# How do you implement AI in a business?

> Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, founder of AINORA, MB – the company behind the Ainora and Impetora brands – and a graduate of Vilnius University.

Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, working with mid-market companies and scale-ups across the UK, EU and US. He builds and operates the production AI systems he advises on, and takes no commission, referral fee or revenue share from any vendor.

**Eight steps, and the order matters more than any of them individually. Most companies start at step five, building something, and discover steps one to four afterwards. That is the single most common reason implementations fail rather than any technical problem.**

Canonical: https://justasbutkus.com/how-to-implement-ai-in-my-business/
Last updated: 2026-08-02

---

## The eight steps, in order

1. **Pick one process and write it down** – In one sentence, with a number. Volume and time per occurrence. If you cannot, measure it before doing anything else.
2. **Check the data behind it** – Not whether it exists. Whether the same field means the same thing everywhere it appears. This is where most projects quietly die.
3. **Name who accepts the risk** – One person who can say yes to it running in front of customers. Do this before building, because you will not find them afterwards.
4. **Define what working means, numerically** – Agreed in writing, in advance. Without it, evaluation becomes an argument nobody can win.
5. **Build the smallest version that could work** – Not the full vision. The narrowest slice that proves the chain end to end.
6. **Run it alongside the humans first** – Same inputs, output compared, nothing acted on. This is where you find out what it actually does, cheaply.
7. **Put it live with a rollback you have tested** – Tested, not documented. An untested rollback is a hope.
8. **Decide who owns it on an ordinary Tuesday** – Systems degrade. Someone must notice. Without an owner, quality decays until a customer finds it.

## Which step do companies skip?

Steps two and six, almost always.

Step two is skipped because everyone assumes their own data is fine. It usually is not, in a specific way: the same concept recorded differently across systems that were never designed to agree. Any system learns that inconsistency faithfully.

Step six is skipped because it feels like delay. Running the system alongside people, with nothing acted on, is the cheapest information you will ever buy about how it behaves on real inputs. Skipping it means finding out in front of a customer.

## How long should this take?

**Realistic timings for a first implementation in a mid-sized company**

| Step | Typical time | Common failure |
| --- | --- | --- |
| Choose and write down the process | Days | Choosing the interesting one |
| Check the data | One to two weeks | Assuming it is fine |
| Name the risk owner | Hours, if anyone will | Nobody will, and it is not noticed |
| Define success numerically | Hours | Left vague, so it can never fail or succeed |
| Build the smallest version | Two to six weeks | Building the full vision instead |
| Run alongside humans | Two to four weeks | Skipped to save time |
| Go live with rollback | Days | Rollback documented but never tested |
| Ongoing ownership | Forever | Nobody owns it, quality decays unnoticed |

A first implementation reaching production within a quarter is reasonable. Substantially longer usually means the constraint was selection, data or ownership rather than engineering.

## What does it cost to keep running?

The number most proposals leave out, and over three years often the larger one.

- **Usage.** It costs something every time it runs, and volume grows.
- **Someone owning it.** Not full time, but not zero either.
- **Re-checking when your process changes.** Your business changes, and the system does not notice on its own.
- **Fixing what drifts.** Quality moves over time, and someone has to be watching for it.

Ask any supplier for this figure explicitly. Most quote a build cost and go quiet on running cost.

## Frequently asked questions

### How do I implement AI in my business?

Eight steps in order: choose and write down one process with a number, check the data behind it, name who accepts go-live risk, define success numerically, build the smallest working version, run it alongside people, go live with a tested rollback, and assign ongoing ownership.

### What is the most commonly skipped step?

Checking the data, and running the system alongside people before it acts on anything. Both feel like delay and both are the cheapest information you will get about whether it works.

### How long does an AI implementation take?

A first one reaching production within a quarter is reasonable for a mid-sized company. If it is taking much longer, the constraint is usually selection, data or ownership rather than engineering effort.

### What does an AI system cost to run after it is built?

Usage that grows with volume, someone owning it, re-checking when your process changes, and fixing quality drift. Over three years this is often larger than the build cost, and most proposals omit it.

### Should we build the full system or start small?

Start with the narrowest slice that proves the chain end to end. Building the full vision first means learning to operate AI and solving the hard part of the use case at the same time, which is what produces stalled projects.

## Related

- [Where to start](/ai-for-my-company/) – Before the sequence, how to choose.
- [If one already failed](/answers/our-ai-pilot-failed/) – The post-mortem before trying again.
- [What to automate](/ai-automation-for-my-company/) – Which processes are worth it.

## About the author

**Justas Butkus** – a fractional AI officer based in Vilnius, Lithuania, founder of AINORA, MB – the company behind the Ainora and Impetora brands – and a graduate of Vilnius University.

## If you are at step one

The first four steps need no supplier and no budget. If they come out convincingly, that is worth a conversation.

Contact: justas@ainora.lt · [LinkedIn](https://www.linkedin.com/in/justas-butkus/)