Which department should you start with?

The one where a written task comes back most often: customer support, sales administration, marketing, catalogue management. The right starting point is a task someone does every week reluctantly.

Do you need to train the whole team before starting?

No. You need one lead person per function, one use case and simple rules about confidential data. Broad training comes after the first result.

How long before a result?

A well-chosen first use case gives a measurable result within a few weeks. Projects that run for months without delivering usually started too big.

GuidesAI in business

Bringing AI into a small business: where to start

By the Banzai Labs teamUpdated on November 11, 20262 min read

The starting mistake: buying a tool

Many companies start with licences. Six months later, a few people use them to rephrase emails, and nothing has changed in the business.

AI is brought in through the work, not the tool. The right question is: which task, done by whom, how often, with which data?

Finding the first use case

A good first use case ticks four boxes:

1It comes back oftenEvery day or every week.
2It relies on text or dataYou already own.
3Its result is quick to checkSomeone in the role can tell whether it is right.
4A mistake is not seriousNothing irreversible at the start.

Common examples:

sorting and pre-qualifying incoming requests;
producing a weekly report from several tools;
answering internal questions from the documentation;
writing and translating product pages, reviewed by a human.

Our guide to the AI audit gives the method for drawing up this list.

Preparing the data

An agent is only as good as what it reads. Before building:

identify where the reliable data is, and who keeps it up to date;
clean out what is wrong or outdated, because the AI would repeat it with confidence;
decide what must never leave: customer data, contracts, salaries.

The three levels of integration

Level	What it is	Effort
General assistant	A chat tool, used by hand	Low, limited gain
Connected assistant	The same, linked to your documents and tools	Medium, clear gain
Autonomous agent	It runs alone, on a schedule or on an event	Higher, lasting gain

You move up a level once the previous one has proved itself. The step to the second level often goes through an MCP server or a RAG.

Measuring

Without measurement, the project stays an impression. Before launching, write down:

the time spent on the task today;
the current error rate;
the expected monthly running cost.

After a month, compare. If the gain is not clear, stop or change use case.

The rules to set right away

What may be given to a model, and what is forbidden.
Who reviews what the AI produces before it leaves the company.
Who is responsible for each agent.

The legal framework is summarised in AI, GDPR and the AI Act.

Getting help

The first project goes faster with someone who has already put some into production. See AI digital agency: how to choose one.

Frequently asked questions

Which department should you start with?: The one where a written task comes back most often: customer support, sales administration, marketing, catalogue management. The right starting point is a task someone does every week reluctantly.
Do you need to train the whole team before starting?: No. You need one lead person per function, one use case and simple rules about confidential data. Broad training comes after the first result.
How long before a result?: A well-chosen first use case gives a measurable result within a few weeks. Projects that run for months without delivering usually started too big.