# Drafted, cited answers inside your helpdesk

> A support knowledge assistant is a retrieval-augmented generation (RAG) system that searches your help docs, policies and past tickets, then drafts an answer with citations inside your helpdesk for an agent to check and send. Once the drafts prove accurate, selected topics can be answered for customers directly. The Method Crew builds it in 6 to 12 weeks for $20,000 to $60,000.

- Page: https://methodcrew.com/blueprints/support-knowledge-assistant
- Last updated: 2026-10-28
- Publisher: [The Method Crew](https://methodcrew.com/)
- Blueprint: Support knowledge assistant

Your support team gets a drafted reply on every ticket, written from your help center, policies and past tickets, with a citation for each claim. Customers can be answered directly once the drafts earn it.

## Key facts

- Build time: 6 to 12 weeks
- Price: $20,000 to $60,000 per project. Agent-assist in one helpdesk sits in the lower half. Infrastructure and model usage passed through at cost.
- Typical stack: Zendesk, Intercom, Freshdesk or Help Scout, Confluence, Notion, Google Drive or SharePoint, OpenAI or Anthropic API, OpenAI or Cohere embeddings, Postgres with pgvector, Pinecone or Qdrant, Cohere Rerank, Microsoft Presidio for redaction, Langfuse for tracing and evaluations
- Runs on your accounts; model and tool usage billed at cost

## How it runs, step by step

1. **Sync the sources.** Help center articles, internal policies, macros and resolved tickets are pulled on a schedule from your helpdesk and from Confluence, Notion, Google Drive or SharePoint, each with an owner and a last-updated date.
2. **Redact past tickets.** Names, emails, phone numbers, order numbers and payment details are removed with Microsoft Presidio and patterns for your own ID formats before anything is stored.
3. **Chunk and index.** Documents are split along their own headings and tagged with product, audience (internal or public) and version, then indexed for both keyword and embedding search.
4. **Retrieve for the ticket.** When a ticket opens, the assistant builds a query from the subject, the message and the customer's plan, retrieves candidate passages and reranks them.
5. **Draft a cited answer.** The model writes a reply using only the retrieved passages, with a citation on each claim. If the sources do not cover the question, it says so and names the escalation path.
6. **Agent reviews and sends.** The draft appears in a sidebar next to the ticket. The agent inserts, edits or discards it, and can flag a wrong source with one click.
7. **Answer customers directly.** Optional. Once agent-assist meets the bar you set, approved topics are answered in chat or email, with a person one request away.
8. **Evaluate and update.** Every prompt, model or index change runs against a fixed set of real questions first. A weekly report lists accepted drafts, flagged sources and gaps in the docs.

## What the system is made of

- **Source connectors:** Scheduled sync from helpdesk, wiki and drive that keeps update dates and access tags.
- **Redaction step:** Strips personal data from tickets before they are stored or embedded.
- **Search index:** Postgres with pgvector, Pinecone or Qdrant, plus a keyword index for hybrid search.
- **Retriever and reranker:** Finds candidate passages, filters them by audience and product, and orders them by relevance.
- **Answer prompt:** Rules for tone, citing, declining and escalating, versioned alongside the code.
- **Helpdesk sidebar app:** Shows the draft and its sources inside Zendesk, Intercom or Freshdesk.
- **Evaluation set:** 150 to 300 real questions with approved answers, run before every change.

## What could go wrong and what stops it

- **Answers without a source.** Every claim needs a citation. A second check rejects drafts whose cited passage does not support the sentence. When retrieval finds nothing relevant, the assistant says so and suggests who to escalate to.
- **Internal content reaching customers.** Every chunk is tagged internal or public. In customer-facing mode the search query itself is filtered to public chunks, so internal notes cannot be retrieved, whatever the prompt says.
- **Outdated policies.** Superseded versions are removed on each sync, and drafts show the date of the source they cite. A weekly list of articles behind flagged drafts goes to each article's owner.
- **Customers trying to steer it.** Customer messages are treated as text to answer. The assistant has no tools that change accounts, issue refunds or edit orders. Requests like these go to a person with the conversation attached.

## How does the assistant keep answers grounded in your documents?

It is a [retrieval-augmented generation](https://methodcrew.com/glossary/retrieval-augmented-generation) system: it looks up the relevant passages first, then writes from those passages only. The weak point in most RAG systems is retrieval, so that is where the build spends its time.

Documents are split along their own headings, so each chunk answers something complete, and each carries its product, audience, version and last-updated date. Retrieval uses [hybrid search](https://methodcrew.com/glossary/hybrid-search). Keyword matching catches exact terms such as error codes and plan names, embeddings catch questions phrased differently from the docs, and a reranker orders the top candidates.

Past tickets need different handling. They show how your team phrases things and which questions arrive together, but they also hold one-off exceptions and promises nobody should repeat. So tickets guide tone and point to the right article, while policy answers must cite a help article or policy document.

Each draft cites every claim, and a second check confirms that each cited passage supports its sentence. If nothing supports an answer, the assistant says it could not find one. For the longer version, read [how to build a RAG knowledge assistant](https://methodcrew.com/insights/build-rag-knowledge-assistant), or see how we scope [RAG knowledge systems](https://methodcrew.com/services/rag-knowledge-systems).

## What does a support knowledge assistant cost to build and run?

The build is a fixed price between $20,000 and $60,000, and it takes 6 to 12 weeks. An agent-assist sidebar in one helpdesk, drawing on the help center and one internal wiki, sits in the lower half and the shorter end of that range. Customer-facing answers, several products or languages, permission-aware internal content and ticket history push it toward the top.

It starts above a single-source [RAG build](https://methodcrew.com/services/rag-knowledge-systems) ($15,000 to $35,000, 5 to 7 weeks) because it always reads at least 2 sources, redacts past tickets and adds a sidebar app inside the helpdesk. Per-user permissions across several sources follow that service's multi-source timeline of 8 to 12 weeks.

Running costs are billed at cost on your accounts:

| Item | What drives it | Notes |
|---|---|---|
| Model usage | Tickets drafted | Usually cents per draft |
| Embeddings | Corpus size and sync frequency | Small after the first full index |
| Vector database | Number of chunks | No license cost with pgvector on your Postgres; paid tiers on Pinecone or Qdrant Cloud |
| Tracing and evaluations | Trace volume | Langfuse, self-hosted or cloud |
| Monitoring (optional) | Review and tuning | $1,500 to $5,000 a month |

For comparison, Intercom prices its Fin AI agent at $0.99 per outcome, with a monthly minimum when used alongside another helpdesk ([Intercom pricing](https://www.intercom.com/pricing), checked October 2026). At low volume that is cheaper than a build. A custom assistant makes sense when volume is high, when answers depend on sources outside the helpdesk, or when you need control over citations and access. [AI ops monitoring](https://methodcrew.com/services/ai-ops-monitoring) covers the ongoing evaluation work.

## When should you not build a support assistant?

If your help center is thin or out of date, the assistant will cite wrong articles with confidence. Fix the docs first. A useful test: take 50 recent tickets and check whether a current article answers each one. If fewer than half are covered, the first project is documentation, and [content operations](https://methodcrew.com/services/content-operations) can speed that up.

If your tickets are mostly account actions (refund this order, change my plan, reset my login), retrieval helps little. Those need an agent with scoped tools and permissions, built as a [custom AI agent](https://methodcrew.com/services/custom-ai-agents), or plain automations behind a form.

If you handle fewer than about 500 tickets a month and everything lives in one helpdesk, use the AI features your helpdesk already sells and revisit in a year. A build at this size rarely pays back against per-outcome pricing.

And if a wrong answer could cause real harm, such as medical guidance or regulated financial advice, keep the assistant internal, with a qualified person sending every reply. The assist mode is still worth having there: staff find the right policy faster, and every answer they send carries its source.

## Common questions about this build

### Will the support assistant make things up?

It can, which is why it is built to decline. Drafts are written only from retrieved passages, every claim carries a citation, and a second check rejects citations that do not support the sentence. When nothing relevant is found, it says so. In agent-assist mode staff review every draft, and we measure how often they change the facts.

### Which helpdesks does it work with?

It is designed for Zendesk, Intercom, Freshdesk and Help Scout, where the assistant appears as a sidebar app next to the ticket. Other helpdesks work if they support apps or let us read tickets and post internal notes by API. Knowledge sources can sit anywhere we can read by API, such as Confluence, Notion, Google Drive or SharePoint.

### Can it learn from past tickets that contain customer data?

Yes, after redaction. Names, emails, phone numbers, order numbers and payment details are removed before tickets are indexed, using Microsoft Presidio plus patterns for your own ID formats. Tickets older than your retention period are excluded. The index lives on your accounts, and OpenAI and Anthropic do not train on API data by default.

### How do you know the assistant is working?

We report weekly on how often staff send a draft with light or no edits, how often they flag a wrong source, how often the assistant finds nothing, and handle time on assisted tickets. Before launch and after every change, it also runs against 150 to 300 real questions with approved answers. If accuracy drops, the change does not ship.

### Can customers ask it questions directly?

Yes, once agent-assist has proven itself topic by topic. We open a topic to customers in chat or email after staff have accepted most drafts on it for 4 weeks. Anything outside approved topics, or below the confidence threshold, goes to a person with the conversation attached. Customers can ask for a person at any point.

### Is RAG better than fine-tuning for support answers?

For support, almost always. Docs and policies change weekly. RAG picks up a change on the next sync, while a fine-tuned model has to be retrained. RAG also shows its source, so staff can check an answer in seconds. Fine-tuning can help with tone or format, which a good prompt and a few examples usually cover.

## Services that build and run it

- [RAG and knowledge systems](https://methodcrew.com/services/rag-knowledge-systems): Your documents turned into an assistant that cites its sources. $15,000 to $75,000 per project.
- [Custom AI agents and copilots](https://methodcrew.com/services/custom-ai-agents): Agents that look things up in your CRM or helpdesk and take the next step, with approvals. $15,000 to $90,000 per project.
- [AI ops monitoring and improvement](https://methodcrew.com/services/ai-ops-monitoring): Quality checks, cost control and model updates for live AI systems. $1,500 to $5,000 per month.

## Sectors where it fits

- [SaaS and technology](https://methodcrew.com/industries/saas-technology): When a buyer asks ChatGPT or Perplexity for a shortlist in your category, the answer names a handful of products. We work to get yours named accurately, then build the site and the support assistant those buyers land on.
- [Ecommerce and retail](https://methodcrew.com/industries/ecommerce-retail): If your support queue is mostly people asking where an order is, and your catalog was written in a hurry before launch, both are fixable with systems that read your order data and product data directly.
- [Insurance and financial services](https://methodcrew.com/industries/financial-services): Applications, statements and policy documents arrive as PDFs and someone retypes them. Renewals depend on a CRM nobody has time to update. We build systems that read the documents, keep the CRM current and log every AI-assisted step for your compliance team.

## Next step

Request a 30-minute scoping call: https://methodcrew.com/contact. Or email contact@methodcrew.com.