence] ¹¹

ZF Annotate. Cloud AI service for annotating and validating sensor data in ADAS/AD development — validation-stage AI, unrelated to plastics. [Evidence] ⁹

cubiX. Chassis-dynamics control software with "predictive trajectory control" framing — control engineering with AI branding; treat the AI label cautiously. [Evidence + Marketing] ⁹

ZF-AInA. "AI Negotiation Assistant" for supplier negotiations, powered by the external vendor Pactum AI; 24/7, five languages. Procurement, not engineering — but it confirms the buy-don't-build pattern. [Evidence] ¹²

Teamcenter Product Cost Management (bought in; Siemens). Parametric cost models for products and tools; ~95% faster bottom-up calculations; payback under a year. Classical parametric automation, not AI — but it reveals what ZF's engineering management pays for: fast, quantified payback. [Evidence] ¹³

AI Capabilities¶

1. In-mould cure prediction (sensXPERT) — the study's key case. Description: ZF's Electronics Systems business unit was ramping new lines to encapsulate power-electronics circuit boards (for e-drive systems) in a new grade of low-viscosity reactive epoxy — overmoulding high-value electronics against moisture, chemicals, vibration and temperature. A joint ZF/sensXPERT team ran encapsulation studies with the resin shot into an instrumented mould on a new transfer press; dielectric + temperature + pressure sensors watched the material cure in real time. At the target DoC of 90%, all parts cured within minutes — but some shots reached target much faster than others, headroom a fixed cure timer (set for worst-case material batches and process drift) can never harvest. With the ML models predicting DoC live, the press ejected at readiness. Workflow stage: production / in-process control. Inputs: dielectric, temperature, pressure signals from the mould. Outputs: live DoC prediction; end-of-cure alert. Measured outcome: curing time −4% on average; up to −9% potential on fastest-curing cycles with fully dynamic adaptation. Rico Zeiler, Specialist in Process Engineering at ZF, credited the system with deeper insight into material behaviour. Based on the results, the Synergy Group Electronics operations unit decided to outfit all moulds in all presses on its new production line with sensXPERT monitoring. Limitations: vendor-published case (no independent audit); thermoset transfer-moulding encapsulation, not thermoplastic injection moulding; the −9% is stated as potential, not achieved average; and the vendor has since shut down (§15). [Evidence] ¹

2. TempAI virtual sensors (in-house). ML predicting unmeasurable internal e-motor temperatures; in series production; also compresses development calibration from months to days — so it straddles product and development-process AI. Limitation: e-drive domain; nothing analogous published for plastics. [Evidence] ¹¹

3. AI-assisted requirements analysis. Corporate R&D describes AI systematically reviewing customer requirements. Limitation: no named product, metrics or domain. [Evidence — thin] ¹⁰

4. "AI-powered simulations" for testing/validation. Claimed in the "AI at ZF" story without domain or numbers. No evidence this touches moulding simulation. [Evidence — thin] ¹⁰

5. ZF Annotate (validation AI) and 6. ZF-AInA (procurement negotiation AI, Pactum-powered) — real, shipped, and both outside the moulded-parts chain. [Evidence] ⁹¹²

What was searched for and not found: AI or ML applied by ZF to plastic-part design, DFM checking, mould design, mould-flow simulation or moulding-defect prediction from historical data. The IPA stabilizer-link announcement — ZF's own showcase moulded part — contains no mention of AI, simulation or design method at all. [Evidence of absence] ³

Engineering Workflow Contribution¶

ZF's moulded-part workflow, on public evidence, runs conventionally from requirements through part and tool design to moulding trials — with AI appearing only at the very end:

requirements → part design (CAD; toolchain undocumented) → simulation (conventional; no public moulding-AI) → tool design & costing (parametric, not AI¹³) → moulding trials → production, where sensXPERT ML watches the material cure inside the mould¹

Contrast with the platform vendors in this study: there, AI is arriving top-down (assistants, surrogates) but hasn't reached the press; at ZF, AI entered bottom-up at the press and hasn't climbed upstream. The two ecosystems' AI fronts have not yet met anywhere in the moulded-parts chain. [Synthesis]

Public Customer Evidence¶

For a Tier-1, "customer evidence" inverts: the cases below are evidence of ZF as a customer or practitioner.

Case Study 1 — sensXPERT on power-electronics encapsulation (the AI-buyer case)¶

Practitioner: ZF Electronics Systems / "Synergy Group Electronics."
Problem: fixed cure times with safety margins waste cycle time on new encapsulation lines for an e-drive program with a major European automaker.
Solution: bought-in in-mould dielectric sensing + ML DoC prediction (collaboration from 2020).
Outcome: −4% average curing time, −9% potential; fleet-wide rollout decision (all moulds, all presses, new line). [Evidence] ¹
Caveats: published by the vendor (© 2023); no ZF-side press release found; vendor defunct since 31 July 2025.²

Case Study 2 — IPA stabilizer link (the structural-plastics push, no AI)¶

What: ZF's first Injection Plastic Assembly (IPA) stabilizer link — composite design with IPA ball joint and fibre-reinforced-plastic rod, full plastic housing replacing steel; >15% lighter, cost-optimised, aimed at new-energy-vehicle chassis.
Where/when: Zhangjiagang plant, China; SOP reached in 14 months (mass production from May 2025 per ZF's H1 statement; release dated 13 August 2025); capacity ~1.6 million units/year; further IPA products in trial runs with SOP expected 2026.
Quotes: Steffen Reichelt (Head of Product Line), Ken Si (Head of Chassis Components Asia-Pacific).
Significance for this study: ZF's moulded-parts business is growing (metal→plastic conversion at automotive scale) while its moulded-parts AI remains nil upstream — the release mentions no AI, no simulation, no design method. [Evidence] ³⁶

Case Study 3 — Teamcenter Product Cost Management (quantified, classical)¶

95% faster bottom-up cost calculations (2–3 minutes), 2,400+ calculations and 40 parametric technology cost models in two years, sub-year payback — ZF pays for engineering automation when the payback is measured and fast. Not AI. [Evidence] ¹³

White Papers¶

The sensXPERT/NETZSCH white paper "Industry Trends and Challenges: Investigating Automotive Plastics Manufacturing" (© 2023) is the primary source for Case Study 1 — a vendor document co-branded with ZF's logo and a named ZF engineer. [Evidence] ¹

Technical Architecture (Inferred)¶

Evidence¶

In-mould dielectric + temperature + pressure sensors on encapsulation moulds; edge device gathering data; ML models predicting DoC; cloud service for multi-machine historical data (sensXPERT "Digital Mold" architecture, as deployed at ZF). [Evidence] ¹
Internal AI Tech Center under Corporate R&D (Friedrichshafen); TempAI trained on test-bench calibration data. [Evidence] ¹⁰¹¹
DFKI shareholding and CISPA partnership feeding research capacity. [Evidence] ⁴
External AI vendors for non-core functions: Pactum (negotiation), and formerly NETZSCH/sensXPERT (moulding process). [Evidence] ¹²¹

Synthesis¶

ZF's AI architecture is a hub (AI Tech Center) building product-embedded and development-process ML where the data is ZF's own (test benches, vehicle data), while renting AI where a vendor's data or platform is the asset (moulding material behaviour, negotiation). The moulding case fits ZF's general pattern: the material-science ML moat belonged to the vendor, so ZF bought.

Inference¶

With sensXPERT gone, ZF either (a) inherited orphaned hardware/software it must maintain alone, (b) reverted to fixed cure timers, or (c) found a successor solution — nothing public says which. Anyone pitching moulding AI to Tier-1s should expect this question in every ZF-like procurement meeting from now on. [Inference]

AI Technologies¶

In one list: in-process ML (bought: dielectric-sensing cure prediction); virtual sensors / surrogate models (built: TempAI, trained on test data); NLP/requirements AI and "AI-powered simulation" (claimed, thin); validation AI (ZF Annotate); agent-style negotiation AI (bought: Pactum); research partnerships (DFKI shareholding, CISPA). No LLM of its own; no publicly named foundation-model partner for engineering. [Evidence + Synthesis]

Research Publications¶

No ZF papers, patents or benchmarks specifically on AI for moulded-part engineering were found in this research pass. ZF's public AI communication is corporate storytelling (zf.com "Stories") rather than peer-reviewed publication; its research capacity flows partly through DFKI, where ZF is a shareholder. A dedicated patent sweep was not performed — noted in Appendix C. [Evidence of absence, scoped] ⁴

Open Source¶

Nothing found: no AI-relevant open-source footprint (models, datasets, tools) was identified for ZF in this research pass. As a component manufacturer whose ML is either embedded in products or bought in, this is unsurprising — but it also means the community learns nothing reusable from ZF's deployments. [Evidence of absence — light search; see Appendix C]

Engineering Service / Platform Mapping¶

What is publicly verifiable about ZF's toolchain is thin: Siemens Teamcenter Product Cost Management (with carbon-footprint calculator) is documented via a Siemens case study.¹³ The CAD authoring stack (NX? Creo — a plausible TRW legacy? CATIA for OEM-facing work?) is not publicly documented and this report does not guess. For the study's platform-vendor volumes, ZF is best treated as a multi-vendor enterprise buyer whose one confirmed platform relationship (Siemens) was won on measured cost-engineering payback. [Evidence — narrow + Inference]

Engineering Intelligence Stack Mapping¶

Intent — AI-assisted requirements analysis claimed; thin, unnamed. [Evidence — thin] ¹⁰
Knowledge — no evidence of a searchable engineering-knowledge base or design-reuse system (nothing like a part-library retrieval capability is public). [Evidence of absence]
Reasoning — TempAI is genuine learned reasoning over physics (temperature prediction); nothing equivalent for moulding or part design. [Evidence] ¹¹
Execution — conventional; tool costing automated parametrically (not AI). [Evidence] ¹³
Feedback — the standout layer: sensXPERT closed a real feedback loop from measured in-mould material behaviour to press action, per shot. This is the only layer where ZF's moulded-parts practice is ahead of every platform vendor in this study — and it was bought, and its vendor is gone. [Evidence + Synthesis] ¹²

Strengths¶

Demonstrated willingness to buy moulding ML — piloted, measured, then scaled to every mould on the line. Exactly the kind of adoption behaviour that signals an opportunity. [Evidence] ¹
Real in-house ML competence (TempAI in series production; AI Tech Center; DFKI shareholding since 2019). ZF can evaluate vendors technically — it is a qualified buyer, not a credulous one. [Evidence] ¹¹⁴
Growing moulded-plastics franchise: metal→plastic conversion (IPA) at 1.6M units/year scale with more SOPs coming 2026, plus high-value electronics overmoulding. The addressable surface for moulding AI inside ZF is expanding. [Evidence] ³¹
Pays for quantified payback (Teamcenter PCM: sub-year payback documented). [Evidence] ¹³

Weaknesses¶

No AI upstream of the press in the moulded-parts chain — design, DFM, tooling and mould-flow are conventional on all public evidence. [Evidence of absence] ³⁹
Vendor-viability exposure: the one moulding-ML vendor ZF standardised a line on ceased operations in July 2025, leaving the deployment's future unknown. [Evidence] ²
Financial strain: ~€10.5 billion net debt, sub-4% margins, workforce reductions, divestments, CEO churn. Foundation ownership blocks equity raises, keeping deleveraging the priority — a hostile climate for speculative platform spending. [Evidence] ⁵⁷⁸
AI communication exceeds documented substance outside e-drive/ADAS: the requirements-AI and simulation-AI claims carry no products, domains or numbers. [Evidence — thin] ¹⁰

Current Gaps (largely manual today)¶

On public evidence: mouldability/DFM assessment of new plastic designs (e.g. the coming IPA family), mould-flow interpretation, tool-design decisions, transfer of moulding lessons between programs, and any reuse-oriented retrieval over ZF's historical part/tool designs. Also unresolved: life-after-vendor for the sensXPERT-instrumented line. [Synthesis + Inference]

Future Direction¶

More structural plastics: additional IPA products with SOP expected 2026 — the plastics engineering workload grows. [Evidence] ³
Restructuring continues under Miedreich: deleveraging (3.15× and falling), performance programs, portfolio focus after LIFETEC and the ADAS sale. [Evidence] ⁷⁸¹⁴
My read on buying behaviour (what would a Tier-1 actually buy?): cost pressure at ZF does not switch AI buying off — it shapes it. What gets bought: tools with measured, fast payback, priced as a service (sensXPERT was Equipment-as-a-Service; Pactum is a service; Teamcenter PCM paid back in under a year). What does not get bought: open-ended platforms and unproven promises. A seller to ZF needs a metric, a pilot, and an opex price — and, post-sensXPERT, proof it will still exist in three years. [Inference]

Relevance to Automotive Moulded Parts¶

Capability	State at ZF	Notes
Plastic Part Design	Active, growing — no AI	IPA stabilizer link, >15% lighter, 1.6M/yr; announcement AI-free³
Surface Design	No public evidence	—
CAD Automation	No public evidence	Toolchain undocumented
DFM	No AI, no public method	—
Tool Design	Conventional; costing parametric	Teamcenter PCM (not AI)¹³
Mould Flow	No public AI	No trace of surrogates or AI-assisted moulding sim
Manufacturing Engineering	Conventional	—
Quality / Process	AI — the proven case	sensXPERT in-mould cure prediction, fleet rollout decision¹
Engineering Knowledge Reuse	No evidence	No retrieval/reuse system public

Critical finding. The answer to this report's central question — is there any AI upstream of the press? — is no, not on public evidence. ZF's moulded-parts AI is entirely in-process: sense the material in the mould, learn its behaviour, act within the shot. Design-side AI exists at ZF only in other domains (e-motors, ADAS). The sensXPERT pattern is therefore not just ZF's first moulding AI — it is, so far, its only one. [Evidence of absence + Synthesis]

Why the in-process pattern won first (and what it teaches): the press is where labelled data exists for free — every shot generates sensor curves and an outcome, no data-cleanup project required. Upstream (design, DFM, mould-flow), the data is unlabelled geometry and siloed simulation runs. ZF's adoption sequence is the study's labeled-data bottleneck thesis playing out inside one company. [Synthesis]

Key Takeaways¶

ZF is the study's proven AI buyer: it adopted third-party in-mould ML (sensXPERT) on power-electronics encapsulation — collaboration from 2020, −4% average curing time (up to −9% potential), then fitted all moulds on the new line. [Evidence] ¹
The vendor it bought from is dead — sensXPERT (a NETZSCH venture) ceased operations 31 July 2025. Vendor viability is now part of every Tier-1 AI sale. [Evidence] ²
No AI upstream of the press: nothing public in ZF's plastic-part design, DFM, tool design or mould-flow. The moulded-parts AI is entirely in-process. [Evidence of absence] ³
ZF's moulded-plastics business is growing (IPA stabilizer link: >15% lighter, ~1.6M/yr, SOP 2025, more IPA SOPs due 2026) — the upstream AI gap will widen, not close. [Evidence] ³
ZF has genuine in-house ML where its own data lives: TempAI virtual sensors in series e-motors (May 2026 story). It builds on test-bench data, buys where the vendor owns the data moat. [Evidence] ¹¹
ML in product development has been officially scoped since March 2019 (Saarbrücken AI center, DFKI shareholding) — seven years later it still hasn't reached moulded parts. Intent ≠ deployment. [Evidence] ⁴
Financial distress (net debt ~€10.5B, restructuring, CEO change Oct 2025) shapes rather than stops AI buying: quick-payback, opex-priced, measured tools win; platforms don't. [Evidence + Inference] ⁵⁸
The one documented platform relationship — Siemens Teamcenter Product Cost Management — was justified on sub-year payback and is not AI. [Evidence] ¹³
The ZF adoption sequence (in-process first, upstream never) is live confirmation of the labelled-data bottleneck: AI landed exactly where labelled data is free. [Synthesis]

References¶

Primary Sources — ZF¶

Saarbrücken AI & Cybersecurity Center, Mar 12 2019 — https://press.zf.com/press/en/releases/release_3841.html
FY2024 annual results, Mar 20 2025 — https://press.zf.com/press/en/releases/release_82496.html
IPA stabilizer link SOP (Zhangjiagang), Aug 13 2025 — https://press.zf.com/press/en/releases/release_90944.html
Board of Management change (Klein → Miedreich), Sep 11 2025 — https://press.zf.com/press/en/releases/release_92352.html
Nine-month 2025 results, Nov 21 2025 — https://press.zf.com/press/en/releases/release_96896.html
ZF Artificial Intelligence technology page — https://www.zf.com/mobile/en/technologies/artificial_intelligence/artificial_intelligence.html
"AI at ZF" story, Aug 26 2024 — https://www.zf.com/mobile/en/technologies/artificial_intelligence/stories/ai_at_zf.html
TempAI virtual-sensor story, May 12 2026 — https://www.zf.com/mobile/en/technologies/artificial_intelligence/stories/virtual_sensor.html
ZF-AInA negotiation assistant (supplier board) — https://www.zf.com/site/supplierboard/en/ebusiness/zf_ainatheartificial_intelligence_negotiation_assistantpowered_by_pactum_ai_/zf_ainatheartificial_intelligence_negotiation_assistantpowered_by_pactum_ai_.html

Primary Sources — Vendors¶

sensXPERT/NETZSCH white paper "Industry Trends and Challenges" (© 2023; PDF Feb 2024) — https://sensxpert.com/wp-content/uploads/2024/02/sensXPERT-White-Paper-ZF-Automotive_compressed.pdf
sensXPERT homepage (closure notice, "concluded our business operations as of 31.07.25") — https://sensxpert.com/
Siemens case study: ZF & Teamcenter Product Cost Management (undated) — https://resources.sw.siemens.com/en-US/case-study-zf-teamcenter/

Secondary Sources¶

Wikipedia, ZF Friedrichshafen (ownership, divisions, ADAS sale to Harman) — https://en.wikipedia.org/wiki/ZF_Friedrichshafen
Chinabuses.org, "ZF Releases H1 Financial Report," Aug 1 2025 — https://www.chinabuses.org/news/2025/0801/article_13970.html
electrive.com on the CEO change, Sep 11 2025 — https://www.electrive.com/2025/09/11/zf-ceo-klein-must-leave-miedreich-takes-over/

Appendix A — Timeline¶

Mar 2019 — Saarbrücken AI & Cybersecurity Center; DFKI shareholding; "ML in product development and production" scoped → 2020 — sensXPERT collaboration begins → © 2023 — white paper: −4% avg cure time, all-moulds rollout decision → Aug 2024 — "AI at ZF" (AI Tech Center, requirements/simulation AI claims) → Mar 2025 — FY2024 results: €41.4B sales, €10.5B net debt, restructuring → May 2025 — IPA stabilizer link mass production → Jul 31 2025 — sensXPERT ceases operations → Aug 13 2025 — IPA press release (no AI content) → Sep/Oct 2025 — CEO change to Miedreich → Nov 2025 — 9M results, leverage 3.15× → May 2026 — TempAI story (ML in series e-motors) → 2026 — further IPA SOPs expected.

Appendix B — Glossary¶

Tier-1 supplier — a company selling systems/components directly to vehicle manufacturers (OEMs).
Degree of cure (DoC) — how far a thermoset resin's hardening chemical reaction has progressed (100% = fully cured). The part can be ejected once the target DoC is reached.
Dielectric sensor — measures the material's electrical properties inside the mould; these change as the resin cures, so the curve reveals cure progress.
Encapsulation / overmoulding — moulding polymer around an insert (here: circuit boards) to seal it against moisture, chemicals and vibration.
Transfer moulding — thermoset process where pre-measured resin is pushed from a pot into a closed mould — the process in the ZF/sensXPERT case (related to, but not the same as, thermoplastic injection moulding).
IPA (Injection Plastic Assembly) — ZF's name for its injection-moulded chassis-component design (plastic housing + fibre-reinforced plastic rod).
Virtual sensor — an ML model predicting a quantity that cannot practically be measured directly (e.g. rotor temperature), from other signals.
Equipment-as-a-Service (EaaS) — hardware + software rented as a service (opex), not bought as capex — how sensXPERT was sold.
SOP — start of production.
Net debt / leverage — borrowings minus cash; leverage = net debt ÷ EBITDA.
DFKI — German Research Center for Artificial Intelligence; ZF is a shareholder.
DFM — design for manufacturability: checking a design against the rules that make it mouldable (draft angles, wall thickness, undercuts…).

Appendix C — Notes (thinnest-evidence areas to revisit)¶

The sensXPERT case is vendor-published only. No ZF press release or independent audit of the −4%/−9% figures was found; the numbers rest on the © 2023 white paper (with a named ZF engineer, Rico Zeiler). Also note the figures describe curing time (the dominant cycle segment), and −9% is potential, not achieved average.
What happened to the deployment after sensXPERT's closure (31 Jul 2025) is unknown — orphaned, replaced, or internalised. Worth revisiting.
ZF's CAD/PLM authoring stack is not publicly documented; only Teamcenter Product Cost Management is verified (undated Siemens case study). Do not cite ZF as an "NX shop" or "CATIA shop" without new evidence.
"AI-powered simulations" and requirements AI (Aug 2024 story) have no named products, domains or metrics — treat as claims, not capabilities.
No dedicated patent sweep was performed for ZF AI/moulding patents; §10's negative is scoped to this pass. Same for open source (§11).
Wikipedia-sourced items (ownership percentages, Harman ADAS sale price, division list) should be re-verified against ZF filings before final publication.
"Synergy Group Electronics" is the white paper's name for the ZF operations unit; its exact position in ZF's current division structure (post-LIFETEC, post-reorg) was not independently verified.

Executive Summary¶

Who they are. ZF is one of the world's largest automotive Tier-1 suppliers — driveline, chassis, safety and electronics systems. 2024 sales were €41.4 billion, with about 161,600 employees and R&D spend of €3.6 billion (8.6% of sales). It is foundation-owned (Zeppelin Foundation 93.8%). [Evidence] ⁵¹⁴
Why they matter for this study: ZF is the proven AI buyer. Its Electronics Systems unit ("Synergy Group Electronics") adopted sensXPERT — a third-party machine-learning system with dielectric sensors inside the mould that predicts the resin's degree of cure in real time — on power-electronics encapsulation moulding. Measured result: curing time cut by 4% on average, with up to 9% potential on the fastest-curing shots; ZF then decided to fit all moulds in all presses on the new production line with the system. Collaboration began in 2020; the white paper is © 2023. This is the study's clearest documented case of a Tier-1 buying moulding ML instead of building it. [Evidence] ¹
The sting in the tail. sensXPERT — a venture of the German instruments group NETZSCH — ceased business operations on 31 July 2025. The proven buyer's chosen vendor is dead. That is a first-class finding in its own right: Tier-1s will buy niche moulding ML, but the vendors selling it may not survive. [Evidence] ²
The central honest finding. ZF has real, shipping engineering AI — an ML "virtual sensor" (TempAI) in series production inside its e-motors, AI-assisted requirements analysis, an AI validation service (ZF Annotate), even an AI procurement negotiator. But no public evidence was found of AI upstream of the moulding press — nothing in plastic-part design, DFM, mould design or mould-flow simulation. ZF's moulded-parts AI is entirely in-process (the sensXPERT pattern). Its flagship 2025 plastics-engineering win — the injection-moulded IPA stabilizer link — was announced with no AI or simulation content at all. [Evidence of absence] ³⁹
Context that shapes buying behaviour. ZF is mid-way through "the most comprehensive restructuring plan in its history": net debt ~€10.5 billion, a CEO change in late 2025, divestments, and German job cuts. Cost pressure has not stopped AI purchases — but it selects for quick-payback, opex-priced tools, not platform bets. [Evidence + Synthesis] ⁶⁸

sensXPERT / NETZSCH Process Intelligence white paper, "Industry Trends and Challenges: Investigating Automotive Plastics Manufacturing — including a case study on ZF Friedrichshafen AG" (© 2023; PDF generated Feb 7, 2024). Case-study facts used: collaboration from 2020; ZF Electronics Systems / Synergy Group Electronics; encapsulation of circuit boards in reactive epoxy on a new transfer press for a program with a major European automaker; dielectric + temperature + pressure sensors; target DoC 90%; curing time −4% average, up to −9% potential; quote from Rico Zeiler; decision to outfit all moulds in all presses on the new line. Vendor-level claims (50% scrap, 23% energy, 30% cycle) are cross-customer marketing. — https://sensxpert.com/wp-content/uploads/2024/02/sensXPERT-White-Paper-ZF-Automotive_compressed.pdf ↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩
sensXPERT homepage closure notice: "We have concluded our business operations as of 31.07.25" (viewed Aug 1, 2026; © 2026). The white paper identifies the operating company as NETZSCH Process Intelligence GmbH, Selb, Germany. — https://sensxpert.com/ NETZSCH corporate site: netzsch.com. ↩↩↩↩↩↩
ZF press release, "First Injection Plastic Assembly stabilizer link SOP in Zhangjiagang Plant," Aug 13, 2025 — >15% lighter than steel design, ~1.6M units/yr, SOP in 14 months, further IPA SOPs expected 2026; no AI/simulation content. — https://press.zf.com/press/en/releases/release_90944.html ↩↩↩↩↩↩↩↩↩↩↩
ZF press release, "ZF establishes Technology Center for Artificial Intelligence and Cybersecurity" (Saarbrücken), Mar 12, 2019 — ~100 hires, DFKI shareholding, CISPA partnership; scope explicitly includes "machine learning in product development and production." — https://press.zf.com/press/en/releases/release_3841.html ↩↩↩↩↩↩
ZF press release, FY2024 annual results, Mar 20, 2025 — sales €41.4B, adj. EBIT €1.5B (3.6%), net debt €10.5B (year-end 2024: €10,467M vs €9,982M in 2023), adj. FCF €305M, R&D €3.6B, headcount 161,631. — https://press.zf.com/press/en/releases/release_82496.html ↩↩↩↩
Chinabuses.org, "ZF Releases H1 Financial Report," Aug 1, 2025 — H1 sales €19.7B, adj. EBIT €874M (4.4%), Klein quote on "most comprehensive restructuring plan in its history," ZF LIFETEC spin-off completed. — https://www.chinabuses.org/news/2025/0801/article_13970.html ↩↩↩↩
ZF press release, nine-month 2025 results, Nov 21, 2025 — sales €28.9B, adj. EBIT margin 3.7%, net debt ~€10.6B, leverage 3.27×→3.15×, adj. FCF €683M. — https://press.zf.com/press/en/releases/release_96896.html ↩↩↩
ZF press release, Board of Management change, Sep 11, 2025 — Holger Klein leaves Sep 30, 2025; Mathias Miedreich (board member since Jan 2025, Electrified Powertrain Technology) becomes CEO; Peter Laier departs over "differing strategic perspectives." — https://press.zf.com/press/en/releases/release_92352.html; corroborated by electrive.com, Sep 11, 2025 — https://www.electrive.com/2025/09/11/zf-ceo-klein-must-leave-miedreich-takes-over/ ↩↩↩↩↩↩
ZF "Artificial Intelligence" technology page (viewed Aug 1, 2026) — Gollewski quote, AI Tech Center (Dr. Manuel Götz), cubiX, ZF Annotate, virtual sensors. — https://www.zf.com/mobile/en/technologies/artificial_intelligence/artificial_intelligence.html ↩↩↩↩↩↩↩↩
ZF story, "AI at ZF," Aug 26, 2024 — AI Tech Center in Friedrichshafen under Corporate R&D; applications claimed in requirements analysis, design optimization, "AI-powered simulations," virtual sensors, manufacturing/logistics. — https://www.zf.com/mobile/en/technologies/artificial_intelligence/stories/ai_at_zf.html ↩↩↩↩↩↩↩↩↩↩
ZF story on TempAI virtual sensors, May 12, 2026 — self-learning temperature prediction for rotor/stator in series-production e-motors; trained on ~30 test-bench calibration points; claimed 6–18% efficiency gain, development from months to days; names Alexander Hoffmann and Dr. Stefan Sicklinger (Head of AI, Digital Engineering, and Validation, R&D). — https://www.zf.com/mobile/en/technologies/artificial_intelligence/stories/virtual_sensor.html ↩↩↩↩↩↩↩↩↩
ZF Supplier Board page, "ZF-AInA — the Artificial Intelligence Negotiation Assistant, powered by Pactum AI" (viewed Aug 1, 2026; undated). — https://www.zf.com/site/supplierboard/en/ebusiness/zf_aina___the_artificial_intelligence_negotiation_assistant__powered_by_pactum_ai_/zf_aina___the_artificial_intelligence_negotiation_assistant__powered_by_pactum_ai_.html ↩↩↩↩
Siemens case study, "ZF" (Teamcenter Product Cost Management; undated) — 95% faster bottom-up calculations (2–3 min), 2,400+ calculations and 40 parametric technology cost models in two years, payback typically under one year; quote attributed to Simon Roth. — https://resources.sw.siemens.com/en-US/case-study-zf-teamcenter/ ↩↩↩↩↩↩↩↩↩↩
Wikipedia, "ZF Friedrichshafen" (viewed Aug 1, 2026) — Zeppelin Foundation 93.8% / Ulderup Foundation 6.2%; 162 locations in 29 countries; division list; 2025 sale of passenger-car ADAS business to Harman (Samsung) for €1.5B. Re-verify against ZF filings before publication. — https://en.wikipedia.org/wiki/ZF_Friedrichshafen ↩↩↩↩↩
DFKI (German Research Center for Artificial Intelligence) — Kaiserslautern, Germany — https://www.dfki.de ↩
CISPA (Helmholtz Center for Information Security) — Saarbrücken, Germany — https://www.cispa.de ↩
Pactum AI — Mountain View, USA — https://www.pactum.com ↩