02 — Applied AI
A source you can read. A proposal you can assess.
Reading a document, finding information or preparing a draft can be a useful task for a model. The work is to define that task, examine its errors and make the result verifiable by the responsible person.
From source to review
An example of document preparation with an explicit human review step.
Authorised source
An inspectable passage, connected to its document and version.
Proposal
An extraction or draft, with references and uncertain elements identified.
Human review
The responsible person confirms, corrects or rejects the proposal.
Information without sufficient source material remains unverified. A proposal is not a decision.
Choose a task that can be assessed
The first use case should be possible to describe and evaluate: extract fields, find a passage or draft a response from authorised documents. Test examples need to include missing information, ambiguity and contradictions.
Initial exploration can use synthetic data. Real documents require an explicit processing scope and authorisation; removing a name does not necessarily make a document anonymous.
Design the review alongside the result
A proposal should point to the material needed to assess it. An extracted field can show its source passage; a draft can distinguish documented facts from missing information. An answer without adequate sources remains unverified.
An authorised person confirms, corrects or rejects the proposal. Any actions an agent could perform need defined permissions and approval conditions. Compliance decisions remain human decisions.
A scope to agree together
- A defined use case, authorised inputs and explicit limits
- Reference examples and evaluation criteria
- A journey for reviewing proposals, sources and corrections
- Integration of the models or tools within the agreed scope
- Documentation of processing, dependencies and review arrangements
Where it can help
- Teams reviewing documents or preparing recurring responses
- Professionals who need to trace information to its source
- Organisations testing a specific use case before expanding it
Possible technical choices
- TypeScript
- Validation schemas
- Document retrieval
- Versioned evaluation sets
Common questions
- Does the model need to be trained on our documents?
- Training is not a prerequisite. Retrieval from authorised documents, defined instructions and business rules can address some needs. The choice depends on the use case and the evaluation results.
- Where is the data processed and retained?
- These conditions must be checked for each provider and configuration before real data is used. The project specifies authorised data, processing locations, retention and access. A general no-retention promise cannot replace that verification.
- How can someone examine an incorrect answer?
- The interface should let them inspect the available sources and correct or reject the proposal. Observed errors become test examples. Any confidence score is an indicator, not evidence or automatic approval.
What would you like to make simpler?
A client journey, an internal tool, a document task: tell us about the context and the change you want to make. That first conversation helps define the work ahead.
Discuss a project