When a client asks where to start with AI, our answer is usually duller than they hoped. Start with your documents. Specifically, start with the documents people retype by hand every day.
Every business we work with has a stack of them. Invoices, purchase orders, delivery notes, certificates, application forms, supplier statements. Someone opens the PDF, reads a number, types it into a system, then does it again forty times before lunch. It is slow, it is error prone, and nobody wants the job. That is exactly why it is a good first project.
Document extraction is a strong starting point for three plain reasons.
The problem is real and measurable
You can count it. How many documents a week, how many minutes each, how many typos slip through. Before you write a line of code you know what success looks like and roughly what it is worth. That clarity is rare in AI work, and it keeps the project honest.
The risk is contained
Extraction does not act on the world. It reads a document and proposes values for a human to confirm. If it gets a field wrong, a person catches it at the review step, the same step they were already doing when they typed everything manually. You are not handing the keys to anything. You are taking the typing away and keeping the judgement.
It teaches you about your own data
This is the part people underestimate. The moment you try to extract fields cleanly, you discover that your suppliers use nine different layouts, that half your forms have a handwritten note in the margin, that the date format is never the same twice. That is not a setback. That is the real shape of your operation, finally visible. Sorting it out makes every later AI project easier, because the mess was always going to surface eventually.
In practice we treat the first version as a draft assistant, not an oracle. The model reads the document and fills the fields. The person reviews a clean side by side view, fixes anything off, and confirms. Over a few weeks you watch where it stumbles. Maybe one supplier's layout trips it up, so you handle that case explicitly. The accuracy you can trust is something you measure, not something you assume.
A Discovery Sprint suits this work well. In two to four weeks you can take one real document type, build a working prototype, run it against your own back catalogue, and see honest numbers. No theory, just your invoices and a tool that reads them.
We are not against ambitious AI projects. We just notice that the teams who start here build the habits, the trust, and the clean data that make the ambitious projects actually work later. Begin where the pain is obvious and the risk is low. Document extraction is usually that place.
Facing something similar in your business?
Talk it through with our AI guide, or send the team a note. We will tell you straight whether and how we can help.