From pilot to production: a case study on digital twin implementation in CDMOs

Strategy is easy to agree on. Implementation is where organisations diverge.

The gap between an “approved digital twin business case” and a “validated twin in GxP production use” is wider than most teams anticipate — and it is littered with the wreckage of projects that started with enthusiasm and ended as expensive shelfware.

Here is what the journey actually looks like, in concrete terms.

Four phases. Realistic timelines. Real numbers.

Phase 0 — Discovery (2–4 weeks)

Before selecting a vendor, before scoping a pilot, before committing a budget — answer four questions honestly:

1. Is bioreactor process data automatically logged to a historian at ≥1-minute intervals, or still transcribed from paper charts?
2. Can engineers retrieve 50 historical batches programmatically, or does it require a two-week IT ticket?
3. Are online process data, offline analytical results and batch disposition records linkable by a common batch ID?
4. Does your QMS have a CSV procedure aligned with GAMP 5 that covers AI/ML-based tools?

Score these. If you are below 20/30, data infrastructure work must precede the pilot by 3–6 months. Skip this step and you will discover it mid-pilot, at maximum cost.

Phase 1 — Pilot (3–6 months)

Single unit operation. Shadow mode only. Define success criteria before you start — defining them after seeing results is a validation red flag.

Phase 2 — Deployment (6–12 months)

Full integration: OPC UA connectivity to the historian, a LIMS API for automated lab result ingestion, MES event integration. Formal qualification: IQ/OQ/PQ. User training across four groups: operators, engineers, QA reviewers and management.

Phase 3 — Scale (12–24 months)

Multi-operation, multi-product, then multi-site.

The IQ/OQ/PQ essentials

IQ confirms correct installation:

• Software version matches the approved URS specification
• OPC UA connectivity to the historian confirmed
• Audit trail configuration enabled and functional
• System clock NTP-synchronised

OQ confirms the system operates as designed:

• Online sensor data appears in the dashboard within 60 seconds of a process event
• A prediction returns with uncertainty bounds from a 7-day process history input
• An anomaly alert is generated and logged within 30 minutes of fault-pattern detection
• The model produces identical results on identical inputs (determinism check)
• Role-based access prevents a read-only user from modifying model parameters

PQ confirms performance in operational context:

• Titer MAPE ≤ 5% on 8 prospective PQ batches
• Anomaly detection sensitivity ≥ 80% on a retrospective validation set
• System availability ≥ 99.5% over a 90-day evaluation period
• Data latency ≤ 2 minutes, 95th percentile

Change control for model updates

Every model update flows through change control. The model’s validated state is a specific, documented configuration. (The typical pilot KPIs and the three-tier change-control framework are shown as tables in the original post.)

The CDMO case study: what actually happened

Starting point (Q1 2023): 8 × 2,000 L bioreactors, ~60 batches per year, 30% of batches requiring a deviation investigation, an 18-month average tech transfer timeline.

Months 1–2: The data audit revealed historian data for all bioreactors, but only 34 of the expected 50+ batches had complete offline metabolite profiles — the rest sat in paper-based sample tracking never entered into LIMS. QA was engaged to prioritise LIMS remediation. The pilot was scoped to the single largest-volume product (22 batches available).

Months 3–8: A hybrid model trained on 12 clean batches achieved a titer prediction MAPE of 6.8%. Two anomaly false positives were traced to pH calibration events — resolved by adding a calibration exclusion flag.

Months 9–18: Full IQ/OQ/PQ. Final PQ results: MAPE 5.9%, anomaly sensitivity 90%, a false-positive rate of 1 per 14 batches.

First 6 months after go-live:

• Batch investigations requiring senior scientist input: 40% reduction (30% → 18% of batches)
• Mean time to anomaly detection: 28 minutes (vs. more than 8 hours with manual monitoring)
• One tech transfer compressed from 18 to 11 months

Total programme cost: €680,000, including data infrastructure remediation, vendor licence and services, validation and ~1.8 FTE internal effort. Projected break-even: 24 months from go-live.

Five failure modes — all preventable with upfront design

1. Data quality surprises mid-pilot.
2. Scope creep.
3. QA not engaged from Phase 0.
4. Operators who see the twin as surveillance rather than support.
5. Over-reliance on the vendor without building internal capability.

What is your organisation’s current GAMP 5 readiness for AI/ML validation — and has it been stress-tested against an actual pre-approval inspection?

Originally published on LinkedIn on 11 May 2026.