Where artificial intelligence works, and where not
Most of the market will say yes to everything. This is the page where we also say no, because that is what makes the yes worth anything.
Four uses that hold up in production today
Document classification and data extraction. Invoices, case files, forms, contracts. When volume is high and the format repeats, pulling fields automatically and leaving human review for the doubtful cases saves real time.
Automatic summarisation. Minutes, long email threads, lengthy files. It exists so you read less, not so you decide without reading.
Automated interaction quality evaluation. In a contact centre, reviewing one per cent of calls by hand leaves the other ninety-nine unexamined. Automated evaluation covers all of it and frees the analysts for the cases that need a human ear.
Internal assistants over company documentation. An assistant that answers with your own policy, procedure or contract, and cites where the answer came from.
Four cases where it only adds risk
When the decision carries legal or financial consequence and no human reviews it afterwards. A sworn translation, a credit approval, a dismissal. Speed was not the problem there.
When the volume is low. If the process runs forty times a month, the cost to build and maintain never comes back. A template and a person do it better.
When the data cannot leave the company and there is no budget to deploy it inside. We say so before the contract, not after.
When the process is not defined. Automating ambiguity produces faster ambiguity. That is an A2 project, not this one.
Short scope, measurement, clean exit
Every project in this vertical starts with a bounded case and a numeric success criterion agreed before anything is built. If it is not met, we say so and stop.
Human review is designed in from the start, not bolted on when something breaks. And everything built is documented so it can be switched off without breaking the process that existed before it.
Frequently asked questions
Do my documents train somebody else's model?
Not in the deployments we build. We use configurations that neither retain nor train on client content, and where the requirement is stricter, the model runs inside the client's own infrastructure.
What happens when it gets something wrong?
It will, which is why the design specifies where and how it is reviewed. The right question is not whether it fails, but what happens when it does and who notices.
What does it cost to keep running?
Running cost is estimated before building, at real volume. If that number does not come back, we say so in the diagnostic and it does not get built.
Does it work well on Spanish-language documents?
Yes, including Guatemalan legal and clinical documents. The care goes into each client's own terminology, and that is solved with a glossary, not with optimism.
At what point would you tell me no?
In the diagnostic, which is free, and before quoting. Saying no there costs an hour. Saying it six months in costs the relationship.