Trusting demo accuracy over production accuracy
A model that scores 98% on a curated test set may drop to 70% under sensor drift, lighting changes or batch variation. Always request accuracy measured on live plant data over at least 30 days.
§ Due diligence · Technology assessment · Vendor selection
Industrial AI due diligence checklist: what to assess before investing.
Before committing capital to an industrial AI company, platform or internal initiative, assess six dimensions: technology maturity, data infrastructure, team capability, integration with plant systems, IP and data ownership, and deployment evidence. The checklist below structures that assessment into a work sequence that an investment committee, acquisition team or technology selection panel can apply directly.
Whether you are acquiring an AI startup, selecting a platform vendor or approving an internal initiative, the same six dimensions determine whether the technology will survive contact with a real plant. This checklist is drawn from practitioners who have conducted technical due diligence on industrial AI deployments across automotive, pharmaceutical, food and energy manufacturing.
ForInvestors, acquirers, technology selection committees
ScopeTechnology · Data · Team · Integration · IP · Deployment
UseStructured checklist with go / no-go criteria
01 The assessment framework
Six dimensions that determine whether industrial AI is production-ready.
Industrial AI due diligence differs from generic software due diligence because the technology must function inside a physical production environment. A model that achieves high accuracy in a notebook is irrelevant if it cannot run on the plant network, tolerate sensor drift, or integrate with the MES and SCADA systems that operators already use. The framework below organizes the assessment into six dimensions, ordered from technology fundamentals to deployment evidence.
Each dimension produces a go / conditional / no-go signal. A single no-go on data ownership or deployment evidence is typically deal-stopping. Conditional signals on technology or team can be mitigated with a post-acquisition plan, provided the buyer has the internal capability to execute it.
For a deeper treatment of the investment perspective, see our industrial AI project ROI guide. If you are selecting a vendor rather than acquiring a company, the same checklist applies — simply replace team assessment with vendor capability assessment.
02 The checklist
Due diligence checklist by dimension.
| Dimension | What to verify | Go / No-go signal |
|---|---|---|
| 1. Technology maturity | Model architecture, training methodology, inference latency on target hardware, model versioning and retraining pipeline. Is the model reproducible from source data and code? | Go: reproducible pipeline with versioned models. No-go: model exists only as a trained artifact with no retraining path. |
| 2. Data infrastructure | Source of training data, sensor contracts and licensing, data labelling process, data pipeline reproducibility, storage and compute costs at scale. Who owns the data after acquisition? | Go: owned or licensed data with clear provenance. No-go: data sourced from third parties without transferable AI usage rights. |
| 3. Team capability | Has the team deployed AI in a production plant before? Are ML engineers paired with automation and process engineers? Is there documented ownership of the model in production? | Go: team includes members with plant-floor deployment experience. No-go: team is research-only or has never shipped to production. |
| 4. Integration with plant systems | Does the AI connect to MES, SCADA, PLC and ERP through documented interfaces? Is edge or cloud deployment specified? What happens when the network or the model is unavailable? | Go: integration tested with live plant data. No-go: integration demonstrated only on synthetic or historical data. |
| 5. IP and data ownership | Patents, trade secrets, model weights ownership. Open-source dependencies and their licences. Data generated by the system in production — who owns it? | Go: clean IP with no copyleft risk to the acquirer's stack. No-go: core model built on open-source weights with restrictive licences. |
| 6. Deployment evidence | Is there a live deployment in a real plant? What is the measured impact on throughput, quality or downtime? Can you visit the plant and see it running? | Go: live deployment with quantified results and site visit possible. No-go: deployment claimed but not demonstrable. |
03 Risks and common errors
Where due diligence goes wrong.
Ignoring data ownership and licensing
Training data sourced from a customer or partner may not be transferable after acquisition. Sensor data contracts often restrict AI use. This is the most common deal-stopper discovered late.
Assessing the team on CVs instead of deployment history
A strong research pedigree does not equal production capability. Ask for specific deployments the team members personally shipped, including the plant, the integration and the measured outcome.
Skipping the plant visit
Architecture diagrams and slide decks can hide integration gaps that a one-hour plant walkthrough exposes. If the target cannot arrange a site visit to a live deployment, treat that as a signal.
Overlooking model retraining and maintenance
A model deployed today will degrade as equipment wears and processes change. Without a retraining pipeline and an owner for model maintenance, the technology becomes a liability within months.
Underestimating integration complexity
Connecting AI to MES, SCADA and PLC is not a software task — it involves safety, network segmentation and operational procedures. Due diligence must verify that integration has been tested, not just designed.
04 Frequently asked questions
About industrial AI due diligence.
How long does an industrial AI due diligence take?
A technology-focused assessment typically takes two to four weeks. A full M&A due diligence covering team, IP, data and deployment can extend to six or eight weeks depending on access to documentation and the target's cooperation.
What is the most overlooked area in industrial AI due diligence?
Data infrastructure and ownership. Many targets demonstrate a working model but cannot show who owns the training data, whether sensor contracts permit AI use, or whether the data pipeline can be reproduced after acquisition.
Should due diligence cover model explainability?
Yes. In regulated industries such as pharmaceutical or aerospace manufacturing, model explainability is a compliance requirement. Even in less regulated sectors, an opaque model that operators cannot interpret will face adoption resistance on the plant floor.
Can due diligence be done remotely?
Document review, code audit and architecture assessment can be done remotely. A site visit is recommended for evaluating deployment maturity, integration with physical infrastructure and the reality of day-to-day operations.
What red flags should stop a deal?
A model that only works on historical data with no live retraining path, a team with no production deployment experience, data sourced without clear licensing, or integration claims that cannot be demonstrated in a live plant walkthrough.
Related resources: our industrial AI due diligence service for private equity provides independent expert assessment for investment decisions, and our guide to choosing an industrial AI consultant covers vendor selection from the buyer's perspective. For sector-specific considerations, see our pharmaceutical manufacturing and automotive industry pages.
05 Request experts
Need an independent expert for your assessment?
Tell us the target, the sector and the stage of the transaction. We propose practitioners who have conducted industrial AI due diligence in comparable plants.