# Implementing AI in your company, in three verifiable steps > How AI implementation works in a company, step by step: paid discovery, a pilot with written acceptance criteria, then controlled rollout. URL: https://thenichesociety.ro/en/ai-engineering/ai-implementation Implementing AI in your company almost always means following a clear process: discovery, a pilot with written criteria, then rollout — not just picking a good model. **We structure every project this way**, whether it's training, an agent, or an integration with your existing ERP. ## You can't implement what you haven't measured The first step isn't choosing an AI model, it's understanding the company's real processes and data. A day of discovery with the leadership team produces an opportunity map with estimated ROI for each, not a generic list of "what AI can do." Without this step, the risk is building something technically impressive that doesn't solve the company's real problem — or worse, something that needs data the company doesn't have in structured form. AI discovery for companies isn't a generic list of ideas — it's a working day with the leadership team that produces an opportunity map with real priorities. ## Acceptance criteria get written before, not after The most common mistake in an AI pilot is judging success "by feel" after building something. We write the acceptance criteria before implementation starts: what the system needs to do, with what accuracy, in how much time. If the criteria are not met at the end of the pilot, the decision is clear — you adjust or stop, you do not continue on the strength of a good impression from a demo. How this process actually played out at FOX — from the first training to an SAP digitalization discovery — is documented step by step in the case study. [how this process actually played out at FOX](https://thenichesociety.ro/en/case-studies/fox). ## Go/no-go, based on data After the pilot, the decision to scale isn't made out of enthusiasm, but by comparing the real results against the criteria written at the start. A pilot that doesn't meet its criteria isn't a project failure — it's the information that tells you what needs to change before investing more. This is the step missing from most failed AI projects: they jump straight from „the demo worked” to „let’s roll it out everywhere”. A go/no-go decision on criteria written down in advance is what separates real AI implementation in a company from an experiment with no clear ending — if the pilot does not pass the criteria, the project stops there, it does not get extended out of inertia. When choosing tools for the pilot, we explain [why the AI model matters less than you think](https://thenichesociety.ro/en/blog-modelul-ai-conteaza-mai-putin) compared with how well what the system can do is controlled. ## Scaling up requires training, not just code An AI system that works technically but that the team doesn't know how to use produces no value. The rollout includes training for the affected teams, documentation, and usually a maintenance retainer — models and integrations change, the system needs upkeep. This is where governance comes in too: who's allowed to see what data, what gets logged, what needs manual approval. The same process applies to [implementing AI agents](https://thenichesociety.ro/en/ai-engineering/ai-agents): one measurable flow first, then expansion. ## AI prepares, the human decides In every implementation we've done, we've kept it explicitly clear what stays a human decision: the signature on a filing, pressing a submit button, judgment on an unclear case. The argument is as much commercial as it is prudent — "AI does everything on its own" is a hard thing to stand behind when legal liability stays with the company. The "copilot, not autopilot" model was validated directly in an accounting automation project: the system reads, calculates and flags; the human presses the button. Expansion goes faster when the team goes through an [AI training for employees](https://thenichesociety.ro/en/ai-engineering/ai-training) built on its real cases. ## From training to a complete platform, at the same client We've seen the full path play out at a single industrial company: team training, then a small platform (ticket redemption), then a discovery for production digitalization on SAP, then a proposal for a full AI platform. Each step was built on the trust earned in the one before. It is the kind of progression we recommend: not „everything at once”, but verifiable steps that build up. At FOX, the AI discovery for the company started with training, not with a complex system — only once the team trusted the tools did we move to a real automated process. ## A fixed price, after a short discovery call. - process and data audit - opportunity map with estimated ROI - recommendations report - written acceptance criteria - 1-2 cases in production - handover + documentation - multi-department rollout - team training - monthly retainer Every project has a different context, workflows and infrastructure, so the price is set after a short, paid discovery and does not change along the way. ## Frequently asked questions ### How is AI implemented in a company? In three steps: discovery (one day, an opportunity map), a pilot on 1-2 cases with acceptance criteria written beforehand, then rollout with training and maintenance. Skipping discovery or not having written criteria are the most common causes of failure. ### What does AI implementation actually mean? It's the process of connecting an AI model or agent to a company's real data and processes, with rules, permissions and an audit trail — not just a few employees using a chatbot ad hoc. It includes both the technical side and team training and internal usage policy. ### How do I implement AI in my own business, without an in-house technical team? The most realistic option is an external partner who does the discovery, builds the pilot with clear criteria, and hands over a documented system, with training included. You don't need an in-house AI department from day one — you need a first use case taken all the way through and measured. ### How long does an AI implementation take, from discovery to production? Discovery usually takes a day; a pilot on 1-2 use cases takes between two and six weeks, depending on the complexity of the data and the integrations required. Full rollout varies a lot and depends on how many departments are involved. ### What happens if the AI pilot doesn't work as it should? If the acceptance criteria written at the start aren't met, you adjust the approach or stop the project before investing more — the decision is made on measured data, not on a good impression from the demo. That's why the criteria get written before you start, not at the end. ### Who makes the final call in an AI system implemented at a company? Decisions with legal, financial or contractual impact stay with people — the system prepares, checks and flags, but doesn't press the final button on its own. This separation is set explicitly during discovery, not added on later. ### What CAEN code applies to AI implementation / custom software development? Custom software development (including AI implementation as a service) generally falls under CAEN code 6201 — Custom software development activities. Always confirm the exact classification with an accountant, based on your actual invoicing structure. ### Roughly how long does an AI discovery for a company take? Usually one working day with the leadership team, followed by a written opportunity map. The pilot that follows, if discovery confirms a clear direction, usually takes 2–6 weeks on 1–2 use cases. ### Let's see what can be automated in your business. A free 30-minute session: we'll tell you what can be automated, how long it takes and what it costs, with a fixed price after discovery. We reply the same business day.