02 · AI opportunity map: where AI pays off in Halaman's operations
Drafted 22 September 2026 from the public record (doc 01, the research/ notes) and
general industry knowledge. Revised the same day (v2) against research/ai-in-print-packaging.md
and its facts file, research/competitors.md and research/regulation-compliance.md, which add
vendor, price and case-study evidence with URLs, and against the owner-supplied page extraction
(research/owner-supplied-page-extraction.md: the full machine park, the certificate page and twelve
reference logos from halaman.com). Nothing else here relies on knowledge of Halaman's internal systems. Where the notes are silent a statement is still marked "[to verify]". Public
list prices are quoted with their source and capture date (22 September 2026); they are anchors,
not quotes. Every money figure for Halaman is an estimate and is labelled as such. The website is
out of scope; this chapter covers operations only. Regulation detail lives in doc 07 and is
cross-referenced, not repeated.
Özet (TR)
Bu bölüm, yapay zekânın Halaman'ın üretim ve ticari süreçlerinde nerede para kazandıracağını gösterir; web sitesi kapsam dışıdır. Bu sürüm (v2), tedarikçi, fiyat ve vaka araştırmasıyla (research/ai-in-print-packaging.md) güncellenmiştir; her senaryonun altında kaynaklı bir "Evidence" satırı vardır.
Makine parkı artık biliniyor (halaman.com): dört Heidelberg ofset (CX 102-6 LX, CD 102-6 LX, CX 75-6-LX-UV, CX 75-5-LX), üç dijital (HP Indigo 35K HD, HP Indigo 7000, Heidelberg Versafire CV), yedi kesim makinesi, beş Omega katlama-yapıştırma, beş GTP sıcak yaldız, Scodix, Focusight etiket kontrol makinesi. Prinect bağlantısı ve lisans durumu hâlâ doğrulanmalı (A3, A4). ERP/MIS, CRM ve prepress yazılımı kamuya açık değildir; plan hiçbir varsayımda bulunmaz. Her senaryonun veri ön koşulu yazılıdır ve 09 numaralı keşif anketiyle doğrulanır.
On fonksiyonda 18 kullanım senaryosu. Sıralama değişmedi: önce satış ve teklif (gelir), sonra uyum (PPWR 12 Ağustos 2026'dan beri yürürlükte, EmpCo 27 Eylül 2026, EUDR 30 Aralık 2026).
Hızlı kazanımlar (0-90 gün): ICP tabanlı müşteri adayı havuzu, e-posta/PDF'den RFQ okuma, beş dilde teklif ve fuar takibi, EmpCo iddia denetimi, PPWR uygunluk beyanı şablonu, Türkçe SOP asistanı. Mevcut n8n + Google Sheets + OpenAI kurulumu yeterlidir (n8n Cloud Pro 50 €/ay; RFQ başına token maliyeti sentlerle ölçülür).
Çekirdek (3-9 ay): MIS verisiyle fiyat tahmini (PrintVis 59.000 USD'den başlıyor; Cerm Lexis e-postayı siparişe çeviriyor; Tharstern fiyat açıklamıyor), otomatik preflight (Esko Automation Engine; PitStop Pro 432 USD/yıl; PACKZ 12 yapay zekâ asistanı), HP Indigo için iş gruplama (Esko Phoenix vakası: ayda 22 saat planlama tasarrufu), sipariş durumu asistanı (WhatsApp utility mesajı 0,004-0,046 USD), ihracat evrakı (MEDOS e-A.TR), e-Fatura mutabakatı (Logo/Netsis için Ilura), karton talep tahmini (Faller: tahmin doğruluğu +%50).
Yeni sonuç: HP PrintOS Print Beat, Indigo'larda OEE'yi kurulum gerektirmeden verir ve REST API ile dışa aktarır; Heidelberg Print Shop Analytics, Prinect portala bağlıysa ücretsizdir. Bu nedenle OEE panoları (UC-10) İleri'den Çekirdek'e alındı; kestirimci bakım İleri'de kaldı (anket A4 onayı şart).
İleri (9-18 ay): CX 75 / Indigo / Scodix / KAMA arasında çizelgeleme ve kestirimci bakım (Heidelberg Predictive Monitoring: plansız duruşta %20'ye kadar azalma iddiası). Zaikio 2024'te tasfiye edildi; plana dahil edilmedi.
Temel şartlar: tek ürün ana verisi, MIS/ERP veri erişimi, CRM, belge deposu, marka artwork'ü için erişim kontrolü, AB'de barındırılan model seçeneği ve KVKK Üretken Yapay Zekâ Rehberi'ne (24 Kasım 2025) uyum: yurt dışı LLM için 9. madde aktarım şartları, botların yapay zekâ olduğunu bildirmesi.
12 aylık bütçe tahmini bölüm 6'dadır (tahmindir; tedarikçi teklifleri alınmalıdır; kamuya açık liste fiyatları kaynağı ve tarihiyle verilmiştir). KOSGEB Dijital Dönüşüm, TÜBİTAK 1507 ve 1711 (2027 çağrısı); ayrıntı 07 numaralı belgede.
Başlıca riskler: marka artwork'ünün sızması, yapay zekânın uydurduğu fiyatlar, Almanya'da UWG §7, KVKK, sahada değişim direnci, tedarikçiye bağımlılık.
Önerilen ilk adım: sponsor ve günlük sorumlu atayın, anketi iki hafta içinde doldurun, ilk 90 günde altı hızlı kazanımı başlatın; aynı anda PrintOS ve Heidelberg Portal erişimini MatSet ve Heidelberg Grafik Ticaret Servis ile açın.
1. How to read this
The plan assumes nothing about Halaman's internal systems. The ERP/MIS, CRM, prepress
software and press connectivity (Prinect, HP PrintOS) are not public. Each use case
therefore lists the data it needs, and the discovery questionnaire in doc 09 confirms
feasibility. Question IDs from doc 09 (A1, C2, D3 and so on) are cited where they decide a case.
"Problem today" is inferred from the public record and from how comparable folding-carton
plants of this size usually work. The owner should correct it.
Effort is given in calendar weeks and people. "Builder" means one person who can work
with n8n, spreadsheets and an LLM API; this can be internal or a contractor.
Cost bands are external spend only (software, API usage, integrator days), excluding
internal staff time. They are rough estimates, not quotes:
Band
External spend (estimate)
A
Under EUR 5k one-off; running cost under EUR 200 per month
B
EUR 5k to 25k one-off; running cost under EUR 500 per month
C
EUR 25k to 75k one-off, usually with an integrator
D
Above EUR 75k one-off, typically a system purchase (MIS, vision)
Payback logic is written in words. No savings figures are invented; the KPI column
names what to measure so the real number appears within the pilot.
Vendors are named where they are standard in the sector or where the research notes hold
evidence. The list is not a recommendation to buy; it is a shortlist to test.
Each use case ends with an "Evidence" row: what the named tools do, results they or their
users published, the public list price or the pricing model, and the source URL. The evidence
comes from search snippets captured on 22 September 2026, not from fetched pages; confidence
follows the research notes. Where a vendor publishes only a pricing model, the model is stated
and the number is left to a quote. Nothing in an Evidence row is a Halaman figure.
2. Use-case catalogue
Eighteen use cases in ten functions. Numbering (UC-01 to UC-18) is reused in the
prioritisation table (section 3) and the budget (section 6). A peer-evidence subsection
closes the catalogue.
(a) Sales and customer acquisition
UC-01 · ICP-driven prospecting pipeline and account-signal monitoring
Field
Detail
Problem today (inferred)
New business comes from fairs (Cosmopack, FachPack), brand nominations and referrals. No CRM is visible. Prospect research is manual and stops when the fair season ends.
What the AI does
Scores prospect lists weekly against the ideal customer profile (doc 04), enriches each company from its homepage, and writes a ranked shortlist with reasons and a suggested channel. A second workflow watches public signals (plastic-to-paper pledges, new sourcing offices in Istanbul, packaging-developer job posts, PPWR announcements) and flags accounts to call.
Data it needs
A prospect sheet (company, website, country, segment). The ICP rubric. The twelve reference logos on halaman.com (Lacoste, Benetton, Etam, Salomon, Under Armour, H&M, Hugo Boss, Le Bourget, Tesco, Monoprix, Boots, Burger King; research/owner-supplied-page-extraction.md section 1) as the first lookalike seed. Later: the real top-20 customer list (B1) to recalibrate the rubric.
Tools / vendors
The existing n8n workflow in tools/n8n/ (Google Sheets + OpenAI, validated, 31 nodes). Claude or OpenAI models. HubSpot or Pipedrive once a CRM is chosen. RSS/news nodes in n8n for signals. Apollo, Clay or LinkedIn Sales Navigator for DE/AT contacts [licence terms and consent handling to verify].
Effort
2 to 4 weeks. One sales owner plus one builder part-time. The scoring workflow already exists; the signal watcher is new.
Cost band
A (estimate). The n8n README measures about 0.27 M input and 0.1 M output tokens per 100 prospects; at the GPT-5 list price that is about USD 1.40 per 100 prospects, and about USD 0.30 on a mini-class model.
Payback logic
A single won header-card or folding-box programme runs to hundreds of thousands or millions of units per year (see M&S and Decathlon volumes in research/market-textile.md). One such account covers years of running cost.
KPI
Shortlisted accounts per week; meetings booked per 100 prospects; win rate by segment.
Risk
Scores are LLM estimates from thin public data; use for triage only. German and Austrian outreach must respect UWG §7 (section 7). Dirty or duplicate rows.
UC-02 · RFQ intake and parsing from e-mail and PDF
Field
Detail
Problem today (inferred)
RFQs arrive as e-mails with PDF or Excel attachments, sometimes WhatsApp or phone (A6). Someone re-keys the spec. Missing fields (board, finishing, quantity breaks) cause a second and third e-mail before a quote can start.
What the AI does
Reads the shared sales mailbox, separates RFQs from other mail, extracts a structured request (product family, dimensions, board, colours, finishing, quantities, delivery place and date, language), stores it, and drafts the clarification questions in the customer's language for a human to send.
Data it needs
Access to the sales mailbox (read). A spec schema, ideally the product master (section 4). 50 to 100 past RFQs with their final specs as test set.
Tools / vendors
n8n (IMAP, Gmail or Outlook nodes) with Claude or OpenAI structured outputs. OCR for scanned PDFs, for example Azure Document Intelligence or Google Document AI [to verify]. WhatsApp Business Platform if WhatsApp RFQs are frequent. Cerm Lexis if a Cerm MIS is chosen (UC-04).
Effort
3 to 5 weeks. One sales or customer-service owner plus one builder.
Cost band
A to B (estimate). Token cost is negligible; the cost is integration and the test set.
Payback logic
Minutes saved per RFQ times monthly RFQ count (A6), plus a faster first response. In packaging, the first complete quote often wins; speed is the main lever.
KPI
Time from RFQ arrival to a complete structured request; share of RFQs parsed without manual correction.
Risk
Extraction errors on unusual specs. A human must check before any quote is sent.
Evidence
Cerm Lexis (Belgium) reads customer e-mails, PDFs and Excel files, converts them into structured CERM7 sales orders "ready for confirmation" for label, flexible-packaging and folding-carton converters; in use at 20 customers, global release 15 September 2026; Onpack (AU) cut large-order entry from 2 to 3 hours to under 1 hour (https://www.cerm.net/lexis ; https://www.labelsandlabeling.com/news/workflow-mis-inspection/cerm-launches-lexis-ai-agent). Cerm prices by turnover or headcount, no public number (https://www.cerm.net/pricing); Lexis only works with the Cerm MIS, so for Halaman it is an MIS decision, not a bolt-on. Build option: the same pattern with n8n, Gmail, Airtable and an LLM is documented for a manufacturer's RFQ flow (https://arnasoftech.com/case-study/ai-powered-rfq-automation/). Extraction benchmarks on invoices and purchase orders: vision-based reading 92.71 percent on scanned documents against 64.03 percent for parsed text; GPT-5 class models above 96 percent on clean documents; best-in-class straight-through rate about 80 percent, so the human confirmation step stays (https://parseur.com/blog/ai-invoice-processing-benchmarks). Token cost: about 3,000 tokens per RFQ e-mail plus PDF; 1,000 RFQs per month is roughly 3 M input tokens, about USD 4 at the GPT-5 list price (derived in the research notes). WhatsApp Business Platform: utility messages USD 0.004 to 0.0456; customer-initiated service replies free within 24 hours until 1 October 2026, then billed per message; 1,000 free service conversations per month per account (https://developers.facebook.com/documentation/business-messaging/whatsapp/pricing ; https://blueticks.co/blog/whatsapp-business-api-pricing-2026). OCR engines: the notes are silent [to verify].
UC-03 · Multilingual outreach and fair follow-up (TR/DE/EN/FR/IT)
Field
Detail
Problem today (inferred)
The factory works in Turkish, the Hamburg office in German, customers in French, Italian and English. Fair leads are followed up unevenly because writing five languages well is slow.
What the AI does
Drafts personalised follow-ups and short proposals from the fair lead notes, in the recipient's language, using an approved messaging and claims library. Sequences are prepared in the CRM; a human presses send. Also translates spec sheets and quotes with a terminology list.
Data it needs
A fair lead sheet (name, company, interest, consent status, language). Approved product sheets and claims (UC-16).
Tools / vendors
n8n plus Claude or OpenAI. HubSpot or Pipedrive sequences. DeepL for translation cross-check. A mandatory consent field for DE/AT contacts.
Effort
2 to 3 weeks. One sales owner plus one builder.
Cost band
A (estimate).
Payback logic
Fairs are a fixed cost (stand, travel, staff; actual figure from the company). Converting a few more fair leads to meetings pays the workflow many times over.
KPI
Fair leads contacted within 5 working days; reply rate by language; meetings per fair.
Risk
Wrong tone or a wrong technical term in a language nobody in-house reads. Unsolicited e-mail to German companies without consent (UWG §7). Claims must pass the EmpCo check.
UC-04 · AI-assisted estimating and quote drafting with MIS data, with repeat-job detection
Field
Detail
Problem today (inferred)
Quoting depends on a few experienced people and possibly Excel (A2). Repeat jobs are re-estimated from scratch. Route choice is wide: B1 offset on the CX 102 or CD 102, B2 on the two CX 75s (UV or conventional), digital on the Indigo 35K HD, Indigo 7000 or Versafire CV, Scodix against hot foil on five GTP platens or the Varimatrix CSF inline unit, and seven die-cutters. It is done by habit.
What the AI does
Takes the structured RFQ (UC-02), finds the nearest past jobs (same customer, same die, similar spec), proposes a production route with a cost comparison, pulls the cost rates from the MIS, and drafts the quote with its assumptions listed. Repeat jobs are flagged and their specs copied. The estimator approves; the MIS does the arithmetic.
Data it needs
MIS or ERP job history: specs, quantities, run times, material costs, sold price. Cost rates per machine. The product master. If there is no MIS: 12 to 24 months of Excel quotes.
Tools / vendors
MIS estimating modules: PrintVis (Dynamics 365 Business Central, Copilot built in), Cerm (with Lexis intake), ePS Radius or Pace, Tharstern, Optimus dash, Heidelberg Prinect Business; Turkish print MIS (Matbaasoft, EraOFSET, Pusula, TAM) [fit and Turkish localisation to verify]. GelatoConnect Estimator as an AI quoting layer on top of an MIS. Claude or OpenAI for retrieval and drafting. n8n. A vector index over past jobs.
Effort
8 to 12 weeks on an existing MIS. Estimator, MIS administrator, one builder. If an MIS must first be introduced, add 4 to 9 months.
Cost band
B on an existing MIS; C to D if an MIS is bought (estimate; quotes required; PrintVis is the only packaging MIS with a public anchor, from USD 59,000).
Payback logic
Quote turnaround from days to hours, a higher hit rate, and fewer under-priced jobs. Value scales with monthly job count and repeat share (C1).
KPI
Quote turnaround time; quote-to-order rate; margin variance actual vs estimated.
Risk
Hallucinated prices if the model is allowed to invent rates. Rule: the model retrieves and drafts, the MIS calculates.
Dielines and artwork arrive as PDF or Illustrator files from brand agencies. Blank size, style, window, Euro hole and finishing layers are re-typed, and mismatches surface late in prepress. Windows and trays are fitted on Heiber-Schroeder and Esatec machines, so window geometry matters early.
What the AI does
Reads dieline PDFs and artwork layers, extracts blank size, panel count, glue flap, window and hole positions, spot colours and foil, emboss or spot-UV layers, compares them with the RFQ and outputs a spec sheet in the product-master format.
Data it needs
100 dielines and artworks with known specs. Layer-naming conventions. Which CAD is in use (A3).
Tools / vendors
Esko ArtiosCAD or Automation Engine for structural data [what is installed to verify]. Enfocus PitStop for PDF inspection. Vision-capable Claude or OpenAI models for layer and annotation reading. Python PDF libraries called from n8n.
Effort
6 to 8 weeks. Prepress lead plus one builder.
Cost band
B (estimate).
Payback logic
Fewer spec errors reaching the press; faster estimating because UC-04 receives clean input.
KPI
Spec fields auto-filled correctly; spec corrections after order confirmation.
Risk
Non-standard files. Keep human sign-off; treat the output as a draft.
UC-06 · Automated preflight, artwork QA and dieline check
Field
Detail
Problem today (inferred)
Preflight is manual or semi-manual. Artwork from many agencies carries bleed, font, overprint, barcode and text errors. Dieline-to-artwork mismatch is found on the proof, not on receipt. Plates come off a Heidelberg Suprasetter A106 and a CRON thermal CTP, so a Prinect prepress workflow is likely but unconfirmed (A3).
What the AI does
Rule-based preflight (resolution, bleed, fonts, overprint, ink coverage, barcodes) plus AI comparison of artwork versions (text differences, language check, Braille for pharma) and dieline alignment. Writes the report to the customer in their language.
Data it needs
Preflight profiles per product family. The dieline library. Approval history.
Tools / vendors
Enfocus PitStop Server and Switch. Esko Automation Engine. Hybrid Software PACKZ 12 and PACKZFLOW. Heidelberg Prinect Production Manager (PPM Packaging) if Prinect is already the workflow. EyeC Proofiler Graphic or GlobalVision Verify for artwork comparison. GS1 barcode verification. LLM for report drafting.
Effort
6 to 10 weeks. Prepress lead, one builder, vendor support.
Cost band
B to C (estimate). PitStop Pro is USD 432 per seat per year; the workflow engines are quote-only subscriptions.
Payback logic
Fewer reprints and delays; fewer ISO 9001 corrective actions; faster approvals.
KPI
Errors caught before proof; remakes caused by artwork; hours from artwork receipt to proof.
Risk
Over-strict rules block jobs. Vendor lock-in. Brand NDAs on where artwork may be stored (D3).
UC-07 · Colour management to ISO 12647-2 and proof-approval assistant
Field
Detail
Problem today (inferred)
Halaman is ISO 12647-2 and Mellow Colour certified. Brand colours must match across four gamuts: UV offset (CX 75-6-LX-UV), conventional offset (CX 102, CD 102, CX 75-5), HP Indigo liquid ink (35K HD, 7000) and Versafire CV toner. Proof approvals wait on customers.
What the AI does
Monitors press-side measurements against ISO 12647-2 tolerances, predicts drift, recommends recalibration, and matches spot colours across offset and digital. An assistant packages proofs, explains ΔE results to customers in plain language and tracks approvals.
Data it needs
Spectrophotometer readings and control-strip data, including Prinect Inpress Control logs from the Speedmasters. The colour library (brand references). A proof approval log.
Tools / vendors
Heidelberg Prinect Color Toolbox, Inpress Control and Color Assistant Pro [licence status to verify, A3]. HP PrintOS Color Beat via the Print Beat API. GMG or X-Rite tools. LLM for the assistant. n8n for approval tracking.
Effort
6 to 8 weeks. Colour lead plus one builder. Depends on A3 and A4.
Cost band
B (estimate).
Payback logic
Fewer colour rejections and reprints; shorter approval cycles; protects a certification that is part of the sales pitch.
KPI
Share of jobs within tolerance on the first pull; proof approval cycle time.
Risk
Press connectivity unknown. Customers may distrust automated colour decisions; keep the operator in charge.
UC-08 · Job ganging for HP Indigo and digital-vs-offset routing rules
Field
Detail
Problem today (inferred)
The 35K HD set a 1.7 million-click single-shift record in January 2026, so volumes are high and runs are short. Three digital presses (Indigo 35K HD in B2, Indigo 7000, Versafire CV toner) sit next to four offset presses. Ganging of header cards and small boxes is manual. Routing follows rules of thumb.
What the AI does
Proposes gang layouts by board, finishing and due date. A routing rule engine (quantity, board, colours, embellishment, deadline) compares cost per route across the seven presses using UC-04 rates and learns from actual results.
Data it needs
The open order list with specs and due dates. Board stock. Press cost models. Run-time history.
Tools / vendors
Esko Phoenix (AI planning and imposition). HP PrintOS Site Flow and HP Nio for the Indigos. Kodak sPrint One inside Prinergy. Scodix Studio W2P for embellishment jobs. Or a custom optimiser in Python called from n8n.
Effort
8 to 12 weeks. Planner, one builder, vendor.
Cost band
B to C (estimate). Phoenix and Site Flow are quote-only.
Payback logic
Board waste and click cost per job fall; more jobs per shift on the same press.
KPI
Sheet utilisation percent; jobs per shift; share of jobs routed to the cheaper feasible route.
Risk
Needs structured order data. Planners must trust and be able to override the suggestion.
UC-09 · Scheduling with changeover minimisation across presses, die-cutters, folder-gluers and finishing
Field
Detail
Problem today (inferred)
The machine park (halaman.com) is wide: four Speedmasters, three digital presses, seven die-cutters (Promatrix 106 CSB, Varimatrix CSF 105 with hot foil, two Varimatrix CS 105, Easymatrix 106 CS, Bobst Speria 106 E, KAMA ProCut 76), five Omega Allpro folder-gluers (110, 90, two 70, 55), five GTP hot-foil platens, Scodix, Top Spot UV, a Lotus laminator, window patching and box setting. Changeovers (plates, dies, foil, glue, board grade) dominate short-run economics, and the same job can take several routes through die-cutting and gluing. The schedule probably lives in Excel or on a board (C1, C2).
What the AI does
A constraint-based scheduler that sequences jobs across presses and postpress to minimise changeovers while respecting due dates, die and tooling availability, gluer capacity and shift patterns. Re-plans on disruption and answers what-if questions from the planner in plain language.
Data it needs
Live job list with routings, setup matrices per machine, machine status, shifts. Ideally press and postpress feedback (JDF/JMF, PrintOS, Prinect).
Tools / vendors
MIS scheduling modules (PrintVis, Cerm, ePS PrintFlow 4D, Tharstern) [to verify]. Heidelberg Prinect Scheduler (included in Prinect Production Manager) and Push to Stop on the Speedmasters. Custom optimiser with OR-Tools. LLM as planner interface.
Effort
12 to 20 weeks. Planner, production manager, one builder or integrator. Shorter if Prinect Production Manager is already licensed (A3).
Cost band
C (estimate).
Payback logic
Setup minutes recovered become capacity without new machines; on-time delivery improves.
KPI
Changeover minutes per shift; on-time-in-full; schedule adherence.
Risk
Garbage in, garbage out. Little value without machine feedback. Shop-floor adoption.
UC-10 · OEE dashboards from press data and predictive maintenance
Field
Detail
Problem today (inferred)
Press data exists in HP PrintOS for both Indigos and, if licensed, in Prinect for the four Speedmasters and the Suprasetter CTP, but may not be used. Downtime causes are anecdotal. Maintenance is reactive.
What the AI does
OEE dashboards per machine, anomaly detection on downtime and waste, alerts on consumable wear and recurring faults, and a maintenance planner that drafts work orders.
Data it needs
PrintOS Print Beat API feed. Prinect data via the Heidelberg Portal. KAMA, Scodix and Versafire logs. A maintenance log (paper or CMMS).
Tools / vendors
HP PrintOS Print Beat, Automatic Alert Agent and HP Nio. Heidelberg Print Shop Analytics with Performance Advisor Technology, AI Performance Chat, Predictive Monitoring and the Heidelberg Assistant. Grafana or Power BI. LLM summaries. HP Maintenance Beat [to verify].
Effort
4 to 8 weeks for dashboards, because both OEMs already compute OEE; 12 weeks or more for predictive alerts. Production manager, one data person, OEM support.
Cost band
A to B for dashboards (free OEM tiers plus a BI seat); C for predictive (estimate; service contracts [to verify]).
Payback logic
Uptime on the 35K HD and the Speedmasters, the most expensive assets in the plant.
UC-11 · Inspection data mining, complaint analytics and first-article checks
Field
Detail
Problem today (inferred)
A Focusight label inspection machine is installed (halaman.com machine park), so reject data exists but is probably not stored or analysed. None of the four Speedmasters is documented with inline inspection. Complaints are logged in e-mail or spreadsheets (A7). First-article checks are manual, with customer representatives present.
What the AI does
Mines Focusight rejects by job, press, board and finishing. Clusters complaint texts (TR/EN/DE) into root causes. A digital first-article checklist compares photos with the approved proof. Drafts CAPA records for ISO 9001 and BRCGS. Later: an AI classifier that separates critical from cosmetic defects, trained on Halaman's own reject images.
Data it needs
Focusight logs and images (export format and licence [to verify]). Complaint records. Job data. Approved proofs.
Tools / vendors
Export from the Focusight machine [to identify]. Python plus LLM. n8n. Vision LLM for first-article photos. For a later inline or offline AI station: EyeC ProofRunner, Lake Image, Cognex, or a Turkish vision vendor (Leosay, Mongery Soft, infabit, Codexper).
Effort
4 to 8 weeks. Quality manager plus one builder.
Cost band
A to B for analytics (estimate). A new AI vision station is Band D and is not in the 12-month budget.
UC-12 · Board demand forecasting, price tracking and EUDR-ready supplier data
Field
Detail
Problem today (inferred)
Board is the largest material cost. Purchase quantities are decided by experience (C3). Prices are volatile. EUDR requires mill-level data for board placed on the EU market as product from 30 December 2026.
What the AI does
Forecasts board demand by grade from the order pipeline and seasonality. Tracks supplier quotes and index prices. Maintains a supplier register with DDS reference numbers, FSC claims, recycled share and geolocation where needed, and flags expiring documents.
Data it needs
Purchase history, stock levels, order pipeline, supplier certificates and mill declarations.
Tools / vendors
ERP purchasing data (Logo, Netsis, SAP B1 or other; A1). n8n. A database or sheet. LLM for document extraction. FSC "Aligned for EUDR" module (live). EUWID or Fastmarkets price feeds. A light EUDR tool only if the register outgrows a sheet.
Effort
6 to 10 weeks. Purchasing lead plus one builder. The EUDR register alone is 2 to 3 weeks and should start now.
Cost band
B (estimate). Forecasting SaaS is not needed at Halaman's scale; a statistical model on 12 to 24 months of history is enough to start.
Payback logic
Less dead stock and fewer rush purchases. EUDR readiness protects EU orders: buyers are already refusing orders without traceability data (research/regulation-compliance.md 2.5).
KPI
Forecast error; stock days by grade; share of board lots with complete EUDR data.
Risk
Mills are slow to supply data. Forecasts need a year of history to be trusted.
UC-13 · Order-status assistant for the Hamburg office and customers
Field
Detail
Problem today (inferred)
"Where is my order" questions from Hamburg and from customers are answered by sales people by phone and e-mail, across time zones and languages.
What the AI does
Answers order-status questions from MIS or ERP status, proof-approval state and shipment tracking, in DE/EN/FR/IT/TR. Sends proactive delay notices. Internal use first (Hamburg office), then a customer channel.
Data it needs
Read-only order status from MIS or ERP (A2). Shipping data. Approval state.
Tools / vendors
n8n plus LLM. WhatsApp Business Platform for utility notices. Intercom Fin or HubSpot Breeze if a managed agent is preferred. Microsoft Teams or Slack for the internal channel [to verify]. Later a customer portal.
Effort
4 to 6 weeks internal; 8 weeks or more for the customer channel. Customer-service owner plus one builder.
Cost band
A to B (estimate). Metered channels: cents per message or under one dollar per resolved conversation.
Payback logic
Sales time returned to selling; better service for EU accounts.
KPI
Status questions answered without a human; response time.
Risk
A wrong status costs trust. Read-only integration; the assistant never promises dates that are not in the system. Bots must disclose that they are AI (KVKK).
Each export shipment needs a commercial invoice, packing list, an electronic A.TR for the EU customs union, an origin statement for the UK, and an export-type e-Fatura with the 12-digit GTİP code that the customs declaration references. Export staff prepare these with a broker (C5).
What the AI does
Generates the document set from order and shipment data, checks consistency (weights, HS codes such as 4819 and 4821, quantities, incoterms), prepares the broker package, pre-fills the MEDOS A.TR request and tracks issuance. Attaches the EUDR pass-through data and PPWR DoC reference (UC-17) to the packing list.
Data it needs
Order, shipment, HS code master, customer master, incoterms. The export e-Fatura XML (UBL-TR).
Tools / vendors
ERP export module and e-Fatura or e-İrsaliye integrations (Logo, Netsis, SAP B1 partner add-on; A1). TOBB MEDOS for e-A.TR. Broker portal. n8n. LLM for consistency checks.
Effort
4 to 8 weeks. Export staff, one builder, the broker.
Cost band
A to B (estimate).
Payback logic
Hours saved per shipment; fewer customs holds and corrections.
KPI
Minutes per shipment document set; customs queries per 100 shipments.
Risk
These are legal documents. Errors are expensive; human approval stays. Rules change (UK FTA renegotiation to 2027, EUDR references on documents).
UC-15 · Invoice OCR, e-Fatura reconciliation and DSO chasing
Field
Detail
Problem today (inferred)
Supplier invoices (board, ink, foil, logistics) are matched to purchase orders by hand. Inbound e-Fatura is reconciled manually. EU receivables are chased by e-mail when someone remembers.
What the AI does
Three-way matching of supplier invoices to PO and goods receipt, OCR only for foreign and paper invoices, anomaly flags, e-Fatura XML reconciliation with the ERP, and polite multilingual reminders drafted by ageing bucket for a human to send.
Data it needs
ERP payables and receivables. The e-Fatura inbox (UBL-TR XML). PO data. Bank statements.
Tools / vendors
ERP e-Fatura integrator (Logo, Netsis, SAP B1 partner; A1). Ilura agent on Logo/Netsis. Turkish invoice OCR (Papirus.ai, Softay, Data Market/OnBase) or a general Document AI. n8n. LLM for reminders.
UC-16 · EmpCo-compliant claims checker and customer questionnaire assistant
Field
Detail
Problem today (inferred)
The "climate neutral printing" USP (ClimatePartner since 2012) conflicts with the EmpCo directive from 27 September 2026 and already with Türkiye's Reklam Kurulu guide. About 70 percent of customers ask about certifications, and PPWR, LkSG, EcoVadis and Sedex questionnaires are answered by hand, repeatedly.
What the AI does
Scans web, sales-deck, label and blog copy in EN/TR/DE for banned or generic claims and proposes specific, substantiated wording. Keeps an evidence library (certificates, policies, energy data) and drafts questionnaire answers with references to the evidence. The rule set and the generator design are specified in doc 07, section 6.2; this chapter builds them.
Data it needs
All marketing copy. Certificates confirmed on halaman.com: ISO 9001, ISO 12647-2, ISO 14001, FSC, BRC/IoP and BRCGS, SMETA, Amfori BSCI, Sıfır Atık, Mellow Colour, ClimatePartner (research/owner-supplied-page-extraction.md section 4). Policies. Solar and energy data.
Tools / vendors
Claude or OpenAI with a document store. n8n. A rule list derived from EmpCo's banned-claim examples, the UWG amendment, the French décret 2022-539, the CMA Green Claims Code and the Reklam Kurulu guide. ClimatePartner's replacement labels. Legal review of the final wording.
Effort
2 to 4 weeks. Marketing, quality, one builder.
Cost band
A (estimate).
Payback logic
Removes a legal exposure that customers will push upstream. Faster questionnaire turnaround keeps nominated-supplier status.
KPI
Banned claims remaining across all copy; questionnaire turnaround days.
Risk
This is not legal advice. Counsel signs off the claim list.
UC-17 · PPWR Declaration of Conformity generator and EUDR declaration pack
Field
Detail
Problem today (inferred)
PPWR applies since 12 August 2026. Importers must obtain the Declaration of Conformity (Annex VIII) and technical documentation (Annex VII) from the manufacturer. EUDR due-diligence statements apply from 30 December 2026 for cartons placed on the EU market as products. No product-specification database is visible.
What the AI does
From a per-SKU product master (board grade, grammage, coating, ink system, foil or lamination share, adhesive, weight, dimensions, heavy-metal and PFAS test references, FSC claim) it generates the DoC and technical file, pre-screens recyclability risk (lamination, foil, Scodix polymer) at quotation, and assembles the EUDR supplier declaration with mill DDS reference numbers per shipment. Requirements, data model and generator design: doc 07, sections 6.1 and 6.5.
Data it needs
The product master (section 4, must be built). Laboratory test reports. The supplier register (UC-12).
Tools / vendors
Template generation with an LLM and n8n. Document store. Off-the-shelf DoC tools (PPWR Copilot, Complydex, Sunhat) as a benchmark or fallback. The portals customers use (IntegrityNext PPWR module, Sunhat, Ideagen). Annex VIII DoC template. RecyClass free self-assessment for the recyclability pre-screen.
Effort
Minimum version 3 to 4 weeks (template plus spreadsheet master). Full version 8 to 12 weeks. Quality or compliance owner plus one builder.
Cost band
A for the minimum version; B for the full version (estimate). Tool anchors: GBP 49 to 499 per month or EUR 9.90 per packaging type.
Payback logic
Keeps EU orders flowing and avoids blocked shipments. Deadlines are fixed by law, so this is risk avoidance rather than savings.
KPI
Share of active SKUs with a DoC; share of EU shipments with an EUDR pack; customer requests answered within the agreed SLA.
Risk
Delegated acts are still moving (recyclability grades, labelling). A wrong declaration creates liability; legal review of templates.
UC-18 · SOP assistant in Turkish, onboarding and machine-manual Q&A
Field
Detail
Problem today (inferred)
About 200 staff. ISO 9001, BRCGS and SMETA procedures exist as documents. New operators on the Indigos, Versafire, Scodix, the seven die-cutters and five Omega gluers learn from colleagues; OEM manuals (HP, Heidelberg, Scodix, KAMA, Bobst, Omega, Steinemann, Heiber-Schroeder) are in English or German.
What the AI does
A Turkish-language assistant over SOPs, work instructions and machine manuals, used from a phone or tablet on the shop floor. Builds an onboarding path per role with short checks. Points operators to the OEM assistants (HP Nio, Heidelberg AI Performance Chat) where they exist.
Data it needs
SOP documents. Manuals (check the OEM licence before indexing). Training records.
Tools / vendors
Claude or OpenAI with retrieval over the document store; Kumru (VNGRS) on-premise where Turkish-only, in-house hosting is required. n8n or a small web app. WhatsApp interface (UC-13 prices) or Teams [to verify]. Role-based access.
Effort
3 to 5 weeks. Quality or HR owner plus one builder.
Cost band
A (estimate).
Payback logic
Shorter time to autonomy for new operators; fewer procedure errors; audit readiness (BRCGS Issue 7, SMETA 7.0).
KPI
Questions answered per week; onboarding time to autonomy; audit findings on procedures.
Risk
Outdated SOPs are amplified. Every document needs an owner and a review date.
Measured results where the notes give numbers; otherwise the statement is what the company
or vendor published. Confidence is that of the research notes (search snippets).
DS Smith and Pusterla 1880: DS Smith built an AI and integration platform with Hitachi Digital Services (May 2025), no KPI (source in the notes: hitachids.com, page not captured); Pusterla 1880 standardised global processes on Oracle Cloud ERP (October 2025), no KPI. https://www.businesswire.com/news/home/20251003867603/en/
What this means for Halaman: the measured wins sit in forecasting (Faller, MCC), planning
(the Phoenix case), order intake (Cerm, GelatoConnect) and press automation (Push to Stop).
The large groups talk about AI in design and process, not in numbers. A Turkish converter
that publishes one measured AI result would stand alone in its market.
3. Prioritisation
Scales: impact 1 (low) to 5 (high). Effort 1 (under 4 weeks, one builder) to 5 (over 3
months with an integrator). Data readiness 1 (not captured) to 5 (structured and available
today, as far as the public record suggests; doc 09 will correct these).
UC
Use case
Impact
Effort
Data readiness
Tier
Note
01
Prospecting and signals
4
1
4
Quick win
Workflow exists; twelve reference logos as seed; needs a sales owner
02
RFQ intake and parsing
4
2
3
Quick win
Mailbox access is enough to start
03
Multilingual outreach
3
1
3
Quick win
Consent field first
16
EmpCo claims and questionnaires
4
1
4
Quick win
Deadline 27 Sep 2026; SMETA and BSCI certificates exist
17
PPWR DoC and EUDR pack
5
2
2
Quick win (minimum), then Core
PPWR live; EUDR 30 Dec 2026
18
SOP assistant
3
1
3
Quick win
ISO 9001 and BRCGS documents exist
12
Board forecasting and supplier data
4
3
2
Core (EUDR register now)
Needs ERP purchase data
04
Estimating with MIS data
5
4
2
Core
Depends on A2
05
Spec extraction
3
3
3
Core
Feeds UC-04 and UC-06
06
Preflight and artwork QA
4
3
2
Core
Depends on A3
07
Colour management assistant
3
3
2
Core
Depends on A3, A4
08
Ganging and routing
4
3
2
Core
Needs structured orders
10
OEE dashboards
4
3
2
Core (dashboards)
Print Beat OEE is out of the box; Heidelberg Portal free tier; confirm A4
11
Quality analytics
3
2
2
Core
Start with Focusight exports and complaint texts
13
Order-status assistant
3
2
2
Core
Read-only MIS access
14
Export documents
3
2
3
Core
e-Fatura and MEDOS data exist
15
Invoice OCR and DSO
3
2
3
Core
ERP integrator needed
09
Scheduling optimiser
4
5
1
Advanced
Needs machine feedback across presses and postpress
10
Predictive maintenance
4
4
1
Advanced
OEM service contracts; needs months of history
Tiers: Quick wins 0 to 90 days; Core 3 to 9 months; Advanced 9 to 18 months.
Change from v1. UC-10 is split. The dashboard half moves from Advanced to Core with effort
4 to 3 and data readiness 1 to 2, because HP PrintOS Print Beat computes OEE on the Indigos
without setup and streams it by API, and Heidelberg Print Shop Analytics is free once Prinect
is linked to the portal (UC-10 evidence). The predictive half keeps its v1 scores. No other
score changed: the new evidence confirms tools, prices and peer results, not Halaman's data.
Why sales, quoting and compliance come first. Sales and quoting touch revenue directly
and need only data Halaman already controls: its mailbox, its past quotes and public prospect
data. Compliance has dates set by law, not by Halaman: PPWR since 12 August 2026, EmpCo from
27 September 2026, EUDR from 30 December 2026. Missing them risks blocked shipments and
customer liability, which no efficiency gain elsewhere can offset. Both areas also produce
the product master and the CRM that every later use case needs.
Why press-side AI still waits for the questionnaire. Ganging, scheduling and predictive
maintenance depend on whether the four Speedmasters report through Prinect, whether the Print
Beat API subscription is active on the Indigos, whether the die-cutters, gluers and Focusight
machine export anything, and whether an MIS holds routings (A2, A4). The machine list is now
known; the licences are not. If the answer is yes, UC-08 and UC-09 move forward by months. If
no, the first step is connectivity, which is an OEM conversation (MatSet for HP and Scodix,
Heidelberg Grafik Ticaret Servis for the Speedmasters, die-cutters and CTP), not an AI project.
4. Foundations
Prerequisite
Why it matters
Minimum version (0 to 90 days)
Target version
One product master
UC-04, 05, 12, 17 all read SKU-level specs: board, grammage, style, dimensions, finishing layers, weight, inks, adhesives, test references
A controlled spreadsheet for active SKUs, one owner
Master data in the MIS or ERP with change history
MIS or ERP data access
Quotes, orders, costs, stock, invoices (A1, A2)
Nightly export to a database or sheet, read-only
API or database view with access log. If an MIS is bought, require JDF/JMF and an API; Cerm, PrintVis, ePS Radius, Tharstern and Prinect Business all claim JDF (research notes, section 14)
CRM
One place for leads, RFQs, conversations, consent (A5)
HubSpot or Pipedrive with a consent field per contact (prices in UC-01)
CRM linked to MIS orders and to the n8n workflows
Press and postpress connectivity
Feeds UC-07 to UC-11 (A4)
Confirm PrintOS logins and the Print Beat API subscription with MatSet; link Prinect to the Heidelberg Portal for the free Print Shop Analytics; ask Heidelberg Grafik Ticaret Servis which of the four Speedmasters, the Suprasetter and the Heidelberg die-cutters are licensed and connected; ask KAMA about Cockpit/Job-Manager export and the Focusight supplier about data export; export one month of data
Live JDF/JMF and API feeds into a data store. Zaikio is defunct, so plan on CIP4 JDF/JMF and vendor APIs only
Document store
SOPs, certificates, dielines, proofs, test reports
Google Drive or SharePoint with folder permissions and retention rules
Indexed store with per-customer access rights
Access control and NDA constraints
Brands often forbid artwork leaving approved systems (D3)
Inventory of NDA clauses; artwork never sent to consumer chat tools
Segregated storage per brand; approved AI vendors listed in the NDA
Model choice, data residency and KVKK
Turkish KVKK, EU GDPR (Hamburg), customer NDAs. KVKK's generative-AI guide (24 Nov 2025) requires Article 9 cross-border transfer compliance for foreign LLM providers and disclosure by chatbots (UC-03, UC-13 sources)
API access with no training on submitted data, EU-region hosting where offered: Claude via AWS Bedrock or Google Vertex AI EU regions, OpenAI or Azure OpenAI EU residency, or an EU vendor such as Mistral [availability and terms to verify]; Kumru on-premise for Turkish-only documents where transfer is not wanted
Written data-handling policy per data class (public, internal, customer artwork, personal data); KVKK transfer instrument on file
Guardrails
Prevent wrong quotes and wrong claims reaching customers
Human approval on anything priced or sent externally; logging of every model call; a claims library
Test sets per use case; monthly review of errors; kill switch per workflow
5. Build, buy or partner
Tier
Recommendation
Reason
Quick wins
Build with the existing n8n + Google Sheets + OpenAI setup (add Claude where EU residency or long documents matter)
Cheap, reversible, already validated by the prospect-enrichment workflow in tools/n8n/. Same pattern reused: sheet as database, LLM chain with JSON parser, error rows, human send step. Tokens are cents per document; n8n Cloud Pro is EUR 50 per month
Core
Buy where a mature product exists: an MIS with AI order intake and JDF (Cerm with Lexis, PrintVis with Copilot, ePS Radius; Tharstern and Turkish MIS as price comparators), PitStop or Switch, Automation Engine or PACKZ, Phoenix, a CRM. Build only the glue in n8n. Partner with a Turkish integrator for ERP and e-Fatura
Custom-building an MIS or preflight engine is a multi-year mistake; the glue is where Halaman's specific rules live. Only PrintVis and PitStop publish prices; everything else needs a quote
Advanced
Partner with the OEMs (HP and Scodix via MatSet, Heidelberg Grafik Ticaret Servis in Güneşli, direct since 2001) and, for scheduling, vision or predictive work, an SME AI vendor plus a university lab under TÜBİTAK 1711 (2027 call) or 1507
Press data access, Push to Stop options and Predictive Monitoring depend on the OEM; 1711 requires exactly this consortium (doc 07, section 7)
When to involve a Turkish integrator: any work that touches the ERP or e-Fatura (UC-12, 14,
15), MIS selection and migration (UC-04, 09), on-site training in Turkish, and the KOSGEB
application file. Choose a solution partner of the ERP vendor in use (A1): Logo partners and
Ilura for Logo or Netsis agents, SAP Business One partners such as Cloudspark, Hitsoft, Be1 or
Logosoft if SAP, rather than a general software house. Turkish AI developer rates were TL 700
to 1,500 per hour in 2025, with mid-size projects quoted at TL 50,000 to 200,000
(https://yapayzekalar.org/blog/yapay-zekanin-maliyeti-ne-kadar/). Türkiye's 2026-2030 AI Action
Plan sets up "AI factories" aimed at SMEs, worth a call once they open
(https://www.aa.com.tr/tr/bilim-teknoloji/yapay-zeka-fabrikalari-kobi-ve-girisimcilere-seviye-atlatacak/4040829).
Gaps in the local supply base: no Turkish representative was found for KAMA or EyeC, and the
Esko Türkiye distributor status is unclear (UC-06).
Internal roles needed regardless of tier: one sponsor (board level), one day-to-day
programme owner (D4), one function owner per use case, and one builder. If there is no
developer in-house (D1), contract the builder for the first 90 days and train an internal
successor.
6. Costs and funding
6.1 Public price anchors
List prices as published by vendors or reseller and comparison sites, captured as search
snippets on 22 September 2026 (research/ai-in-print-packaging.md, section 14). Currencies as
published. They size the budget lines below; they are not quotes and must be re-checked before
any purchase.
Item
Public price or pricing model
Source
PrintVis packaging MIS
Implementation from USD 59,000; USD 140 per user per month (full PrintVis plus Business Central user)
Rough 12-month budget, external spend only. All figures are estimates to be replaced by
quotes; internal staff time is extra. Two scenarios: "lean" keeps the current ERP and adds
tools only; "with MIS" includes buying and implementing a print MIS. Changes from v1 are
explained in the notes column.
Item
Lean (estimate, EUR)
With MIS (estimate, EUR)
Notes
Six quick wins (UC-01, 02, 03, 16, 17 minimum, 18)
10k to 25k
10k to 25k
Builder days plus tools; Band A each; developer rates in 6.1
LLM API and n8n hosting
2k to 6k per year
2k to 6k per year
Lowered from 3k to 8k: n8n Cloud Pro is EUR 600 per year and tokens are cents per document (UC-02); the top of the range covers EU-hosted models and document-heavy retrieval
CRM licences (5 to 10 users)
3k to 12k per year
3k to 12k per year
Pipedrive Premium USD 59 to HubSpot Sales Pro USD 90 per seat per month plus USD 1,500 onboarding; metered AI agents on top
Core use cases without system purchase (UC-05, 11, 12, 13, 14, 15)
30k to 70k
30k to 70k
Band A to B each; integrator days included; no forecasting SaaS or vision station
Preflight and ganging software (UC-06, 08)
10k to 40k
10k to 40k
PitStop Pro USD 432 per seat per year is the only public number; AE, Switch, Phoenix, PACKZ and PPM are quote-only subscriptions
MIS purchase and implementation (UC-04, 09 basis)
0
60k to 150k
Band D; PrintVis anchor USD 59,000 plus about USD 17k per year for 10 users; Tharstern, Cerm and ePS quote-only; Turkish MIS likely cheaper, no public price
Press connectivity and dashboards (UC-07, 10)
3k to 15k
3k to 15k
Lowered from 5k to 20k: Print Shop Analytics free tier, Print Beat OEE included in PrintOS; PrintOS API subscription, BI seats and Predictive Monitoring contracts [to verify]
Training and change management
5k to 10k
8k to 15k
Shop-floor sessions in Turkish, Hamburg sessions in German
Contingency (about 15 percent)
10k to 25k
20k to 50k
Total 12 months (estimate)
75k to 205k
145k to 385k
Ranges, not budgets; refine after doc 09
Optional, not in totals: AI vision station (UC-11 extension)
0
0
Market anchor about USD 50,000 per station; only after Focusight and complaint data show where defects cluster
Funding that may apply (conditions, amounts and sources in doc 07, section 7):
KOSGEB KOBİ Dijital Dönüşüm Destek Programı: bank investment credit of 1 m to 20 m TL
for new software, hardware and machinery with KOSGEB paying part of the interest; usable
once; requires KOBİ status (headcount under 250 passes; the 1 bn TL turnover or balance-sheet
test must be checked against the 2025 accounts). Fits the MIS and colour-management purchases.
TÜBİTAK 1507 KOBİ Ar-Ge Başlangıç: 75 percent grant for an R&D project with technical
uncertainty, for example the defect classifier on Halaman's own reject images (UC-11) or
colour prediction across the four gamuts (UC-07). Calls in January and July.
TÜBİTAK 1711 Yapay Zekâ Ekosistem Çağrısı: Halaman as customer organisation with an
SME AI vendor and a university lab. The 2026 window closed on 18 September 2026; plan for
the 2027 call. "Smart Manufacturing Systems" is a priority area, so UC-09 and the predictive
half of UC-10 are the natural project.
7. Risks and how to manage them
Risk
What could happen
Control
Data leakage of brand artwork
Customer artwork or unreleased designs reach a third-party model or a public tool; NDA breach; loss of a nominated-supplier status
Inventory NDA clauses (D3). Artwork only in approved systems. API terms with no training on data; EU-region hosting where offered. No consumer chat tools on prepress machines. Access logs
Hallucinated quotes and specs
A drafted quote contains an invented rate or a wrong spec and is sent; margin loss or a customer dispute
The model drafts, the MIS calculates. Every price comes from a rate table. Human approval before sending. Assumptions listed on the quote. Test set of past quotes. Extraction benchmarks show about 80 percent straight-through at best (UC-02), so the check is permanent
Over-automation of customer contact in Germany
Automated e-mails to German or Austrian companies without prior express consent breach UWG §7; competitor or association claims
Consent field per contact enforced in the workflow (already built into the n8n rubric). Prefer LinkedIn, fairs and referrals for DE/AT. Human sends every first contact. Counsel reviews the outreach policy
KVKK and GDPR in AI tools
Personal data of prospects, customers or staff sent to a foreign LLM without a transfer instrument; a bot that does not say it is a bot
Follow the KVKK generative-AI guide (24 Nov 2025): Article 9 transfer basis on file, disclosure on every chatbot, data minimisation in prompts. EU residency for Hamburg data. Kumru or another on-premise model for Turkish HR documents
Change management on the shop floor
Planners and operators ignore or work around the tools; data quality falls; the project is blamed
Start with tools that remove typing, not judgement. Turkish-language interfaces. Operators name the KPIs. Weekly short review with the production manager. No tool goes live without its function owner
Vendor lock-in and vendor failure
Data trapped in a proprietary MIS, preflight or OEM cloud; price rises; no export; a platform disappears, as Zaikio did in April 2024
Contract for data export in open formats and JDF/JMF. Keep the glue in n8n, which is open source and portable. Prefer models behind a switchable API layer. No dependency on a single start-up for core data
Metered-price drift
WhatsApp bills service replies per message from 1 Oct 2026; HubSpot changed agent pricing in April 2026; token prices move
Budget per-message and per-resolution lines separately; monthly usage review; cap volumes in the workflow
Regulatory drift
PPWR delegated acts, EUDR guidance and EmpCo enforcement change after the templates are built
Version the templates. One owner watches the regulation calendar quarterly (doc 07). Legal review of declarations before first use
Programme without an owner
Pilots stall after the first demo
Sponsor plus day-to-day owner named before any build (D4). Monthly steering with the KPI table from section 2
8. Sources
Every URL cited in this chapter, listed once under the section where it first appears. All were
captured as search-engine snippets on 22 September 2026 by the research sessions (WebFetch was
blocked); none was re-fetched for this revision. Repository sources: research/ai-in-print-packaging.md
and .facts.json, research/competitors.md, research/regulation-compliance.md,
research/owner-supplied-page-extraction.md, docs/07-regulation-and-funding.md, tools/n8n/README.md.
One statement (DS Smith with Hitachi Digital Services) rests on hitachids.com without a captured page.