Frequently asked questions about industrial AI applied on the plant floor

No. We work on top of the infrastructure you already have. We connect to existing PLCs, SCADAs, MES and ERPs over OPC-UA, Modbus, MQTT, REST or flat-file exports.

If your plant runs a mix of old and modern systems, we build an orchestration layer that respects what already works and adds intelligence where it actually moves the needle. The technology adapts to you, not the other way round.

We connect to SCADAs (Siemens, Schneider, Wonderware, Ignition), historians (PI System, Wonderware Historian), MES, ERPs (SAP, Microsoft Dynamics, Odoo), maintenance CMMS, cloud IoT platforms (AWS, Azure, GCP), and relational and time-series databases.

For every client we design the lightest connector that will do the job: sometimes a SQL view is enough, other times we install a dedicated edge gateway. The question is never "can it be integrated?" but "which integration delivers the most value with the least operational risk?".

It's the most common situation on the plant floor, and it's part of the initial diagnosis work. Before modelling anything, we map the data's sources, quality, frequency and semantics.

Cleaning, enriching and contextualising what the plant already produces usually delivers a huge share of the project's final value. If needed, we instrument what isn't being measured yet and design the data model that will support not just this use case, but the ones that follow.

Both. We run 100% on-premise deployments, hybrid deployments with edge on the plant floor and cloud for reporting, and fully cloud deployments when the client wants operational lightness.

For critical, low-latency processes (in-line machine vision, real-time control) we recommend edge; for analytical or predictive use cases with a horizon of hours or days, cloud or hybrid is usually the better fit.

Wherever your IT and security team decides. By default, all sensitive data stays inside the client's perimeter. If anything is uploaded to the cloud, it's anonymised metrics for training or operational telemetry, never production recipes or competitive information, unless expressly agreed.

We sign an NDA and meet the CISO's requirements before the first data flow.

Yes. Our processes are aligned with NIS2 (the EU cybersecurity directive for essential sectors), ISO/IEC 27001 practices, and GDPR requirements for personal data.

We provide processing documentation, access logs, encryption at rest and in transit, and granular role management. If your internal compliance team needs specific matrices filled in, we do that before kick-off.

No. Every client operates in an isolated environment, with its own credentials, its own model and its own history. We never cross-reference data between clients, and we never reuse a model trained on your plant at a competitor's plant.

The intellectual property of the model and the data belongs to the client, and that's set out in the contract.

There's no artificial vendor lock-in. We hand over the model's source code, deployment documentation, data pipelines and retraining scripts. If you want to bring the system in-house, we train your team to run it without us.

Our contract lives or dies on the value we deliver every month, not on a locked door behind you.

It depends on three variables: the complexity of the use case, the quality and volume of the data, and the level of integration required. Every project starts with an exploratory meeting where we understand the process and scope the use case; that produces a diagnosis with a defined scope and deliverables.

From there, pilots usually fall within a contained range, and full deployments are modelled as an investment with measurable returns, not as a fixed cost.

In energy and thermal optimisation and predictive maintenance cases, we've seen returns in under 12 months once in production. What matters isn't when the pilot starts, but when the plant starts acting on the model's recommendations.

That's why we measure not just technical accuracy but operational impact: kWh saved, downtime avoided, OEE recovered.

With a bounded 6-to-10-week proof of concept on your plant's real data. We define a measurable KPI (for example, "reduce electricity consumption per tonne on a specific line"), deliver the model and validate the result against historical data.

If the pilot doesn't demonstrate impact, we don't move to deployment. It's honest, and it saves months of internal budget discussions.

Yes. The pilot is designed so both sides come out with a clear answer: either the use case delivers value and we scale it, or we learn together what doesn't work and redirect.

Either answer is a valid one. What kills projects is the ambiguity of "let's see how it goes".

We combine whatever the problem needs. Classical time series and gradient-boosting models for tabular predictions, neural networks for machine vision and complex patterns, and physics-informed models (PINNs) when there are physical laws the model has to respect: thermal, fluid, energy. On top of that, causal analysis, neuro-symbolic AI and knowledge graphs when the question isn't just what's going to happen, but why it's happening.

We don't do AI because it's trendy: every architecture answers a specific business question.

We apply explainable AI (XAI) techniques so every model recommendation comes with its reasoning: which variables weighed most, which conditions trigger the alert, how confident the model is in the prediction.

We go a step beyond correlation: we look for the physical and operational cause that explains what's happening, because that's what lets you act on the process. It's key in regulated sectors and, above all, for the operator to trust what the system suggests on the plant floor. A model nobody understands is a model nobody uses.

Less than people usually think. In stable processes with good history, 6 to 12 months of data is usually enough for a first production-ready version.

In processes with little history or frequent changes, we combine physics-based models with limited data and enrich them as real history accumulates. Starting with the right use case is worth more than waiting to have "all the data".

Every DDS solution includes monitoring of the model's performance in production: we detect data drift, concept drift and accuracy drops before they turn into operational errors.

When degradation is detected, we run a controlled retraining with recent data and validate it against criteria set from the design stage. The model evolves along with your plant.

We work in one specific category: Industrial Process Intelligence. We don't apply generic models to an industrial process: we start from that process's real physics and operations to make it explainable, optimisable and sustainable. Hence our principle: understand before you optimise.

That work is concentrated on three fronts: energy and thermal optimisation, process stability and quality and industrial decision intelligence. To support them we combine causal analysis, physics-informed models, neuro-symbolic AI and knowledge graphs.

The practical consequence is that we don't hand over a dashboard with loose predictions, but an explanation of why the process behaves the way it does and which levers move the outcome.

A pilot runs for 6 to 10 weeks. The move to production depends on the complexity of the deployment and the client team's availability, but we usually close out production deployments between month 3 and month 6 from the pilot's kick-off.

Our goal is to show value early, not to build endless laboratories.

Not to get started. For the pilot and the first use cases, a business sponsor and a technical point of contact (plant, IT or energy) plus access to the data is enough.

To scale to several use cases, we recommend designating someone in-house as the AI programme owner, and we support them as an extended technical arm. That's exactly the model behind our Outsourced AI Department service.

Focused time from a sponsor and a technical point of contact, access to the data, and availability to validate results on the plant floor. We don't ask them to learn AI: we ask them to bring process knowledge.

It's precisely the client's own domain knowledge that separates a model that works in production from a nice-looking experiment with no real traction.

Three layers: continuous monitoring of the model's health and the data flows, reactive support with an agreed SLA, and periodic optimisation cycles where we propose model improvements or related new use cases.

The idea isn't to hand you a system and disappear. It's to stay with you while the system keeps learning from your plant and your business.