00The need
Some stretches of a process require reading and interpreting: a document that arrives as a PDF, an email that has to be understood, a text that has to be drafted from some data. Until now a person did that, and that is exactly where AI adds value, if it is used with limits.
A good fit if…
- Someone reads documents to pull data out of them
- Repetitive texts are drafted from data
- You want to use AI without losing control of what it does
01Extractor, not autonomous agent
The difference is not one of technology: it is one of design. We give it a specific stretch and clear limits, not the whole process.
- What it receivesA goal, and it decides what to look for
- If the data isn't thereIt fills in the most likely answer
- If it doesn't succeedIt keeps trying until something comes out
- When auditing itHard to tell at which step it went wrong
- What it receivesThe source, and which data to pull from it
- If the data isn't thereIt returns empty
- If it doesn't succeedLimited retries, then a person decides
- When auditing itThe output follows a defined schema
Key idea
If the process can be solved with fixed rules, AI isn't needed.
02What we use it for
Two uses, both on documents, and both with AI placed inside a flow that validates what it returns.
Structured extraction
Pulling data out of a document
Invoices, delivery notes, work reports: AI extracts the fields and the flow validates them before saving.
Document generation
Assembling a document from a template
AI drafts only the parts that require interpretation, and the rest is assembled deterministically.
Key idea
Never on its own: always inside a flow with rules.
03How we make it reliable
Four rules, the same in every flow we build.
- Defined output schema
The response always has the same shape, and the rest of the flow knows what to expect.
- Empty when there is no data
If the data doesn't appear in the source, it returns empty: it doesn't infer or approximate it.
- Limited retries
A limited number of attempts, not a loop until something comes out.
- Escalation to a person
What can't be resolved with confidence goes to someone who decides.
Key idea
Reliable doesn't mean infallible: it means that when in doubt, it alerts.
04Which models
We are not tied to any one model. The one used in each process is chosen for what that process requires (cost per run, extraction quality, data limits), and it can be swapped without rebuilding the flow.
Key idea
The model is one more piece, not the center of the system.