AI guides · 3 of 6

Where to start with AI in your company: a plan for the first 30 days

A 30-day plan for bringing AI into a company: how to pick the first task, what to prepare before launch, why a silent run matters and what to decide on day 30.

22 November 20268 minMRKT.ERA

The first books printed after the Gutenberg Bible in the mid-fifteenth century worked hard to pass for manuscripts. Printers left blank space for capital letters that an illuminator would later paint by hand, used typefaces that copied the scribes' handwriting and printed the same texts that monasteries had been copying for centuries. Title pages, page numbers and the book as we know it came later, once people understood what the new press could do. Companies that start working with AI go through the same stage. The first impulse is to have the model do exactly what people did, in the same steps. The month this guide covers is there to get through that stage quickly: choose one task, test it honestly and by day 30 understand what AI actually changes in your work.

01Days 1–7: find one task

Ask department heads to write down what their people spend the most time on, week after week: summary reports, replies to customers, going through documents, moving data from one system to another. Next to each task, note the hours it takes per week and what a mistake costs.

From that list, pick one task that repeats every day, relies on data in your accounting system or CRM, and has a result you can check in a minute. We explained how to recognise a suitable task in [our guide on whether a business needs an AI agent](/en/blog/do-you-need-an-ai-agent/).

In the same week, give a few employees work subscriptions to a chat assistant such as ChatGPT, Claude or Gemini and ask them to try it on their own tasks. It costs little, and a week later you will have people who understand what the model can do and will suggest for themselves where it could help.

02Days 8–14: prepare the data and the rules

Write down how the task looks today: how long it takes, how many errors it produces and who does it. Without this baseline, nobody will be able to say a month from now whether things got better.

Set up access. The agent should reach only the data the task requires, with read-only rights wherever money and accounting are involved. Collect the rules: the tone for talking to customers, what is off limits, examples of good and bad replies. Twenty real examples are worth more than a page of general wording.

Decide what counts as a good result. For a reply to a review it might be “can be sent without edits”; for an invoice check, “every line item and amount matches the document”.

03Days 15–21: the silent run

In the third week the agent works alongside people and sends nothing outside the company. An employee does the task as before, the agent prepares its own version, and at the end of the day you compare the two. We run a week like this in every implementation we do and call it the silent run.

This week usually brings up things that were invisible on paper: exceptions only an experienced employee knows about, data stored somewhere other than where everyone thought, wording a customer would find odd. Turn each of these findings into a rule for the agent.

At the end of the week, count what share of the agent's results can be used as they are, what share needs edits and what share is unusable. This is the most important number of the month.

04Days 22–30: switch it on and decide

If most results in the silent run could be used without edits, put the agent to work with human approval: the agent prepares, and an employee sends or edits the result with one click. After a week, compare time and errors with the baseline you recorded in week two.

On day 30, make one of three decisions. Expand, if the savings are visible and the agent makes no more errors than people did. Refine, if the result is close and specific problems with clear fixes stand in the way. Close, if the task turned out to be a poor fit. A closed pilot is also a result: the test took a month, and now you know where to go next.

Appoint an owner for the agent: a person who is responsible for its quality, updates its rules and reports once a month on how much time was saved and how many errors there were.

05Rules for the whole team

While the pilot runs, write a short set of AI rules for staff. They should say which services are allowed, which data must never be pasted into a chat (customers' personal data, passwords, contracts with confidentiality terms) and who checks texts that go to customers on the company's behalf. If you serve customers in the EU, the personal data point falls under the GDPR, so settle it together with whoever handles data protection in your company.

A one-page document like this removes most of the risk and spares the team from wondering whether they are allowed to use these tools at all.

06Instead of a conclusion

The printing press changed the world once people stopped making manuscripts on it. The first month with AI serves the same purpose: to stop copying the old process and see where the new technology gives a different result. One task, an honest baseline, a silent run and a decision on day 30 will give you more than a year of discussion.

If you want to go through this month with us, see our [AI automation](/en/services/ai-automation/) page or message us on Telegram @mrkt_era.

In short

  1. Start with one repetitive task whose data lives in a system and whose result you can check quickly
  2. Before launch, record a baseline: how long the task takes, how many errors it produces and who does it
  3. Week three is a silent run: the agent works alongside people and sends nothing outside the company
  4. On day 30, decide whether to expand, refine or close the pilot, and appoint an owner for the agent
  5. One page of AI rules for staff removes most of the risk

Related serviceAI agents for business: bookkeeping, content, SEO

Want to discuss how this would work for you? Message us on Telegram @mrkt_era.

Start

Let's start with a review

A document with priorities and a 45-minute conversation. Then you decide: on your own, with us, or not at all.

book a €250 review
€250 reviewbook