CHAPTER 3 · THE GAP
Twelve systems away from a decision
The science is captured. The enterprise is not. Every scientist pays the tab-switching tax.
SITUATION TODAY
What 'is this formulation worth scaling?' costs today
When Dr Stefan Koen finishes a HELIX trial, deciding 'is this worth scaling?' opens seven tabs in four hours: PfS for the result, ERP for the supplier AVL, PLM for any existing part, QMS for the control plan, REACH/ESPR for hazard flags, the LCA spreadsheet, and an email thread with regulatory. Each tab is a different login, vocabulary and ownership.
7
Tabs opened per scaling decision
Average across 5 observed sessions
4h
Scientist time per decision
Loaded ~$480 per decision
~8,900
Decisions per year (AVY R&D)
1 per result · multiple per programme
$4.3M
Annual tab-switching cost
Pure scientist time burnt
~120
Duplicate experiments / year
Same composition family · different scientist
$2.8M
Cost of duplicates
Lab cost + instrument time
BENEFIT
12 silos → 1 governed graph · $4.3M back
60× faster decisions · always cited · duplicates pre-empted · tribal knowledge preserved
PfS is one node. Twelve more decide the launch.
PfS
System of scientific record
ERP
Material codes · AVL
PLM
Part numbers · BoM
MES
Plant runnability
QMS
Control plans
Supplier ESG
Sustainability rating
REACH / ESPR
Hazard & DPP rules
LCA model
Carbon · water · waste
Complaints
Field failure data
Patents
IP corpus
SharePoint
Memos · slides
Lab PDFs
2019 trials · vendor specs
Retiring brains
3 retire in 12 months
■ Structured (8)
■ Unstructured (3)
■ Tacit (1)
HOW IT WORKS
How SCIKIQ collapses 12 silos into one governed graph
1
Ingest each system on its own clock
PfS via OData (5min), ERP via SAP CDC (15min), PLM nightly, REACH/ESPR feeds daily, complaints + supplier ESG weekly. SharePoint + lab PDFs via embedding-indexed crawl.
▣ Active metadata · change-data-capture · embedding-indexed RAG
→
2
Resolve entities into one graph
A 'Material' in PfS == a 'Material code' in SAP == a 'Part component' in PLM == a 'CAS' in REACH. ML-based entity resolution joins them; lineage edges preserved.
▣ Embeddings + rule-based resolver · property graph (Neo4j-class)
→
3
Govern + serve answers with citations
Every query routes through the semantic layer; every answer is grounded in source rows with click-through to the originating system. No hallucination · no orphan data.
▣ RDF semantic layer · governed RAG · provenance per token
BEFORE / AFTER
Before / After · per scaling decision
| Dimension | Today | With SCIKIQ + Thermo | Δ |
|---|---|---|---|
| Time per decision | 4 hours | 4 minutes | −60× |
| Tabs / systems opened | 7 | 1 (Copilot) | −6 |
| Citations to source data | 0 | Always | + provenance |
| Duplicate experiments caught | rarely | Pre-flight check | ~120/yr saved |
| Cross-region knowledge reuse | ad-hoc | Built-in | Mentor↔Leiden↔Suzhou |
WHO WINS
"I stop being a tab-switcher and start being a scientist. The question 'has anyone done this?' answers itself before I open a notebook."
— Every scientist · every project manager · Across MG + SG + Embelex
⚠ IF WE DON'T
Compounding cost of the gap
- $4.3M / yr in scientist time burnt switching tabs
- $2.8M / yr in duplicate experiments nobody knew about
- $1.2M / yr in regulatory rework when SharePoint memos miss a substance restriction
- + unmeasured: 3 senior scientists retire in 12 months · their tacit knowledge walks out
12systemsTo make one scaling decision
4hper decisionAverage scientist tab-switching
$4.3Mannual costPure tab-switching · 8,900 decisions
1missing layerGoverned graph that joins all 12
After a formulation lands in PfS, the next decision needs twelve other systems — and a few that aren't systems at all. PfS is excellent at the lab; by design it is not where the lab meets the enterprise. That meeting place is where SCIKIQ goes.
12 systems · 8 unstructured sources · 1 missing layer.
AI / ML / LLM IN THIS CHAPTER
DATA SOURCES