Many businesses want to start using AI and do not know where to begin. The most common mistake is beginning with something big and vague. The better approach is to pick one small, specific project that you can measure, learn from and, if it works, extend.
What makes a good first project
It is specific. Reduce the time to process supplier invoices is a project. Use AI in operations is not.
It is repetitive. The task happens often enough that a saving adds up.
It has a baseline you can measure. You know how long it takes and how often it goes wrong today, so you can tell whether it improved.
The inputs are reasonably consistent. Messy, wildly varying inputs make early results disappointing.
A mistake is recoverable. Start where a person can review the output before it matters, not where an error is costly or hard to reverse.
The data is accessible. If getting the information into the system is the hard part, that is the real project.
There is a clear owner. Someone cares about the result and will use it.
The blast radius is small. If it goes wrong, the harm is limited.
What to avoid
Do not start with a vague strategy and no use case. Do not begin with the most critical process in the business. Do not set out to replace a person, because that framing makes the people who know the work reluctant to help. And do not skip the groundwork. If nobody understands the current process well enough to describe it, document it first, as in how to document a process before you automate it.
A simple way to choose
List ten tasks that are repetitive and time-consuming. Score each on how often it happens, how much time it takes, how consistent the inputs are, how costly an error would be and how easy the data is to reach. The best candidates score high on frequency and time, and low on error cost. Pick one.
Run it as an experiment
Define success before you begin: for example, cut handling time by a certain amount while keeping accuracy at or above today's level. Build a prototype and test it on real examples, including the awkward ones. Keep a person in the loop for review. Measure against the baseline, counting the time spent on review and exceptions, not only the machine's part.
Decide on evidence
At the end, you have a result. If it worked, extend it, or move to the next candidate. If it did not, you have learned something cheaply about your data, your process or the technology, which is also a good outcome.
Keep expectations honest
AI is useful in specific places and oversold in general. Our piece on where AI delivers and where it overpromises goes through the realistic picture, and where AI document processing works and where it still needs a human covers one of the most common first projects in detail.
If you are thinking about a first AI project and want help choosing one worth doing, reach out. We will help you pick something small enough to learn from.
Have a question about this topic? We work with businesses worldwide.
Get in Touch