"We should be using AI too" now comes up in almost every management meeting. But what to do after that sentence often remains unclear. The answer to the question of where a business should start with AI actually has more to do with choosing the right task than with technology: which task, who does it, how long does it take, and how will we know it's better? In this post we explain how to keep your first project small, safe and measurable.
Why not start with a "big project"?
The first wave of enthusiasm about AI usually produces ambitious ideas: automating all of customer service, handing sales forecasting entirely to a machine, gathering all company documents into a single smart assistant. These ideas aren't bad, but they're too big for a first step.
In big projects, three problems pile up. First, nobody knows exactly where the data is or what state it's in. Second, success hasn't been defined; "we'll be more efficient" isn't a measurable goal. Third, the team hasn't yet experienced where AI is strong and where it's weak. The result is months of preparation with no concrete gain to show.
Starting with a small task limits both the budget and the risk of disappointment. On top of that, by seeing how AI behaves on a real task, the team builds solid intuition for the next steps.
What does a good first task look like?
A good first AI task usually has these characteristics:
- It repeats often. Not a few times a week, but dozens of times a day.
- It's text-heavy. Reading emails, summarizing documents, classifying form data or drafting proposals are areas where today's language models are strong.
- The cost of errors is low. A person reviews and approves the output; a wrong suggestion doesn't go straight to a customer or to accounting.
- The result is measurable. You roughly know how much time is spent on the task today or how much delay it causes.
- It has a clear owner. The person who does the task today takes part in the project and evaluates the result.
For example, sorting incoming request emails by topic and routing them to the right person, turning long technical specifications into one-page summaries, or compiling frequently asked questions from internal correspondence and preparing draft answers are all good candidates. We cover the email side separately in our post on classifying incoming email with AI.
Your first project, step by step
Running the first project in this order keeps things simple:
- List the tasks. Ask your teams to write down three tasks that "repeat every day and wear you out."
- Pick one. Choose the one that best matches the five characteristics above, not the most exciting one.
- Record the current state. Note the average time the task takes, its monthly volume and the errors that happen often. Without this record, it's hard to talk about gains later.
- Check the data. What data is the task done with, and does it include personal data or trade secrets? Don't skip this question; the details are in our post on giving company data to AI tools.
- Set up a small trial. A trial of a few weeks with a limited number of users is enough.
- Measure the result and decide. Continue, adjust or drop it. All three are valid outcomes.
Off-the-shelf tool or custom solution?
For a first trial, off-the-shelf tools are often enough. The business versions of general-purpose chat tools such as ChatGPT, Gemini or Claude offer a good starting point for a team to get to know AI. What matters here is reading the data terms of the version you use and telling the team clearly what can and can't be entered.
Off-the-shelf tools start to fall short when the task has to connect to your existing systems. For example, automatically logging emails into the CRM, or an assistant that works with ERP data, requires an integration and sometimes custom software. If you've reached that point, the first trial has already proven its value.
Checklist before you start
| Question | Is the answer clear? |
|---|---|
| Which task, who does it? | The task and its owner are written down |
| How long does it take today? | Roughly measured |
| What data will be used? | Personal data and confidential information separated |
| Who will approve the output? | Human review is defined |
| How will success be measured? | A single main metric chosen |
| When will the trial end? | The date is set |
If even half of the rows in this table are blank, filling them in is worth more than choosing an AI tool.
Discuss AI's limits from the start
Language models write fluently but not always correctly. They can produce unsourced information in a confident tone, make calculation errors, and they don't know your company's specific rules. That's why in a first project, the "AI drafts, a person approves" setup is almost always the right one. When the team knows these limits from day one, it prevents both needless disappointment and uncontrolled use.
How we do it at Globya
As a company's digital IT consultant, we start with a list of tasks, not with technology. In our first conversation by phone, we talk through your repetitive tasks together, and as part of the free preliminary analysis we suggest the best first candidate and a way to measure it. If the trial is successful, we handle integration, software, KVKK (Türkiye's Personal Data Protection Law) and hosting as a single point of contact; you don't need to look for separate vendors. Pricing follows in a written proposal once the scope is clear.
Frequently asked questions
Do we need a big data infrastructure to start with AI?
No. First projects are usually done with text you already have, such as emails, documents or forms. A big data infrastructure only comes up at a later stage, once the need is clear.
How long should the first trial last?
For most tasks, a trial of a few weeks with a limited number of users gives you enough insight. What matters is that the start and end dates are set from the beginning.
If the trial fails, is the money wasted?
Even if a small trial fails, it shows which task isn't a good fit for AI. That knowledge protects you from a bigger and more expensive mistake.
What should employees watch out for when using AI?
The two basic rules are not entering personal data or trade secrets into unapproved tools, and not using the output without checking it. For this, we recommend preparing a written AI usage policy.
The Globya assistant is online 24/7; it answers right away and passes your question to the team if needed.