Skip to content

AI

Implementing AI in a company, step by step – what our own projects taught us

Sweet Lava Team Published 3 min read

Key answer: The best way to implement AI in a company is to start with one process, not an "AI strategy for the whole organisation". The sequence that works – audit your data and processes, pick a pilot with a measurable goal, implement in weeks, measure the result, and only then scale. A pilot implementation usually costs €3,500–10,000; larger integrations €10,000–35,000. The hardest part is not the technology but organising the data and the process.

Table of contents
  1. 01The short answer
  2. 02Step 1: audit the data and processes
  3. 03Step 2: a pilot with a measurable goal
  4. 04Step 3: measure and decide
  5. 05Step 4: integrations and scaling
  6. 06Step 5: enable the team
  7. 07How much does an AI implementation cost?
  8. 08When AI isn't worth it

The short answer

Implementing AI in a company starts with one process, not a grand strategy. The sequence that works: data audit, a pilot with a measurable goal, measurement, and only then scaling. We work with AI every day, including in our own tools, and this article describes the process that has survived production at our studio – with prices, and with the situations where we honestly advise against the whole thing.

Step 1: audit the data and processes

Before anyone says "model", two questions need answers: which processes consume the most repetitive work, and what data the company actually has. We run this as a workshop: a review of processes, data sources and systems (CRM, ERP, helpdesk, spreadsheets). The result is a list of use cases ranked by potential return, not by wow factor.

This is where the most common surprise appears: the data is often scattered across inboxes and Excel files. That doesn't block the implementation, but it changes its first stage from "connect the model" to "organise the sources".

Step 2: a pilot with a measurable goal

We pick one process and define the success metric before the start: enquiry handling time, the number of manual data re-entries, the cost of preparing a proposal. The pilot must run in production, on real data, within weeks. A demo on test data proves nothing.

A good pilot has three traits: a narrow scope, a human checkpoint wherever an error costs money, and a clear written list of what the AI must not do on its own.

Step 3: measure and decide

After 2–4 weeks of the pilot we compare the numbers with the starting point. If the process shrank from hours to minutes, we scale. If the effect is cosmetic, we go back to the workshop list or close the project. Yes, sometimes we recommend stopping – it's cheaper to end a pilot than to maintain an implementation that doesn't earn its keep.

Step 4: integrations and scaling

A working pilot gets extended to further processes and connected to company systems through APIs. This is also where the rules are written: who owns the quality of AI answers, how errors are reported, when the model must hand over to a human. Without these rules even a good implementation loses the team's trust over time.

Step 5: enable the team

The best tool earns nothing if nobody uses it. That's why showing your team how to use what we set up is part of every implementation: a live session, concise instructions, good practices. It's not a separate training product – it's a precondition of the return on investment.

How much does an AI implementation cost?

A pilot – one process with a measurable result – usually costs €3,500–10,000. Larger integrations with company systems and databases typically land between €10,000–35,000. On top come running costs: model API fees (usually from tens to a few hundred euros monthly, depending on scale) and technical care. Always precede the quote with a workshop – buying an implementation without a ranked list of use cases is buying a pig in a poke.

When AI isn't worth it

We say this plainly because it saves everyone time. An AI implementation makes no sense when the process is rare (automating something that happens once a month pays back in a decade), when the data is too thin or chaotic and organising it would cost more than the manual work, and when the cost of an error is high and a human has to verify every result anyway. In those situations a simple automation, a better form or simply another pair of hands works better.

Wondering which process to start with? Tell us about your company and we'll work it out together at a workshop.

Related articles

Frequently asked questions

What is the hardest part of implementing AI in a company?

The data and the process, not the technology. The models are mature, but they need well-organised data and a clearly described process to plug into. Companies that skip the data audit end up with an impressive demo and zero business result.

How long does an AI implementation take?

A pilot – one process with a measurable goal – usually takes 4–8 weeks from workshop to a working solution. Extending to further processes and integrating with company systems takes additional weeks or months, depending on the number of integrations.

Can a small company implement AI too?

Yes, and often with a faster return than a corporation, because decisions are made on the spot. Good first steps are automating answers to repetitive enquiries, product descriptions or document summaries. A pilot in a small company lands near the lower end of the price range.

Have an idea in mind? Let's talk.

Tell us about your idea and we'll recommend the best route from concept to a working product.

Get in touch