Insights
Talent Acquisition

AI Experts: Which Roles Take Your Pilot Into Production

Johannes Liebnau
October 30, 2026

Key Points at a Glance

  • An AI pilot rarely fails because of the model. Usually a position is missing: someone who owns the use case, someone who has the data under control, or someone who keeps the system running in production.
  • If you want to hire AI experts, do not look for all-rounders. Look for profiles that fit the stage your project is in right now.
  • We map these profiles to the four steps of the whitepaper "Return on AI": choose, measure, redesign, collect.
  • You will learn when a permanent hire, a freelancer or an interim specialist is the right fit, and how to spot real hands-on experience in an interview.
  • This article is for developers, engineering leads and tech decision makers who need to move a pilot into regular operation.

Quick navigation

Artificial Intelligence in Your Company: Why One Job Is Not Enough

The job ad "AI expert (m/f/d)" exists in a thousand variations. Behind it is usually the hope that a single person covers data preparation, model development, integration, operations and coordination with the business side. In a proof of concept that works. Afterwards, rarely.

The title is also vague. Data scientists train models, ML engineers make them production-ready, and architects design the target picture for the whole system. Each of these positions requires different experience, even if all of them work with artificial intelligence. Lump these profiles together and you get applications that look right on paper and fail in daily work.

Then there is the market. Demand for AI experts is high, and a job ad alone reaches very few of them.

A helpful framework comes from the whitepaper "Return on AI", which CTG Consulting published. Its core thesis: a pilot only pays off once the workflow around it changes. For AI talent acquisition this means you need people who can build, and someone who takes responsibility for the value stream.

The Four Steps and the Profiles Behind Them

An AI project moves through four steps: choose, measure, redesign and collect. Each step places different demands on your team. There is no fixed mapping, but there is a practical rule of thumb for your workforce planning.

Step 1, Choose: Program Lead and Ownership

It starts with the decision for a single workflow with a clear outcome lever, usable data and an accountable person, instead of pursuing ten use cases in parallel. This is where you need two roles: the value stream owner, who runs a workflow end to end with authority across department boundaries, and the AI program lead, who steers the portfolio of use cases by business value and ends pilots without a business case. Technically neither needs the deepest expertise, but both must understand what an ML system needs to work.

Step 2, Measure: Data Ownership and Analysis

Before building, you measure how the workflow runs today, separating working time from waiting time. This does not take a data science department, but event data: timestamps per step from ERP or CRM. Data engineers and data analysts make them accessible, and clear data ownership makes sure they are right. Skimp here and you build on an unreliable basis.

Step 3, Redesign: Architecture, ML Engineering and Integration

Now you build: an AI engineer or ML engineer builds the models and pipelines, an architect decides how the system fits into the existing IT landscape, and software engineers handle interfaces and integration. Order matters. Architecture decisions in an AI project (build or buy, model choice, data flow, security) come early and are expensive to change later. If you can only fill one senior position, put it here.

Step 4, Collect: MLOps and Operations

Finally, the system has to run permanently, be monitored and be tuned, because the business case counts in hit rate and cash, not in hours saved. This is where MLOps comes in: deployment, monitoring, versioning of models and data, quality assurance in production. Without this role, the system slowly drifts after go-live without anyone noticing. Cloud and DevOps skills belong here too.

Four steps, four roles

Who you need in which step to turn the AI pilot into a result.

  1. Step 1: ChooseOwnership
    Value stream owner and AI program leadNo owner, no value stream.
  2. Step 2: MeasureData
    Data engineers and data analystsBaseline from event data.
  3. Step 3: RedesignAI architecture and ML engineering
    Architect, ML engineer, software engineersDecided early, costly to change later.
  4. Step 4: CollectMLOps
    Deployment, monitoring and quality assuranceWithout operations, the system drifts.

Maturity of AI use

  1. Level 1: Fragmented
  2. Level 2: Selectively optimized
  3. Level 3: Process-oriented
    The return starts hereJump from level 2 to 3
  4. Level 4: Value-stream-oriented

Staffing

  • Permanent hireNeeded long term: MLOps, data ownership, architecture.
  • FreelanceClearly scoped build phases: pipeline, prototype, integration.
  • InterimLeadership right now, role not yet permanent: program lead, owner for a limited time.

Source: whitepaper “Return on AI” by CTG Consulting

What This Looks Like in a Team: An Example From Mechanical Engineering

A machine builder wants to send offers faster. The team starts with the workflow from inquiry to offer, because the head of sales wants to own it and the data sits in the ERP. The first effect often comes from a decision by the accountable person, such as raising the approval limit, which needs no AI. After that, AI solutions read inquiries from emails, propose calculations and let a person review and approve.

Transferred to your team, the staffing looks like this: someone from sales takes ownership and decides what may change. A data analyst measures the starting point. An ML engineer and a software engineer build the automation and connect it to existing systems. After go-live, someone keeps the whole thing stable. None of these steps needs research depth, but each needs someone who is accountable for it.

Which Roles Fit Which Maturity Level

Which roles you look for depends on how far your company has already come with AI. At the start, individuals or departments use AI on their own initiative; at the end, one continuous workflow drives the result. In between lies the jump that makes the difference: from selective optimization to a redesigned process.

At the start, a sponsor from the executive team and an owner per use case are often enough. Technically, developers and data engineers who connect tools and make data accessible are usually sufficient.

When an entire process is redesigned, the requirements change. Now you need someone who owns the workflow from start to finish (the value stream owner) and an AI program lead, plus process competence: the ability to map a workflow, redesign it and measure it with metrics. These are rarely pure engineering roles, and in many companies they do not exist internally yet.

Once AI agents take over defined tasks across process boundaries, data ownership and the technical AI roles join in: AI architecture, machine learning and LLM engineering. Clean operations with MLOps usually belong to it, and someone has to define what the agents may do on their own.

The technical profiles are usually the hardest to find. For all roles there are three routes that can be combined: develop internally, hire externally, or bridge the role during the transformation with an interim leader.

Generalist or Specialist? How Much Depth Your Project Needs

The question comes up in almost every hiring process. A generalist covers many topics solidly and is valuable in the pilot because they can close gaps. An AI specialist brings depth, for example in NLP, computer vision or LLM applications, and pays off once requirements are clearly defined.

A rule of thumb: the earlier the phase, the more likely you need a generalist. The clearer the use case and the higher the quality bar in production, the more likely you need an AI specialist.

Agentic AI also shifts the profile. It is less about classic model development and more about orchestration, tool integration and safeguarding.

Technologies in Daily Work: What Recruiting and Engineering Should Settle Together

Before you post a profile, look at the technologies the team will work with. Not every use case needs its own model. Often it is enough to connect existing models through an interface and link them to company data. That changes the profile: whoever integrates AI systems needs solid software engineering and experience with data processing. Whoever trains their own models needs deeper knowledge of machine learning and statistics.

Which AI tools are allowed in house should be settled early too, as should data protection. Anyone planning the implementation of AI projects should name technologies and business processes first and the position afterwards. That keeps the AI transformation part of the company strategy instead of a one-off project.

An example from practice: a chatbot for customer service has different requirements than a demand forecast. Both are AI solutions, but one lives on language processing and integration, the other on clean time series data and monitoring. The profile you look for differs accordingly.

What Specialists Should Bring

There is no single path into AI. A degree in computer science, mathematics or statistics often comes first, but career changers from software development or data analysis frequently bring relevant experience too. Add Python, frameworks such as TensorFlow or PyTorch, and hands-on experience with machine learning. Day to day, they prepare large amounts of data, check model accuracy and take care to avoid bias.

Because the field moves fast, continuous learning comes with the job. Many companies therefore train their own AI experts, for example through courses and workshops. That is a good complement but does not replace practice: if you want to know how a model goes into production, you have to have run one yourself.

Permanent, Freelance or Interim

Which contract type fits depends on how long you need the capability and how much knowledge should stay in house.

Permanent hires fit everything needed long term that builds context: MLOps, data ownership, architecture as a guardrail.

Freelancers fit clearly scoped build phases, such as setting up a pipeline, a prototype or an integration. You get specialist knowledge quickly and can scale back afterwards.

Interim specialists fit when a leadership or program position is needed immediately but should not yet be filled permanently, for example the AI program lead or a value stream owner for a limited time, until the role is filled internally.

The decision is rarely black and white. A common route: interim or freelance to start, building permanent hires for the core positions in parallel.

How to Choose AI Experts: Steps and Interview Guide

A resume full of AI buzzwords says little. These steps help you structure the selection:

  1. Clarify the need: Which task should the person take on, and which of the four steps does it belong to? Write it down in two sentences before you draft a job ad.
  2. Prioritize requirements: Separate must-haves from nice-to-haves. Not every position needs research depth, and production experience often counts more.
  3. Choose sources: Networks, communities and active sourcing reach people who do not show up on job boards. For scarce profiles, a specialized recruiting partner helps.
  4. Run a working session: A real problem from your stack says more than a question catalog. Give insight into your systems and watch how the person asks questions.
  5. Back up the decision: Bring in someone from the business so that communication and business understanding are assessed too.

In the working session, look out for these points:

  • Production experience: Was a model actually put into operation and maintained, or did it stay in a notebook?
  • Data understanding: Does the person know data quality, cleaning and bias, and handle them cleanly?
  • Communication: Can they explain technical matters to the business? Good AI experts are often the interface between business and technology.
  • Willingness to learn: The field moves fast. Ask what the person learned most recently and how.

Plan for Technical Debt Early

Many pilots emerge under time pressure and grow into systems nobody wants to touch. That is classic technical debt, with added risks from data and models. Plan time for tests, documentation and refactoring from the start, and give the person who looks after the production system a mandate to say no.

For recruiting this means: do not only look for people who build a prototype fast, but also for people who keep systems maintainable in the long run. Ask how it was tested and how a model was monitored after go-live.

Frequently Asked Questions

Can I build AI skills in house?

Yes, with training, mentoring and a first project that actually goes into production. Expect it to take time and plan for at least one experienced person who passes on knowledge.

What is the difference between an ML engineer and a data scientist?

An AI engineer, often called an ML engineer, builds and runs systems, with a focus on production, integration and scaling. Data scientists focus more on analysis and model development. In practice the two profiles overlap.

How long does it take to fill an AI position?

There is no blanket answer. For rare profiles the search usually takes longer than for classic developer roles, especially when requirements are unclear or senior production experience is required. Plan a buffer and, if needed, bridge the time with an interim or freelance solution so the pilot does not stand still.

Which position is missing most often when a pilot does not reach production?

Often it is the tasks without glamour: ownership of the value stream, data ownership and MLOps. The model itself is rarely the problem.

Your Next Step

Tell us briefly which of the four steps your AI project is stuck on and what stack sits behind it. You get an honest technical assessment of which role to fill first and whether a permanent hire, freelancer or interim fits. The full framework with all four steps is in the whitepaper "Return on AI".

Which role is missing in your AI project?

We classify the profiles technically and search with you for the right experts for your AI pilot. On equal footing and without obligation.

Key Takeaways

  • Profiles instead of a wish list: Do not look for AI experts as a catch-all, but for the capabilities your project needs right now.
  • Four steps, four focus areas: Choosing needs ownership, measuring needs data, redesigning needs architecture and ML engineering, collecting needs MLOps.
  • No owner, no value stream: The program lead decides whether a pilot becomes a result.
  • Contract type by duration: Permanent for the lasting, freelance for build phases, interim for short-notice leadership.
  • Plan operations early: MLOps and technical debt belong in the pilot phase.
  • What challenges are you currently facing?

    Take short survey

    ‍