RAG chatbot — off-the-shelf comparison¶
Internal reference for the "why don't we just buy Microsoft Copilot?" conversation, which will happen on almost every RAG deal. The honest answer is that several vendor products do most of what a client is asking for, and on the right-fit deal we should sometimes recommend they buy one of them instead of building. Saying no to a wrong-fit deal is how we stay trustworthy on the right-fit ones.
This page is intended for:
- The pre-sales team, to be ready for the question on a discovery call.
- Optionally, to be shared with a client who asks the question directly — better than improvising.
The matching playbook documents are the quote guide and the internal checklist.
The credible off-the-shelf options¶
| Product | What it is | Typical price shape | Fits best when |
|---|---|---|---|
| Microsoft 365 Copilot + Copilot Studio + SharePoint Agents | Chat over SharePoint / OneDrive / Teams content. "Agents" you build in a low-code studio. Honours existing M365 permissions. | [~£24–30] per user per month on top of M365 |
Heavy M365 user, documents live in SharePoint, permissions already modelled in M365 |
| Google Gemini for Workspace + NotebookLM Enterprise + Vertex AI Search | Equivalent for Google Workspace tenants. Vertex AI Search is the more configurable enterprise RAG engine for non-Workspace corpora. | Per-user on Workspace; consumption-based on Vertex | Google-shop clients; documents in Drive |
| AWS Q Business | Managed RAG over connected sources (S3, SharePoint, Confluence, Salesforce, ServiceNow, etc.) with role-based access | Per-user; "Q Business Lite" and "Pro" tiers | AWS-shop clients; multi-source enterprise search |
| Glean | Independent enterprise RAG / search vendor. Polished UX. | [~£30–60] per user per month |
Premium budget, many SaaS apps to unify, search-everywhere expectation |
| Writer / Moveworks / Mendable | Adjacent independent vendors with different specialisms — Writer focuses on content generation, Moveworks on IT helpdesk, Mendable on docs portals | Variable | Narrow vertical use cases |
| Azure AI Search + Azure OpenAI "Chat with your data" | Microsoft's reference pattern for building a custom RAG app on Azure. Deployable from a template. | Consumption-based (Azure AI Search + OpenAI tokens) | This is actually very close to what we'd build — just self-deployed on the client's Azure |
| ChatGPT Enterprise / Team + connectors | Managed connectors to SharePoint, Drive, Confluence, etc. with org-level admin. | [~£25] per user per month |
"We just want chat over our docs without a project" |
| Claude for Enterprise + MCP connectors | Anthropic equivalent. Model Context Protocol (MCP) servers connect to internal systems. | [~£25–50] per user per month |
Quality-of-answer matters most, smaller power-user audience |
Prices are illustrative and move; check current rate cards before quoting against them.
When DIY beats off-the-shelf — the honest list¶
If none of the points below apply to a client's situation, we should consider recommending an off-the-shelf product instead of building. If two or more apply, a custom build starts to make sense.
- Per-role document permissions that don't map to existing ACLs. Copilot enforces SharePoint / OneDrive permissions; if the policy the client actually wants is "HR sees disciplinary procedure, others don't" but that's not modelled in SharePoint, Copilot will not fix it. Glean, AWS Q and the rest are similar — they inherit the source system's ACLs, they don't invent new ones.
- Documents that live outside the big SaaS suites. Mixed estate — file shares, paper, legacy intranets, vendor portals, internal apps, Notion, ad-hoc Confluence. Off-the-shelf connectors cover the easy half well and the hard half not at all.
- Strict data-residency or on-prem requirements. Copilot, Gemini, AWS Q, Glean and ChatGPT Enterprise all flow data to the vendor's tenant. Some regulators / clients will not accept that; a self-hosted build is the only honest answer.
- Custom UX, custom workflows, custom guardrails. "Refuse to answer X", "always render side-by-side citations", "log every Q&A for compliance review with tamper-evident storage", branded internal portal, role-specific landing pages, multi-turn workflows that span more than chat — the off-the-shelf tools have ceilings on every one of these.
- Specific corpus quality issues. Off-the-shelf retrieval is generic. When the corpus needs domain-aware chunking, OCR cleanup on a specific scanner output, glossary expansion for industry jargon, deduplication of versioned documents, or hybrid keyword-plus-semantic search, the managed products do not give us the levers we need.
- Total cost of ownership at scale. Per-user pricing dominates
beyond ~200 users. At 1,500 users, Copilot at
[~£25]/user/monthis[~£450k]per year. A custom build with frontier API tokens, plus our managed-service fee, typically lands at a fraction of that. The crossover point is roughly 200-300 users — below that, per-user pricing is cheaper than a build. - Independence from a single vendor. Some clients explicitly want to be able to swap the LLM provider (e.g. move from Anthropic to a local model when their corpus grows sensitive). Off-the-shelf products lock model choice.
When off-the-shelf beats DIY¶
The other side of the same honesty — situations where we should either decline the deal or recommend a vendor product:
- Client is a small (~50-200 staff) M365 shop, documents are tidy in SharePoint, permissions are already correct, no regulator in the room → Microsoft 365 Copilot + Copilot Studio.
- Client wants search across many SaaS apps (Slack, Notion, Jira, Salesforce, Drive) with a polished generic UX, has the budget, and does not need bespoke workflows → Glean.
- Client is already deep in AWS and wants a fast enterprise-search experience over their existing S3 / Confluence / ServiceNow with minimal engineering → AWS Q Business.
- Client wants "ChatGPT but on our documents" with the lowest possible project effort and the quality of the answers matters more than corpus complexity → ChatGPT Enterprise or Claude for Enterprise with connectors.
Where we add value in those cases is delivery support: the client still needs help with rollout, content cleanup, change management, training, governance. None of these are zero-cost even on a managed product — they are the work we should be quoting for, honestly, rather than dressing them up as a custom build.
Questions to ask early¶
The questionnaire (client questionnaire) captures most of this implicitly, but on a first call we should ask explicitly:
- "Do you already use Microsoft 365 Copilot, Google Gemini for Workspace, ChatGPT Enterprise, or any similar tool? If yes, what is working and what isn't?"
- "Have you evaluated the off-the-shelf options? If you ruled them out, why?" (Their answer either confirms the build is right or surfaces a misunderstanding we can correct.)
- "Are your document permissions accurately modelled in the system the documents live in today?" (This is the single most common reason an off-the-shelf tool falls short.)
If a client says "we never thought about Copilot" we should bring it up ourselves. Hiding the option is short-term thinking.
Where this sits in the proposal¶
When the off-the-shelf comparison matters, the proposal should include a brief "alternatives considered" section naming the relevant vendor product and why it was set aside. Two paragraphs is plenty. It signals to procurement that we did the diligence; it also gives the client's CFO a clean line to defend the spend internally.
Self-hosted / FOSS option¶
There is a third path between off-the-shelf SaaS and a fully custom build: take a mature open-source RAG platform, host it for the client, configure it for their corpus, and wrap it in a managed service. This sits between "buy Copilot" and "build from scratch" on both cost and lead time. See the companion document FOSS RAG platform shortlist for the current candidates we have evaluated and which one to reach for in which situation.