Service 04
Applied AI
A language model solving a defined task, with a verifiable result — not a generic chatbot.
Context
AI pays off when it's pointed at a narrow, boring task: reading a document and extracting a field, classifying a request, drafting the first version of a standardised text, summarising volume nobody has time to read.
The hard part isn't calling the model — it's deciding what to do when it gets things wrong. Every piece of work here includes a way to measure accuracy and a human review path where mistakes are expensive. Without that it isn't automation, it's luck.
Signs this is your case
- Someone reads dozens of documents a day to pull out five pieces of information
- Triaging messages or tickets consumes the team before the actual support starts
- You tried an AI tool and it worked in the demo and failed at volume
What you're left with
- An automated task with accuracy measured on your real cases
- Human review where mistakes are expensive, automatic where they aren't
- Cost per run estimated before it goes to production
What it isn't
- A corporate chatbot that answers anything about the company
- A promise to replace an entire team
A workflow in practice
Read documents with review where it matters
An illustrative workflow, not a client case study or a promised result.
See the monthly retainerExtract
AI proposes fields from the document and keeps a source reference for verification.
Validate
Rules check format and consistency. Uncertain fields and sensitive decisions go to a human reviewer.
Use
Only approved output reaches the system. Your team can trace the document and review behind each record.
Where I start
I test a representative sample of your documents, including difficult ones. We agree on an accuracy criterion and what always needs a person.
What continues every month
New formats and real failures enter the evaluation. I adjust the workflow and check quality before changing models or increasing autonomy.
Questions about this front
How do I know the AI will not make things up?
Because accuracy is measured on your own cases before it ships, and because where mistakes are expensive there is human review in the path. The hard part is never calling the model — it is deciding what to do when it gets things wrong. Without that it is not automation, it is luck.
What does it cost to run every month?
Cost per run is estimated before production, using your volumes. It is one of the few things in technology you can predict with reasonable precision, which is why it makes no sense to find out afterwards.