Implementing AI in business: why projects stall between pilot and production

From pilot to production: where AI projects get stuck

A pilot proves that a model can do a task. Production proves that the company can rely on it: inside the tools the team already uses, on the company’s own data, with results that someone can check before acting on them, at a cost that still makes sense when the task runs every day.

Between the two lies work that has little to do with the model itself. When it is skipped, the pilot stays a pilot – used by a few enthusiasts, trusted by nobody, and quietly abandoned when the next model version changes its behaviour.

Six gaps around the model

AI projects usually stall when one of these questions has no clear answer:

Typical mistakes when implementing AI

The same patterns appear again and again:

• Starting with a tool instead of a task, and then looking for a problem it might solve.
• Testing on clean sample data, then meeting the real, messy data in production.
• Treating a fluent answer as a correct one, without sources or checks.
• Letting an agent act without clear limits – or blocking it with so many approvals that nobody uses it.
• Sending every step to the largest model and discovering the bill after go-live.
• Handing the system over with no one responsible for it once models and requirements change.

Step 1: an honest assessment

Before the first project, find out which tasks bring real value with AI, which data exists for them, and what regulation requires. This includes the team: since 2 February 2025, Article 4 of the EU AI Act requires organisations that use AI systems to take measures so that their staff have sufficient AI literacy. Frameworks such as the NIST AI Risk Management Framework or the ISO/IEC 42001 standard for AI management systems help to structure this assessment. The result is a roadmap with one clearly scoped first project – the job of AI consulting.

Step 2: a clean implementation

AI has to be connected to the systems where the work actually happens. Results need a source, critical steps need human approval, and before go-live the system is tested with the people who will use it. Data protection is settled here, not afterwards: where data is processed, whether a provider may use it for training, and whether a data processing agreement under Article 28 GDPR is in place. This is AI engineering.

Step 3: reliable operation

Models change, costs grow with use, new requirements arrive. Without ongoing care, even a good solution loses value quickly. Reliable operation means watching quality and cost together, testing new model versions before they go live, and having someone to call when something breaks. That is what managed AI is for.

Step 4, sometimes: a product of your own

When no existing solution fits the task, an idea can become an AI product of its own – through discovery, a prototype and an MVP to something people actually use.

What a good first AI project looks like

A good first project is small enough to finish and real enough to matter. It covers one task that the team does often, uses data the company already has, and produces a result a person can check in minutes. It runs inside an existing tool rather than a new one, so nobody has to change how they work to try it.

It also has a clear end: before it starts, the team agrees what “good enough for daily use” means, which test cases it must pass, and who decides whether it goes live. If it passes, the same pattern – task, data, checks, approval, cost, owner – is reused for the next task. If it does not, the company has learned what is missing at low cost.

Risks and limits

Not every task should be automated. Where errors are expensive, the data is thin or the rules are unclear, AI should prepare and a person should decide. And no assessment replaces legal advice on specific obligations under the GDPR or the EU AI Act.

NodeByNode supports companies in all four steps – from the first question to daily operation.