AI produktivitet mätning kräver systematiskt framework visar IU International University studie (DIGITAL 2025): många företag proklamerar productivity gains men saknar structured approach calculate real values. För dig som projektledare betyder detta: McKinsey 2023 studie uppskattade 2.6-4.4 billion dollar worldwide GenAI productivity gain + 0.1-0.6% yearly rise, men 84% German companies förväntar 12% national gains medan 70% managers estimerar endast 8% own company (lägre än nationwide). Framework klassificerar use cases TRE dimensioner: (1) TID (planning, design, production, testing, delivery, support time), (2) KOSTNAD (material usage, waste, space), (3) KVALITET (inspection, anomaly detection, diagnosis, proofreading). TCO (Total Cost of Ownership) approach täcker AI system lifecycle: design/development, customizing, integration controlling, deployment/documentation, user training, operation, training cycles, risk management, certification compliance. Case studies visa: Analytical AI (Salling Group energy 700 supermarkets millions saved, Holstebro municipality 146,000$ savings, orthopedical insoles >70% plastic material saved), Generative AI (heise 5-15min/picture annotation time saved 300€ total, Fieldcode 50% spare parts saved, Klarna 2.3 million chats = 700 employees work + 40 million$ profit increase, BCG 12% more tasks + 25% time savings + 40% quality increase). Integrated framework jämför productivity/profitability gains SUM mot AI system TCO för reliable quantification. Kritisk insight: time/cost savings GenAI may cause quality reductions – journalist produces larger output samma salary men originality/quality viktigt för artefacts.
Problem statement: Proclaimed gains utan structured calculation
Ambitiösa förväntningar men saknad verifiering:
McKinsey 2023 studie (850 jobs, 2,100 tasks):
- Worldwide productivity gain: 2.6-4.4 billion dollar
- Yearly rise: 0.1-0.6%
- Branches mest gynnade: finance, high-tech, media, bioscience
- 75% potential: customer service, sales/distribution, software development, R&D
- Task type clusters: Physical work 70%/34% gains, Data collection 79% (vs 65% analytical AI only), Data management 92% (vs 75% analytical only), Decision making/collaboration 50-55% (nästan 3x analytical AI only)
- Surprising correlation: productivity gains korrelerade med education level – lägre qualifications såg mindre possibilities
ifo German companies 2024 (84% förväntar gains):
- National economy: 12% increase average estimation
- Own company: 8% increase (70% managers)
- Paradox: estimating LOWER values för own company än nationwide
Hammermann et al. 815 companies (2022-2024):
- 45% employees using AI claim productivity gains
- 15% stated OPPOSITE
- Demonstrates lack clear assignability
Demary et al. Institute German Economy (IW):
- GDP rise 0.9% från AI-usage 2025-2030 (moderate, NOT strong growth driver)
- 2030-2040 decade: 1.2% predicted
- AI som complementary till human work
- Question: rises accomplished by AI alone ELLER only if flanked by other digitalization (robotics, software, internet etc.)?
Central research gap: Time savings ofta cited MEN insufficient ask employees general personal impression. Unclear whether attributed till AI alone. Research question: what kind values can be measured and in which dimensions?
Framework tre dimensioner: Tid, kostnad, kvalitet
AI klassificering (three types):
1. Analytical AI: Data-driven decision-making, pattern recognition, predictive analytics. Medical diagnosis, fraud detection, predictive maintenance, algorithmic trading, sentiment analysis.
2. Generative AI: Create new synthetic content (text, images, music, videos). GANs för videos, Transformer-based (GPT4.x, LLMs) för natural language generation.
3. Reactive AI: Real-time inputs + predefined rules, no memory/learning runtime. Chess systems (Deep Blue), older rule-based chatbots.
DIMENSION 1: TID (Time)
- Planning time: Project/product planning, logistic optimization (airports, freight, harbors, railways) – A+G
- Design time: Product/service design, individualization, protein structures (AlphaFold), recipes – A+G
- Production time physical goods: time/piece – A
- Production time immaterial artefacts: text, audio, video creation (journalism, marketing, consulting, arts), program code – G
- Testing time: test cases creation, automatic execution/evaluation – A+G
- Delivery time: demand forecasting, route optimization – A
- Support time: service requests NL recognition, speech-to-text, customer requests/feedback – G
DIMENSION 2: KOSTNAD (Cost)
- Material usage: raw materials, supplies, energy – A
- Waste/offcut: raw materials – A
- Required space: inventory optimization – A
DIMENSION 3: KVALITET (Quality)
- Quality inspection: automatic anomaly detection production, medical diagnosis (skin cancer, tumor detection X-rays/MRTs) – A+G
- Proofreading, stylistic improvement texts, translation – A+G
Kritisk varning GenAI: Time/cost may trend DIFFERENT direction än quality. Example: GenAI saves time producing text/illustrations → employee produce larger output given time (samma salary) → MEN quality/originality också important för artefacts. Just speeding things up may lead counterproductive effects long run.
Case studies documented productivity gains
ANALYTICAL AI examples:
Salling Group (Energy consulting, 700 supermarkets): Smart meters/energy providers data → AI analyses weather data + consumption + optimizes device usage closing hours → Savings millions reported
Holstebro municipality Denmark (Community buildings): AI analyses weather + consumption + optimizes closing hours usage → Savings: 1 million DKK (~146,000$)
SWMS Systemtechnik (3D printing composite materials): AI-supported monitoring: image-based object recognition/segmentation → Material savings 1/3 estimated + energy savings robot/cooling
Orthopedical insoles (3D printing): AI calculates ideal form → Material savings >70% plastic, up to 60% energy
FRAPORT AG Aviation (Ground handling): AI system IDA simulates/optimizes staff planning qualification/availability → No data yet (beta-status)
GENERATIVE AI examples:
heise (IT magazine publisher, 14,000 pictures CMS): System heiseIO (LLM + process-oriented approach predefined prompts) → 5-15min human annotation time saved/picture, total cost 300€. Daily “Botti” newsletter 12 minutes faster = 1.5 person days/month saved
Fieldcode GmbH (Field service planning): LLM-based ticket diagnosis: analyze requests solve remotely instead sending technician → Up to 50% spare parts saved, First fix rate rise
Klarna (Internal knowledge management system Kiki): LLM answers 2,000 questions/day employees, 85% staff usage → Contracts 10 minutes instead 1 hour. Chatbot 2.3 million chats customers = 700 employees work. Estimated profit increase: 40 million$
Boston Consulting Group (General AI usage): 12% more tasks accomplished, 25% time savings, 40% quality increase
Observation: NONE listed case studies reported cost side data (TCO). Only savings mentioned, NO full lifecycle costs.
TCO approach AI systems lifecycle
För get overall TCO picture, cost components organized lifecycle phases:
1. System Design & Development: Modeling, training (supervised/unsupervised/reinforced/deep learning) → large storage/compute (GPUs, AI-chips). Data collection/generation, cleaning, consolidation. Interfaces control/monitor. Integration existing processes. Cost determinants: Programming time, data engineering (salary), on-premises hardware/cloud CPU/GPU time, energy cost. Prefabricated models: covered by subscription fees.
2. Customizing: GenAI finetuning specific tasks/domain. Guard rails responsible/ethical AI use. Cost: Salary (programmers, domain experts), domain-specific datasets acquisition, computing resources.
3. Integration Controlling: KPIs, performance metrics developed monitor system, assess outcomes. Cost: Time choose/agree KPIs (AI engineers, domain experts, executives involved).
4. Deployment & Documentation: Technical setup, integration existing systems/platforms, testing. Documentation facilitates maintenance/troubleshooting/development. Cost: Technical team time.
5. User Training: Initial training AI opportunities/risks/legal compliance (EU AI act required). Curriculum depends tasks assigned (IT department, AI team, HR etc.). Cost: Salary, course fees, materials, travel/accommodation.
6. Operation: Each request causes inferencing costs. Subscription fees (user licenses monthly/per-token basis). Critical systems: human-in-loop ethical requirements. Cost: Subscription fees (fix monthly/annual), token consumption (variable), salary.
7. User Training Cycle: Repeated intervals keep users up-to-date, refresh knowledge. Cost: Salary, course fees, materials, travel/accommodation.
8. Risk Management: Additional insurance potential damage AI failures, monitoring frameworks assess/mitigate risks. Cost: Salary, insurance premiums.
9. Certification Compliance: Legal/regulatory standards (EU AI Act), renewed prescribed intervals. Reassessment/auditing systems. Cost: Certification fees (annual/long-term), internal/external experts salary.
TCO structure användning: Calculate concrete AI project → estimated total costs → calculate amortization time.
Integrated framework: Benefits vs costs comparison
Profitability gains beräkning: Sum all increases measuring difference (∆) between time, material cost, quality × proper computing unit (salary/hour, price/unit). Quality rise prediction difficult – medical environment: follow-up costs incorrect diagnosis would be med AI, eller future purchases more satisfied customer.
AI system TCO calculation: More data needed från past projects + continuous controlling. Companies search benchmarks + share experiences. AI components become standardized “software off shelf” → easier accomplish.
Final comparison: Two sums eller magnitude order compared → judgment whether benefits outweigh costs + by what amount. Prevents companies running blindly into AI projects won’t pay off – too many aspects unnoticed before start.
Praktiska takeaways projektledare
1. Demand structured calculation before investment: NOT rely på employee impressions eller general claims. Use framework three dimensions (time, cost, quality) measure concrete values.
2. Apply TCO lifecycle thinking: Development/customizing costs only början. Include training cycles, risk management, certification, operation subscription fees. Many projects underestimate long-term costs.
3. GenAI quality trade-off awareness: Time/cost savings may cause quality/originality reductions. Journalist larger output ≠ better content. Monitor actual artefact value NOT just production speed.
4. Benchmark against case studies: Compare planned use case till documented examples (Klarna chatbot = 700 employees, heise 1.5 person days/month saved). Realistic expectations prevents disappointment.
5. Collect own company data: Contribute benchmark databases (industry associations, chambers commerce). Measure KPIs från start – enables future projects better estimation.
Källa: “Artificial Intelligence – Myth or Measurable? A systematic framework to determine AI-induced productivity gains” av Sibylle Kunz & Claudia Hess, IU International University of Applied Sciences, publicerad 2025 (DIGITAL 2025: Advances on Societal Digital Transformation).