Guide · Strategy & First Project

What an AI Pilot in Companies

The most common source of failure in AI pilots at companies isn't the technology. It's the assumption that a pilot is primarily something technical. In reality, the quality of scope, ownership and hand-over decides whether the pilot becomes a productive system — or a completed demo project sitting in the showroom.

This guide sums up what we've learned across DACH company projects — compact, without hype, with a clear recommendation for the first weeks.

Short answer

An AI pilot in a company needs five things: a clearly bounded use case with one measurable key metric, an operational owner who takes over the solution afterward, real data flow instead of a showcase dataset, six to ten weeks of implementation time, and a written, agreed hand-over into regular operations. Leave out any one of these five elements and you're not building a pilot — you're building a demo.

1. Scope: better small and sharp than big and vague

The typical misstep: "We want to use artificial intelligence in our sales." That's not a pilot definition — it's a strategic wish. A workable pilot scope answers four specific questions: Which process is being automated? What input data? What measurable result? Which metric is measured before and after the pilot?

Example of a sharp scope: "Automated capture and classification of incoming delivery notes, hand-over to the ERP. Comparison of processing time and error rate before and after the pilot, measured over 30 days of parallel operation." That's concrete, measurable, and can go live in four to six weeks.

The opposite: "AI-driven optimization of supplier communication." That sounds strategic — and is precisely for that reason not a pilot scope. It's a program that can end up anywhere within twelve months, without anyone noticing along the way that, without a focused first step, it would never become productive.

2. Ownership: one person — named before the project starts

A pilot without a named operational owner isn't a pilot. It's a project in a vacuum. The question isn't "Who's leading the pilot?" — the question is "Who takes over the solution after the pilot ends and keeps referring back to the results regularly afterward?"

This person must be named before the project starts, by name, with a concrete time allocation, and with enough authority to make day-to-day decisions. If this person still needs to be found, the pilot will fail — regardless of how good the technology is.

Practical observation: pilots where operational ownership only gets clarified after go-live are highly likely to end up shelved. No one feels responsible, the system loses quality without upkeep, and after three months nobody uses it anymore.

3. Data flow: a real stream, not a curated dataset

The biggest temptation in an AI pilot is to start with a curated training or sample dataset. "We've got a hundred clean documents here, we'll show the feature on those." That's not a pilot — that's a demo. Demos win people over in workshops, but they don't solve an operational problem.

A real pilot works on the company's actual live data stream: the documents coming in today, with all their edge cases, bad scans, unclear suppliers, unusual formats. Exactly these edge cases determine the quality of the eventual solution. Leave them out of the pilot, and you're just pushing them to the day after go-live — where they get far more expensive.

Practical recommendation: better to work in the pilot with lower volume but the real stream, than with high volume from a curated dataset.

4. Timeframe: six to ten weeks of implementation, not six months

An AI pilot that needs more than ten weeks of implementation time is no longer a pilot — it's a disguised large-scale project. Typical implementation time in companies is four to six weeks of build and test plus two to four weeks of parallel operation. Add another two to four weeks of preparation (scope sharpening, data source connection, test set). If it takes longer, the scope was chosen too broad.

There's a practical reason for this: within six to ten weeks of implementation, a company's business situation doesn't shift so much that the initial assumptions become invalid. Over six months, however, processes, people and tools change — and by go-live, the pilot lands in a world that no longer matches its original assumptions.

The solution for larger initiatives isn't a longer pilot, but several small, sequential pilots with clear hand-over points.

5. Hand-over: written, defined before the project starts

The fifth element is also the most commonly forgotten one: What happens after the pilot? Who operates the solution? Who responds to quality drift? Who decides on model updates? Who keeps the interfaces to ERP, CRM and DMS stable?

These questions must be answered before the pilot starts — in writing, with concrete names. Not "the IT department" or "the vendor" — a specific person with a specific mandate. There are three options: the internal team takes over fully, the external partner takes over under a managed-operations contract, or a hybrid model (internal ownership, external backup readiness).

Anyone who only clarifies the hand-over question after the pilot ends loses the moment the external partner leaves. The solution keeps running — but nobody maintains it, nobody responds to new edge cases, nobody checks quality. Three to six months later, the pilot is shut down.

When an AI pilot makes sense — and when it doesn't

Makes sense when
  • • A specific process recurs regularly and ties up manual time
  • • The process has clear input data and a clear result
  • • A measurable key metric exists or can be defined
  • • An operationally responsible person is named
  • • Management stands behind the pilot through to go-live
Doesn't make sense when
  • • The process isn't yet clearly documented or run consistently
  • • No one is willing or able to take operational ownership
  • • The pilot primarily serves a board pitch
  • • The goal is to showcase "the technology AI" without concrete business value
  • • Data flow access isn't guaranteed for organizational reasons

Frequently asked questions about AI pilots

How long should an AI pilot in a company take?
Six to ten weeks of implementation time, plus another two to four weeks of preparation. If it takes longer, the scope was chosen too broad. If someone wants to finish faster, they're building a showcase without real data flow — that's not a pilot, it's a demo.
Which process is suitable as a first AI project?
A recurring process with a clear input, a clear result, a measurable key metric, and an operational owner who wants to take over the solution afterward. Classic examples: document recognition, inquiry classification, standard responses in customer service. Poor candidates: highly strategic decisions, processes with high variability, processes without a clear owner.
What does an AI pilot in a company realistically cost?
For a company, the effort for a clean first pilot — consulting, implementation, integration into an existing system, and one month of accompanied operation — typically falls in the low to mid five-figure euro range. Anyone advertising a flat "from €5,000" is building a showcase. Anyone quoting a flat six figures is building too much at once.
When has an AI pilot failed?
When, after the project ends, three conditions are not met: (1) the pilot runs in real day-to-day business, not in a showroom; (2) a specific person carries operational responsibility; (3) an agreed metric is measured on an ongoing basis. If any one of these three is missing, the pilot isn't a pilot — it's a completed demo project.
Should we start internally or bring in external support?
Companies that already have operational AI experience in-house can certainly start on their own. Those who haven't completed a pilot yet benefit from external support: for defining scope, selecting the first solution, and a clean hand-over into regular operations. External partners should not take over the wheel, though — they should make the path clear so the internal team can run the solution long-term.

Related topics

Pillar 1 · Strategy

AI consulting →

Strategic AI consulting with a roadmap and ROI estimate. The step that precedes every productive pilot.
Free initial call

AI potential analysis →

A structured first assessment: which use cases actually work as a first pilot? Within 48 hours, no upfront cost.
Guide · Decision

AI agent or classic workflow? →

The decision question that shapes every pilot: when is an agentic approach the right choice?
Pillar 5 · Support

Managed AI operations →

The honest answer to the hand-over problem: ongoing operations with monitoring and drift checks.

Want a concrete pilot proposal for your company?

In a free initial call, we identify the one to three processes that are realistically suited for a pilot today — and state clearly which ones you should hold off on.

Request a free initial call