Evaluated 163 in-house AI task ideas, drew up a three-year roadmap of 26 tasks, and delivered three PoCs in optimization, forecasting, and LLMs.
Context
- LOTTE Fine Chemical set up an AI task force to apply AI across the business, facing unstable raw-material supply, tighter safety rules, and an aging plant workforce.
- Spent two months developing the AI roadmap, then ran PoCs for the selected tasks.
Problem
- Staff had proposed 163 AI task ideas in two rounds, with no evaluation criteria yet to rank them.
- Without a company-wide plan, tasks risked running piecemeal, with duplicate investment and no clear owner.
- The new task force had few in-house AI specialists and little hands-on experience with ML forecasting or optimization.
My role
- Led the engagement as project manager in a core team of three, running weekly and milestone reporting with the client's AI task force.
- Led three PoCs, defining the problems, selecting models, and planning how to train them.
Approach
- Interviewed executives by division and set three AI directions, covering employee productivity, business competitiveness, and data-driven decisions.
- Reviewed all 163 ideas, scored them on strategic importance and feasibility, and shortlisted the top 15%, 24 tasks.
- Classified the 24 tasks on a business impact × technology maturity matrix to decide how to run each, and wrote task definition sheets with timeline, cost, and expected benefit.
- Checked for similar tasks to combine, dependencies between tasks, and data readiness when scheduling the 2024–2026 roadmap. Added data governance and a data lake as enabling tasks.
Scroll sideways to see the whole diagram
- Ideas collected: 163 (56 + 107 in 2 rounds)
- Shortlisted: 24 (top 15% by score)
- Classified: 24 (impact × maturity). PoC 12, Proven solution 8, Development prep 4
- On the roadmap: 26 (+2 we proposed). PoC 12, Proven solution 8, Development prep 4, Task we added 2
- PoCs delivered: 3 (2 planned + 1 added). Green mixing, Pulp price, LLM platform
Scroll sideways to see the whole diagram
- Adopt a proven solution (Technology maturity High, Business impact Low): 8 tasks. Easy to do, but lower impact. Bring in a commercial tool for early wins. e.g. smart CCTV, drawing recognition
- Run a PoC first (Technology maturity High, Business impact High): 12 tasks. High impact, with proven use cases. Test the value before building the system. e.g. mixing optimization, cotton-pulp price
- Monitor the technology (Technology maturity Low, Business impact Low): 0 tasks. Lower impact and immature technology. Watch for an opening. No task fell here
- Prepare, then build (Technology maturity Low, Business impact High): 4 tasks. High impact, but the groundwork is missing. Ready the data and infrastructure first. e.g. demand forecast, profit simulation
Key decisions
| Decision | Why | Trade-off |
|---|---|---|
| Score first, then classify | Because scores only ranked the tasks, we used a business impact × technology maturity matrix to choose how to carry out each shortlisted task. | Two rounds of review with the task force. |
| A PoC before any system build | For high-impact tasks with proven use cases elsewhere, a PoC tests the value before the client commits to a full build. | Adds a stage before the benefit arrives. |
| Proven solutions first | We put commercial tools in the first half of the roadmap and scheduled tasks that need new data to start once governance and the data lake are in place. | Some high-impact tasks start only from mid-2025. |
Roadmap
Scroll sideways to see the whole diagram
Proven solution (8 tasks)
- Forklift person detection (Proven solution): 2024 Q1 – 2024 Q4
- Smart CCTV (Proven solution): 2024 Q3 – 2024 Q4
- Driverless forklift (Proven solution): 2024 Q3 – 2024 Q4
- Survey insights (Proven solution): 2024 Q3 – 2024 Q4
- Construction cost estimate (Proven solution): 2024 Q4 – 2025 Q2
- Drawing recognition (Proven solution): 2025 Q1 – 2025 Q2
- Patentability check (Proven solution): 2025 Q1 – 2025 Q2
- Research trend search (Proven solution): 2025 Q3 – 2025 Q4
PoC (12 tasks)
- Green-material mixing (PoC our team ran): 2024 Q1 – 2024 Q2
- Predictive maintenance (PoC): 2024 Q1 – 2024 Q2
- Cotton-pulp price forecast (PoC our team ran): 2024 Q2 – 2024 Q3
- Column automation (PoC): 2024 Q3 – 2024 Q4
- Raw-material price forecast (PoC): 2024 Q4 – 2025 Q1
- Visual quality grading (PoC): 2024 Q4 – 2025 Q2
- Dosing automation (PoC): 2025 Q1 – 2025 Q2
- Wood-pulp supply chain (PoC): 2025 Q2 – 2025 Q3
- Viscosity prediction (PoC): 2025 Q2 – 2025 Q4
- Water yield optimization (PoC): 2025 Q3 – 2025 Q4
- Reaction-time setting (PoC): 2025 Q4 – 2026 Q2
- Motor diagnostics (PoC): 2026 Q1 – 2026 Q2
Data and governance (2 tasks)
- Data governance (Task we added): 2024 Q3 – 2024 Q4
- Data lake and platform (Task we added): 2025 Q1 – 2025 Q4
Development prep (4 tasks)
- Profit by utilization rate (Development prep): 2025 Q2 – 2026 Q2
- Market news digest (Development prep): 2025 Q3 – 2025 Q4
- Demand forecast (Development prep): 2025 Q4 – 2026 Q1
- Competitor monitoring (Development prep): 2026 Q1 – 2026 Q2
Proven solutions go first for early wins
Data platform comes before development
Starts once data and infrastructure are ready
Outcome
- Delivered a three-year roadmap with 26 tasks across four tracks and a definition sheet for each task.
- Ran three PoCs for green-material mixing optimization, cotton-pulp price forecasting, and an in-house LLM service platform with no-code RAG and agent apps.
- Carried two PoCs into the business, with the mixing optimizer moving to a pilot and the price forecast entering quote review.
- Recommended an operating model for the task force, with a dedicated AI team, data-focused tasks, regular roadmap revisions, and an external partner network.