There is no responsible universal price for a custom RAG system. A small internal manual assistant and a multi-department tool connected to permission-heavy business systems are different projects. A proposal should make that difference visible rather than hide it behind an “AI package”.
Separate the initial build from the operating bill
The initial scope usually includes source review, ingestion, retrieval, the answer interface, access controls and evaluation. Ongoing costs may include hosting, search storage, model usage, document processing, monitoring and support. Ask which items are included, which are usage-based and who pays the providers.
Not Devs quotes against agreed requirements. We do not publish a made-up starting price that assumes your sources are clean, your integrations are simple and your access rules can wait.
What changes the development effort?
- Information quality: readable, maintained documents are different from scans, duplicates and conflicting versions.
- Connectivity: a controlled import is different from continuous synchronisation with several platforms.
- Permissions: a single shared knowledge collection is different from document-level access across multiple teams.
- Workflow: an internal answer tool is different from a customer-facing service with escalation and audit requirements.
- Evidence: credible evaluation takes representative questions and time from people who understand the material.
Make the pilot answer a commercial question
Choose a recurring task with an observable baseline. For example: how long does a support colleague spend finding and checking an approved response today? Record the current effort before introducing the assistant, then compare representative work during the pilot.
Do not measure speed alone. Track incorrect answers, missing evidence, unsuitable retrieval and the time people spend verifying responses. A faster answer that causes rework is not a useful improvement.
A transparent example—not a savings promise
Assume 20 people each make four knowledge searches a day, each taking three minutes. Across 20 working days that is 80 hours of searching. If a pilot demonstrated a 25% reduction, the estimate would be 20 hours reclaimed per month.
Those are illustrative inputs. They do not show cash savings, additional revenue or a guaranteed return. Adoption, checking time and changed working practices affect the result. Use the RAG time-value calculator with your own assumptions, then test them.
Ask for a proposal you can compare
- Which source systems and document formats are included?
- Which users and access rules will be supported?
- What does an acceptable answer look like, and how is it tested?
- How are source changes, deletions and failed updates handled?
- Which provider and support charges continue after launch?
- What is handed over, and who operates the system?
Our discovery, pilot and production approach gives those questions a place in the project before a wider rollout.
Sometimes the best next step is not RAG
If the problem is inconsistent data entry, an absent process owner or a disconnected operational workflow, fix that foundation first. A knowledge assistant should solve a defined information problem—not become a costly distraction from it.
For context, start with what RAG can and cannot do for a business, or bring your source list to a scoping conversation.