A guide by SevenLab
Searching a manual, extracting invoice fields and updating an ERP record are different problems. A sound document AI brief separates them before choosing a chatbot or integration approach. This guide helps you assess a document-processing or RAG project with SevenLab without assuming that named business systems already have supported connectors.
Distinguish extraction, search and system updates
Document extraction turns content into fields that can be validated, such as invoice references or contract dates. OCR may be needed when text is embedded in scanned images. Retrieval-augmented generation, or RAG, instead retrieves relevant passages to help answer a question. Neither approach automatically makes the result correct or authorises a change to another system.
Illustrative scenario: a finance team receives recurring batches of supplier PDFs. It needs field definitions, duplicate checks, exception handling and an approved route into accounting, not simply a chat interface. A legal operations team searching contract clauses has a different need: source passages, document versions and professional review. A one-off PDF extraction task may not justify a custom application.
- Choose the output: validated fields, retrieved passages or a proposed action.
- Collect representative scans, tables, attachments and difficult examples.
- Confirm that recurring volume and process needs justify custom development.
Make source quality and access part of the design
A knowledge assistant needs an agreed source collection, current versions and rules for what each user may retrieve. Ask how deleted or replaced documents would leave the index and how answers would refer back to supporting passages. Test unanswerable questions and contradictory documents. RAG can improve access to information, but does not eliminate incorrect answers or replace specialist judgement.
Chemgas is a documented SevenLab example of a maritime procedure RAG assistant. That case reports 90% less lookup time. This figure belongs to the Chemgas case only and is not a forecast for another organisation. Use the case to frame questions about your own procedures, source maintenance and evaluation rather than assuming the same outcome.
- Assign an owner to every source collection and its update process.
- Test access restrictions and references to exact document versions.
- Include questions that should receive no answer or require escalation.
Assess each integration before promising a connected workflow
Treat Dynamics 365, Exact Online, Salesforce, SharePoint and Microsoft Teams as assessment subjects, not confirmed SevenLab connectors. The same applies to proposed ChatGPT connections, n8n implementations or Make workflows. For each requirement, verify available APIs, licences, authentication, rate limits and whether the task needs read access or write access. A product name alone is not an integration specification.
Illustrative scenario: an organisation wants to search SharePoint documents and propose updates to a legacy ERP. Retrieval and record updates should be scoped separately, with permission checks, field mapping and a recovery path for failed writes. If a legacy application has no suitable interface, compare a controlled export, a limited integration layer and modernisation rather than assuming screen automation will be reliable.
- List systems, versions, owners and required read or write operations.
- Verify API access, licensing and permission boundaries with each system owner.
- Specify retries, duplicate prevention and recovery before allowing writes.
Validate with real documents and plan ongoing ownership
Build an evaluation set that reflects the work: poor scans, unusual layouts, missing pages and conflicting versions. For extraction, check individual fields and the downstream consequence of a wrong value. For RAG, assess whether retrieved evidence supports the answer and whether the right user can access it. Keep a human route for cases the system cannot resolve safely.
InShared is a SevenLab case source for document processing, not a basis for an assumed improvement figure. Review that case alongside your own acceptance criteria. Before delivery, agree source updates, access reviews, monitoring and incident ownership. Consult the existing pricing page, then request a proposal that separates document preparation, integration, evaluation and ongoing usage.
- Create an acceptance set with normal cases and costly failure cases.
- Assign owners for document updates, access reviews and unresolved items.
- Use the pricing page and request a scope-specific implementation and operating estimate.
