Introductory call
Thirty minutes to understand the document problem, who is affected, what a good result looks like, and any security or deployment constraints. You leave knowing whether RAG, OCR, or both fit.
How it works
How to order a RAG or OCR project.
Every engagement follows the same path: a short conversation, a look at your real documents, a scoped pilot with measurable acceptance criteria, then production.
We do not sell a fixed package. The scope is built around your documents, the answers or fields your team needs, and the decisions people must keep. Along the way we settle one question early: whether the workflow needs an AI agent at all, or whether RAG, OCR, or both are enough on their own.
1 callto define the document problem
Documentsdrive the scope, not a price list
Pilotmeasured before production
Agentonly when the workflow must act
The ordering process
Six steps from a first conversation to a working system.
Each step ends with something you can read, test, or approve. You can stop after any step with the work so far in hand.
Document and source review
We look at representative files, knowledge sources, formats, scan quality, and the systems involved. This is where we spot layout variation, permission boundaries, and missing information.
Written scope
A short proposal naming the document set, the required outputs, the evaluation approach, the review path, the deployment boundary, and the acceptance criteria the pilot must meet.
Pilot
We build against your real documents and a reviewed test set. Retrieval and extraction quality are measured separately from answer quality, so a fluent response cannot hide weak evidence.
Production deployment
The validated pipeline connects to approved sources, users, and business systems with access controls, monitoring, and the human review points agreed in the scope.
Managed optimisation
After launch we review extraction accuracy, retrieval relevance, exceptions, usage, and new document patterns, and adjust the system as your sources change.
Agent decision
Do you need an AI agent, or just RAG and OCR?
Most document problems are solved without an agent. RAG answers questions from approved knowledge with citations. OCR turns scans and forms into validated data. An agent is added only when the workflow must take an action after the answer or the extraction, and even then a person approves the consequential steps.
Your workflow needs
What we recommend
Answers to questions from approved documents, with sources
RAG only. No agent.
Structured data out of scans, PDFs, forms, or invoices
OCR with validation and exception review. No agent.
Search across scanned archives and return cited answers
OCR feeding a RAG index. No agent.
Someone must act on the result: route an exception, draft a reply, update a system
Add a controlled agent with explicit tools and defined approval points.
Actions with financial, legal, or clinical consequences
An agent may prepare the action. A named person approves it. Never automatic.
Signs RAG or OCR alone is the right scope
- The output is read by a person, not written into a system
- Users ask questions and need cited answers they can check
- Documents need to become fields, tables, or searchable text
- The next step after the result is a human decision
Signs the workflow needs controlled action
- A repeatable step follows the answer or extraction every time
- The step touches another system: ticketing, ERP, email, a case record
- Approval points and audit logging are defined before automation
- Low-risk steps can run; uncertain or high-impact ones must wait for review
What we need from you
Bring the documents and the decision, not a specification.
You do not need a technical brief. These six things let us scope a realistic pilot in the first two steps.
Representative documents
Typical files plus the difficult ones: poor scans, unusual layouts, multiple languages, and examples that should be rejected.
The questions or fields
For RAG, the questions people actually ask. For OCR, the fields and tables the downstream process needs and which ones matter most.
The current process
Who handles the documents today, how long it takes, where mistakes happen, and who owns exceptions.
Connected systems
Where documents arrive and where results must go: shared drives, document management, ERP, CRM, ticketing, or a review queue.
Security constraints
Data sensitivity, residency, retention, and whether processing must stay on your infrastructure. See our deployment options.
What success means
The accuracy, speed, or coverage that would make the pilot worth taking to production, and who signs it off.
Scope factors
What shapes the scope and timeline.
Two projects with the same goal can differ widely in effort. These are the factors we assess in the document review, and they are what we discuss with you before writing the scope.
Volume, variety, and how quality will be judged
Formats, scan quality, languages, layouts, tables, the number of extraction fields, the number and freshness of knowledge sources, permission boundaries, and how large a reviewed test set the risk justifies.
Where it runs and what it connects to
System connections, access controls, review interfaces, and the deployment model: commercial model APIs, private cloud, on-premise GPU, or a hybrid design. Each changes the build and the ongoing operation. Compare the deployment options and the security controls.
After the pilot
Two ways a successful pilot goes live.
A bespoke RAG or OCR system in your environment
The pilot pipeline is hardened and connected to your sources, users, and business systems. You own the architecture, and the deployment boundary is the one agreed in the scope. Most regulated or high-volume workflows take this route.
See a production RAG applicationAn ongoing document and knowledge workspace we run for you
For teams that want the outcome without operating a system, the same RAG, OCR, and controlled-agent capabilities are available inside our managed workspace: role-based access, approved knowledge and context, scheduled document workflows and reports, human approval controls, and an audit log. Volume and access are agreed during scoping.
Explore the delivery platformStart the process
Bring one document workflow to the first call.
Tell us about the documents, the questions or fields, and the decision people make today. We will say plainly whether RAG, OCR, or an agent is the right starting point, and what a pilot would need to prove.