Almost every business has heard that artificial intelligence can save time. Many have tried a ChatGPT-style assistant to draft an email or summarise a document. Going from there to a project that changes how your team works is a different matter, and the usual question is always the same: are we ready, or is it too soon?
The good news is that you don’t need a data department or a large budget to find out. You need to look honestly at three things: which problem you want to solve, what information you have and who will take ownership. This article gives you concrete criteria to answer those questions and a sensible way to begin.
Start with a problem, not with the technology
The most common mistake is to start the wrong way round: “we want to use AI, where shall we put it?” That approach usually ends in an impressive demo that soon stops being used. Projects that work start with a task that costs time, money or mistakes today.
A good candidate task has these features:
- It happens often. Daily or weekly, not twice a year.
- It follows recognisable rules. An experienced person can explain how it’s done, even if there are exceptions.
- It involves text or documents. Emails, invoices, contracts, tender documents, product sheets or customer questions.
- It can be checked. It’s easy to tell whether the result is right or wrong.
- A mistake isn’t serious. Or, if it is, a person reviews the output before it has consequences.
Imagine a wholesaler that receives orders by email in different formats, and someone copies them by hand into the management software. That task ticks almost every box. “Improving the company’s strategy”, on the other hand, ticks none.
Signs that your company is ready
You don’t need all of them, but the more you have, the easier the first project will be.
- You can name a specific task and estimate how many hours your team spends on it.
- The information that task needs is accessible: in email, shared folders, your management software or spreadsheets.
- Someone in the company knows the process well and can spend a few hours explaining it and reviewing results.
- A decision-maker backs the project and accepts that the first version won’t be perfect.
- You know which data is sensitive (customer, health or financial data) and what must not leave the company.
Data: less perfect than you think, more important than it seems
Many people believe their data has to be “organised” before they can start. It doesn’t. Today’s language models, the technology behind conversational assistants, are quite good at reading messy documents, emails or scanned PDFs with legible text.
What does matter is something else:
- Access. The information can be reached automatically, without someone having to look for it by hand every time.
- Real examples. A set of solved cases lets you test whether the AI gets things right. Without examples, there is no way to measure.
- Permissions and privacy. Knowing which data can be sent to an external provider and on what terms. Providers offer different options on where data is processed and whether it is stored; review them before you choose.
What to expect from AI (and what not to)
Current AI is very good at reading, summarising, classifying and extracting data from text. It also writes reasonable first drafts. But it makes mistakes, and sometimes it makes them with great confidence: it can invent a detail that isn’t in the document.
That’s why, in a small or medium-sized business, the most prudent approach is usually this: the AI prepares and a person validates. Over time, once you have measured how often it gets a specific task right, you can decide which parts to automate fully and which to keep under review.
A good question for any AI proposal: what happens when it gets something wrong, and who notices?
Where to start: a four-step plan
- Choose a single task. The one that takes the most time and meets the criteria above. Resist the temptation to tackle several at once.
- Decide how you’ll measure success. For example: time per order, number of errors or how many cases a person has to correct. Write down where things stand today.
- Test with real data. A prototype, meaning a simple first version that works with your actual documents, shows within a few weeks where it gets things right and where it fails.
- Decide based on results. If the test works, it gets connected to your tools and extended. If it doesn’t, you have spent little and you know why.
Questions to ask any provider
Before you sign anything, ask for clear answers to these questions:
- Which specific task are we solving, and how will we measure whether it works?
- What data do you need, where is it processed and who can see it?
- Which AI provider will you use, and could we switch later?
- What happens when the system doesn’t know the answer or gets it wrong?
- How much will it cost to keep running afterwards, including the cost of using the model?
- Who owns the code and the configuration when the project ends?
If the answers are vague, or everything sounds like “the AI will do it on its own”, be wary.
Signs that it isn’t the right time yet
It also helps to know when to wait. It may be too early if you can’t name a specific task, if the process changes every week and nobody can explain it, or if the information you would need exists only in one person’s head. In those cases, the first step isn’t AI but organising and documenting the process. That work is never wasted, and it will make the later project cheaper.
The next step
Try this exercise this week: write down the three repetitive tasks that take the most time in your company, how many hours you spend on them and where the information they use lives. With that list you can already judge whether any of them is a good candidate.
If you’d like a second opinion, at Brains we review lists like this and tell you frankly whether there is a use case worth testing. You can see how we work on our AI projects page or get in touch directly.