Advice about adopting artificial intelligence is mostly written for organisations with a data team, a transformation budget and someone whose full-time job is this. If that does not describe your company, most of it is not merely unhelpful — it is actively misleading about what the first step should be.
Here is a version scaled to a company where the person evaluating AI also has another job.
You do not need a strategy document
Large organisations need strategy documents because dozens of people must coordinate. In a company of thirty, a strategy document is usually a way of postponing the moment when something actually gets tried.
What you need instead is one specific process, described in plain language, with a named person who owns it. That is enough to start and it produces more learning in six weeks than a strategy will in six months.
Start with the assistant, not the integration
There is a sequencing decision here that matters. Giving your staff a proper AI assistant — inside your own environment, under your own contracts — is cheap, fast and requires no process analysis.
It also does something that no amount of planning achieves: it teaches your people what these systems are good and bad at, using their own real work. After two months of that, the conversation about which process to automate becomes dramatically better informed, because people are speaking from experience rather than from headlines.
It has a second benefit. If staff are already using consumer AI tools informally, this gives them a sanctioned alternative and brings that usage back inside your control.
Then pick exactly one process
Choose something repetitive, text-heavy, and where a mistake gets caught by a human before it reaches anyone outside the company. Resist the urge to pick the most impressive candidate. Pick the one where failure is cheapest, because your first project's real output is knowledge about how your company handles this.
Budget for the boring parts
The cost of an integration is rarely dominated by the AI. It is dominated by connecting to your existing systems, handling the inputs that do not fit the pattern, and maintaining it as things change.
A common and avoidable mistake is spending the entire budget on the build and having nothing left for the six months afterwards, when the process reveals what it actually needs. Hold something back.
Settle the data question before you start
Where is the data processed? Who is the processor and under what contract? Is there a data processing agreement under Art. 28 GDPR? Are your inputs used to train anyone's models?
These are quick questions with quick answers from a serious provider. Asking them at the start costs nothing. Discovering the answers after two years of use is considerably more expensive, particularly if a customer asks you first.
Judge it honestly after three months
Set a date at the outset for deciding whether it worked, and define what "worked" means in advance — a share of items handled without correction, a reduction in turnaround time, something checkable.
Projects without a review date do not end. They fade, which means you never find out whether the approach was wrong or just the first attempt.
The realistic timeline
Assistant available to staff: a few weeks. Useful understanding of where AI helps in your specific business: two to three months of real use. First process integration running in production: a few months after that, depending on how tangled the process turns out to be.
Anyone promising substantially faster than that is describing a demo, not a working process.
