Table of Contents
Keep a first industrial data team small and centred on one plant problem: a plant engineer who owns it, a data engineer who makes historian, lab and maintenance data usable, and a data scientist who builds and tests models, with part-time OT and IT support. Add ML engineers and more data scientists once a model is in daily use.
Key takeaways
- Data work dominates early projects. Data scientists in Anaconda's 2020 survey reported spending 45% of their time loading and cleaning data (InfoWorld), and RAND concluded that AI products need "more than just a data science team" (RAND).
- The engineer who owns the problem matters as much as the data scientist. Data scientists "are rarely experts in the topics for which they are building their models" (RAND).
- Start centralised and build on existing teams. Microsoft's guidance is that organisations early in AI adoption benefit from a centralised team, moving to an advisory role later (Microsoft).
- Give the team one problem for at least a year. RAND recommends this commitment before any AI project starts.
- Combine hiring, training and partners. In the WEF's 2025 employer survey, 85% planned to prioritise upskilling and 70% expected to hire staff with new skills (WEF).
Why build a team instead of buying projects?
A vendor can build a model. Someone still has to own it when the feed changes or the contract ends. Google researchers wrote that it is common to incur "massive ongoing maintenance costs in real-world ML systems" (Sculley et al., 2015). Joint guidance led by CISA asks plant owners to understand "the organization's current capacity for maintaining an AI system in their OT environment" before adopting one (CISA et al., 2025).
Knowledge of the data is the other reason. RAND notes that, without good documentation, losing a data engineer can mean that "no one knows which datasets are reliable" (RAND). In a plant, that knowledge covers tag names, replaced instruments, lab timestamps and shutdowns hidden in the historian. It should stay in-house.
Which roles does a first industrial data team need?
A survey of machine learning in the process industries describes a soft sensor project team as a panel operator, a process engineer, a control engineer and a project manager, with the process engineer writing the project charter (Lawrence et al., 2024). A data team adds the people who turn plant data into models and keep them running.
| Role | What they do in a plant | Usual source | When |
|---|---|---|---|
| Problem owner (process, maintenance or quality engineer) | Defines the decision and the KPI, explains tags and events, judges results | Your own staff | From day one |
| Data engineer | Extracts data from the historian, LIMS, CMMS and ERP; cleans it, aligns timestamps, builds repeatable pipelines | Hire, or retrain IT or automation staff | From day one |
| Data scientist | Builds models and tests them against a baseline | Hire, or a partner at first | From day one |
| OT or automation engineer | Data access to the DCS, SCADA and historian; network and safety rules | Your own staff, part-time | From day one, part-time |
| IT and security | Servers, accounts, backups, security review | Your own staff, part-time | From day one, part-time |
| ML engineer (MLOps) | Deploys, monitors and retrains models in production | Hire, or a partner | When a first model reaches a pilot |
| Sponsor | Sets priorities, protects the team's time and budget | A plant or operations manager | From day one |
Microsoft's guidance for an enterprise AI centre of excellence adds AI governance, AI security and AI operations specialists, alongside business leaders who identify use cases and evaluate results (Microsoft). A first plant team will not have all of these. Cover the missing roles part-time, from IT and OT, or through a partner.
The problem owner is not optional. RAND found that data scientists need subject-matter experts to explain what the data means and which fields are unreliable, and 10 of its 50 industry interviewees named a lack of domain understanding as a cause of failure (RAND).
Why hire a data engineer before a second data scientist?
Most of the early work is data work. In Anaconda's 2020 survey, data scientists reported spending 19% of their time loading data and 26% cleaning it (InfoWorld). One of RAND's interviewees said that "80 percent of AI is the dirty work of data engineering", and 30 of 50 discussed persistent data quality problems (RAND).
Plant data adds its own work. Process data is "often abundant but poor in information" because of disturbances, outliers and missing values, and quality samples taken once per 8-hour shift or once a day must be aligned with process data recorded far more often (Lawrence et al., 2024). Maintenance histories are often free text in the CMMS, and some records are still on paper. Our soft sensor guide shows how much of a model project is time alignment.
Deployment is the next gap. Google Cloud notes that data scientists "might not be experienced software engineers who can build production-class services" (Google Cloud). RAND's interviewees recommended hiring ML engineers to build deployment infrastructure (RAND).
Plan for retention. RAND reports that many data engineers aim to move into data science, causing high turnover in some organisations. Give the role a career path and make documentation part of the job.
Where should the team sit in the organisation?
| Placement | Strengths | Risks |
|---|---|---|
| In the IT or digital department | Servers, security, ERP access, IT governance | Far from the plant; OT access is harder; projects chosen for the technology |
| In operations or process engineering | Close to the problems and the operators | Seen as one site's tool; weak on infrastructure and security |
| Central team serving several sites | One set of tools and standards; scarce skills are shared | Becomes a bottleneck; members lose touch with the sites |
| Data staff embedded in each site | Fast, close to users | Duplicated tools; isolated individuals who are hard to retain |
Microsoft's guidance is that companies early in their AI journey "benefit from a centralized CoE to consolidate expertise and foundational practices", then move toward an advisory role as adoption matures. It also advises building on existing teams, and creating a standalone AI team only if current teams cannot support AI adoption (Microsoft). For an industrial company, a practical reading is a small central team, reporting to a manager with authority over both operations and IT, whose members spend regular time on site.
Keep the builders together. Sculley et al. trace integration problems to "overly separated 'research' and 'engineering' roles" and suggest embedding both in the same team (Sculley et al., 2015). The same applies to data staff and OT engineers: the CISA-led guidance recommends cross-disciplinary collaboration and training OT teams to interpret and validate AI outputs (CISA et al., 2025). Our post on where AI fits in the Purdue model covers the network side.
Should you hire, train or use a partner?
Usually all three. Microsoft's AI adoption guidance recommends combining training, hiring and partnerships (Microsoft).
The market for these skills is competitive. The WEF's Future of Jobs Report 2025, based on a survey of over 1,000 employers, puts AI and big data at the top of the fastest-growing skills and big data specialists among the fastest-growing jobs. Of the employers surveyed, 63% see skill gaps as a major barrier to transformation (WEF).
Algeria has a new local pipeline: the National Higher School of Artificial Intelligence (ENSIA), created in 2021, graduated 105 students in June 2026 after five years of training (Radio Algérie). Graduates bring the methods; plant knowledge comes from your engineers.
- Hire for data engineering and data science, where you need the skills full time.
- Train engineers and automation or IT staff who already know your systems, and protect part of their time for data work.
- Use a partner for the first project, ML engineering or specialised models. Write knowledge transfer into the contract, and check how the partner will access and use your OT data, as the CISA-led guidance advises. Our vendor selection guide lists the questions to ask.
What should the team do in its first year?
- Pick one enduring problem. RAND recommends committing each team to one problem for at least a year; a project not worth that commitment is probably not worth starting (RAND).
- Check readiness. An AI readiness assessment shows whether the historian, tags and records can support the problem.
- Define the metric and the baseline. Rule #2 of Google's Rules of Machine Learning is "First, design and implement metrics." Compare with a rule or a calculation before building a model.
- Build repeatable data access. A scripted extract from the historian, LIMS and CMMS, with a data dictionary for every tag used.
- Keep the first model simple. Google's Rule #4 is "Keep the first model simple and get the infrastructure right."
- Pilot with the users. Machine learning "is often met cautiously by the end users", and involving them early builds trust (Paleyes et al., 2022).
- Hand over with an owner. Monitoring, a retraining trigger and a fallback, as described in why industrial AI projects fail.
- Only then pick the second problem. Reuse the pipelines and documentation from the first.
Budget the team's time as its own line; see industrial AI project cost.
Which mistakes slow a new data team down?
- Hiring only data scientists. Nobody owns the data pipelines or the deployment.
- Changing priorities every few weeks. RAND found that projects can then be discarded before they show results, and that many leaders expect AI projects to take weeks instead of months.
- Chasing the newest tool. RAND describes failures where technical staff try new models when established tools fit better. Google's Rule #1 is "Don't be afraid to launch a product without machine learning."
- Keeping the team away from the plant. It misreads tags and builds tools operators ignore.
- Outsourcing everything. If the partner leaves with the knowledge, you start again from zero.
What we have seen at KABAS AI
Our two published projects are formulation R&D projects, but they show where a client's own people sit in the work. At ENAD Shymeca, the Virtual Lab implementation began with a discovery phase in which KABAS engineers worked with the client's R&D team on workflows and data structures, followed by migration of historical formulation data and training of R&D staff. Capturing knowledge before experienced researchers retired was one of the client's own concerns. At ADFC, decades of formulation experience were spread across formats and locations, and structuring them into a reusable repository was part of the project.
Two lessons carry over to a data team. The client's domain experts are part of the team, not an audience. And the structured data and documentation are an asset that should stay with the client.
Where should you start?
Name the sponsor, the problem owner and the first problem before you hire. Our consulting team runs readiness assessments and use case identification, and our data science team can cover data preparation, model development and MLOps while your own team builds those skills. Contact us to discuss your situation.
Sources
- Ryseff, De Bruhl and Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND, 2024
- Lawrence et al., Machine learning for industrial sensing and control: A survey and practical perspective, Control Engineering Practice, 2024 (arXiv 2401.13836)
- Microsoft, Establish an AI Center of Excellence, Cloud Adoption Framework
- Microsoft, Plan for AI adoption, Cloud Adoption Framework
- Google Cloud, MLOps: Continuous delivery and automation pipelines in machine learning
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015
- InfoWorld, For data scientists, drudgery is still job one (Anaconda 2020 State of Data Science survey)
- CISA, ASD's ACSC et al., Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- World Economic Forum, The Future of Jobs Report 2025
- Radio Algérie, Baddari préside la cérémonie de sortie de deux promotions de l'ENSIA et de l'ENSM, June 2026
- Google for Developers, Rules of Machine Learning
- Paleyes, Urma and Lawrence, Challenges in Deploying Machine Learning: a Survey of Case Studies, ACM Computing Surveys, 2022 (arXiv 2011.09926)
Frequently asked questions
How many people does a first industrial data team need?
A small core is enough to start: an engineer who owns the plant problem, a data engineer and a data scientist, with part-time help from OT and IT. Add machine learning engineers and more data scientists once a first model is in daily use and there is more work than the core team can handle.
Should we hire a data scientist or a data engineer first?
In most plants the first months are spent getting data out of the historian, the laboratory system and the maintenance system, and making it consistent. RAND's interviewees described data engineering as most of the work and noted that AI products need more than a data science team. If you can only hire one person, make sure someone with data engineering skills is on the team from the start.
Can our process engineers become data scientists?
Some can, and they bring the plant knowledge that data scientists usually lack. Training existing staff, hiring and working with partners are complementary routes. Engineers who move into data work still need time protected from operational duties, or the training does not turn into projects.
Should the data team report to IT or to operations?
Neither works well alone. IT brings infrastructure and security, and operations brings the problems and the users. Microsoft's guidance for early-stage AI adoption favours a centralised team that builds on existing teams. In a plant, that means a small central team reporting to a manager with authority over both, whose members spend regular time on site.
How long before a new data team delivers results?
Plan in months, not weeks. RAND found that leaders often expect AI projects to take weeks instead of months, and recommends committing each team to one problem for at least a year. A first model can be piloted within that year if the data is accessible, but the team's lasting value is the data pipelines, documentation and skills it builds.



