AI for business: where to start?
The first steps to consider when integrating AI into your operations.
The most common mistake we see in AI projects is starting with the technology. When "which model should we use?" comes before "which job are we trying to fix?", the result is a system that works but nobody uses.
The order needs to be reversed.
1. Start with a repetitive task, not a model
A good first AI project has three properties: it repeats often, its rules are reasonably clear, and its errors are not fatal.
Tasks that meet all three are usually hiding in plain sight:
- Routing incoming requests to the right team
- Drafting the standard sections of proposals and contracts
- Giving the first response to the most frequently repeated customer questions
- Extracting data from invoices, waybills and forms
Equally, the areas to avoid in a first project are clear: processes that produce binding decisions on their own, are hard to reverse, and whose errors reach the customer directly.
2. Define success before you start
"Improve efficiency" is not a measurable goal. A measurable goal looks like this: cut first response time from 4 hours to 15 minutes, or halve the time it takes to produce the first draft of a proposal.
If you don't set that number before the project begins, you'll end up debating whether the system worked at all — and nobody wins that argument.
3. Find out where your data actually lives
Most organizations have data, but it is scattered: some in the ERP, some in email, some in a spreadsheet on someone's laptop. AI does not fix that scattering; it is built on top of it.
So the first week usually goes not into choosing a model, but into mapping where the data sits and who can reach it. That inventory is also the foundation for the next point.
4. Don't leave the legal side until last
The legal dimension of an AI project is not an afterthought to be handled with "let's also look at the contract" at the end. It is part of the design.
The headings that come up most often in practice:
- Personal data: If the content you feed the model contains personal data, processing it falls under data protection law. Does your privacy notice actually cover this processing?
- Cross-border transfer: Where are your provider's servers? If they are abroad, that is a transfer requiring separate disclosure.
- Confidentiality undertakings: Does the NDA you signed with your client permit you to send that client's data to a third-party service?
- Responsibility for output: Who is liable when the system produces a wrong output? This ties directly to where human approval sits in the process.
The answer to these questions is usually not "no, you can't" but "you can, under these conditions." Knowing the conditions upfront is far cheaper than discovering them after the system is built.
5. Build human approval into the architecture
In the first version, the system's output should not go straight to the customer. Put a human approval step in between — it lowers the cost of errors and shows you where the model gets things wrong.
Removing an approval step later is easy. Recovering from never having had one is not.
In short
AI accelerates a well-defined task; it does not define an undefined one. Start somewhere small and measurable, settle the data and legal framework upfront, and place human approval inside the architecture.
If you'd like to work out where to start, get in touch.