In a company of ten or fifty people, AI usually arrives by accident. Someone starts using a chat assistant, someone else pastes a contract into a free tool, and the managing director asks whether "we are doing something with AI". There is no plan, and no one can say whether any of it helps.
This article is for the managing director or owner of a small company who wants a sensible first month. It does not cover tool selection in detail or a long-term strategy. It suggests three small trials, a way to judge them and a list of things to postpone.
The aim of the first month
The aim is learning. After thirty days you should know three things: which tasks in your company AI helps with, how much checking its output needs, and who on your team is willing to work with it. A first month that ends with "we tried it, here is what we saw" is a success, even if one of the trials fails.
A larger audit of where AI could save time is a good next step, and we describe it in how to find out where AI will save your company time and money. The first month comes before it and gives you the experience to run it well.
How to choose the three trials
A good first trial has four properties:
- Frequent. The task happens at least weekly, so you get enough examples in a month.
- Low risk. A wrong result is caught internally and does not reach a customer or change a financial record.
- Easy to check. A person who knows the work can judge the result in a minute or two.
- Contained. It uses information you are allowed to share with the tool you choose. If you are not sure, use a task with no personal or confidential data.
Tasks that often fit: first drafts of routine emails, summaries of long documents or meeting notes, cleaning and sorting a list, preparing a first version of an internal procedure, turning rough notes into a structured page.
Three example trials
Trial 1: drafts of routine writing
One person prepares first drafts of a type of text they write often, for example replies to common enquiries or a weekly internal update. They edit every draft before it is used and note how much editing was needed.
Trial 2: summaries of long material
Choose a regular source of long text: supplier documents, internal reports, meeting notes. The owner asks the tool for a short summary and compares it with their own reading for the first several items. They record whether anything important was missed or invented. On the risk of invented content, see why AI tools produce confident mistakes.
Trial 3: tidying and sorting data
Pick one small, messy list: a contact list, product names in different spellings, a set of free-text answers. The tool proposes a cleaned or categorised version, and a person checks a sample before anything replaces the original. Always keep the original.
Write the success test first
Before each trial begins, write down in two lines how you will judge it. Examples:
- "A draft counts as useful if editing takes less than half the time of writing from scratch, in at least most cases."
- "A summary counts as useful if the reader finds nothing important missing and nothing invented in the first ten items."
- "Sorting is useful if a sample check finds errors in no more than a small, agreed share of rows, and fixing them is faster than doing the job by hand."
Choose your own thresholds. The point is that the test exists before the result, so nobody has to argue afterwards about whether it worked.
Give each trial an owner and a little time
Each trial needs one named person and about an hour a week. Without this, trials fade after a few days. The owner keeps a short log: date, what was tried, what the result was, how long the checking took. A shared page of ten lines is enough.
Agree a few simple rules at the start: which tools may be used, what kind of information must never be pasted in, and that every output is checked by a person before it is used. If you want a more formal document later, a simple usage policy can grow out of these three rules.
What to leave for later
Writing this list down is as useful as choosing the trials, because it protects the month from enthusiasm. Typical items to postpone:
- Anything that talks to customers without a person in between, such as an automatic chat on your website.
- Decisions about people or money, such as ranking candidates or approving payments.
- Large integrations with your accounting, ERP or CRM systems.
- Buying a platform or signing a long contract before you know which tasks matter.
- Training your own model. For a small company this is almost never the right first step.
None of these is forbidden for ever. They need evidence from the first month and usually more care than a small trial can give.
The review at the end of the month
Set a 45-minute meeting in the last week. For each trial, read the log and apply the test you wrote at the start. Then decide one of three things: continue as it is, adjust and try again for another month, or stop. Stopping a trial that did not work is a good result because it saves money later.
If a trial worked, think about what it takes to make it routine: who will do it, what checking remains, and who looks at the results. A structured way to run this decision for a larger pilot is described in how to review an AI agent pilot and expand, adjust or retire it.
What to do this week
List five tasks your team repeats every week. Cross out those that touch customers directly or involve sensitive data. From the rest, choose three that are easy to check, name an owner for each, write the two-line success test and set the review meeting in your calendar for four weeks from now.
Sources
- Our related articles linked above.
- The choice of three trials, the selection criteria and the review steps are a general working approach, not a standard; adapt them to your own company. No figures or client results are claimed in this article.