Surface Relevant Help Center Articles During Chat

The fix is to bring the search to the agent instead of sending the agent to the search. While the agent is working a ticket, an AI agent reads the conversation, searches your knowledge, and hands back the most relevant articles ranked by match — with an offer to drop the steps straight into the reply. That's the use case this post walks through, how to set it up with Macha, and — honestly — where it overlaps with what your helpdesk already does.

What "surface articles during chat" actually means

There are two ways an agent gets knowledge mid-conversation, and they're worth separating:

• Passive suggestions — a panel quietly lists articles it thinks are relevant to the open ticket. The agent glances over and picks one.
• Active retrieval — the agent asks ("any articles on 2FA setup?") and gets a ranked, scored answer back, with the option to act on it.

Macha does the second, and it does it as a configurable agent — a named worker with instructions and tools, not a fixed widget. In the workflow we're describing, the agent's job is narrow and useful: take the question in front of you, search the connected knowledge, and return the best matches with a relevance score so you can trust the ranking at a glance.

Here's the shape of the interaction. An agent, working ticket #8301, types into Macha's chat:

"The customer wants to know how to set up two-factor authentication. Any relevant articles?"

Macha searches the Help Center and replies:

Here are the most relevant Help Center articles: 1. "Setting Up Two-Factor Authentication" (98% match) — covers setup steps for SMS and authenticator-app methods. 2. "Account Security Best Practices" (82% match) — includes 2FA plus password requirements and session management. Want me to include the setup steps in your reply, or share the article link directly?

The relevance score is the part that earns trust: a 98% match means use this one; an 82% match is the "related, but read it first" backup. And the closing question is the part that saves the most time — the agent never has to leave the conversation to fetch the link or retype the steps.

The honest part: your helpdesk already does some of this

It would be dishonest to pitch "surface articles for agents" as something only Macha can do. It isn't. Zendesk ships a Knowledge section in the context panel that auto-searches and suggests Help Center and Community content based on the ticket's brand and language, lets agents link, quote, or flag an article in one click, and even request a new one to fill a gap. (External content and AI quick answers are gated to higher plans, per Zendesk's own docs.) Genesys, Freshdesk, and most modern suites have a version of the same thing.

So if all you need is "suggest Zendesk articles inside the Zendesk agent workspace," turn on the native feature first — it's included, it's well-integrated, and you don't need us for it. Macha is an AI agent layer on top of your helpdesk, not a replacement for it; we'd rather you use the native panel than bolt on a tool you don't need.

Macha earns its place when one of three things is true:

1. Your knowledge lives in more than one place. The answer isn't always in the Help Center. It's split across the public Help Center, an internal Notion wiki, a Confluence space, and a folder of policy docs in Google Drive. The native panel searches one Help Center. A Macha agent searches all of them in a single ranked answer.
2. **You want the agent to act, not just list.** "Want me to include the setup steps in your reply?" isn't something a suggestions panel offers. Because Macha's agent has instructions and tools, it can draft the response, paste the steps, or pull the live link — not just show you a card.
3. The question shows up outside the helpdesk. The same agent that searches the Help Center for a ticket can answer a teammate in Slack who asks the same thing. (That's a sibling workflow — auto-suggesting knowledge base articles in Slack — built on the same retrieval.)

How to build it in Macha

The whole thing is one agent with a tight set of tools. Here's the configuration.

1. Connect your knowledge

Connect Zendesk so the agent can read the ticket and search your Help Center, then add any other source the answer might live in. Macha treats each connected source as searchable knowledge: the Get Articles action pulls from your Zendesk Help Center, and connectors for Notion, Confluence, Google Drive, and uploaded files extend the same search across your internal docs. The more honestly your knowledge is connected, the better the ranked answer.

2. Give the agent its tools and instructions

Create an agent — call it @knowledgeFinder — and enable just the actions it needs: Get Ticket (so it can read the conversation it's helping with) and Get Articles (so it can search the Help Center). The instructions stay short and specific:

You help support agents find answers. When asked about a ticket or a question, search the connected knowledge sources, return the 2–3 most relevant articles ranked by match percentage, and summarise what each covers in one line. Offer to insert the key steps into the agent's reply or share the link. Never invent an article — if nothing scores well, say so.

That last sentence matters more than it looks: a retrieval agent that admits "I don't have an article for this" is far more useful than one that confidently surfaces a 40%-match near-miss.

3. Decide how it runs: on-demand or on-trigger

There are two patterns, and most teams use both:

• On-demand (chat): the agent sits in Macha's chat. An agent @-mentions it with a question or a ticket number and gets the ranked answer back instantly. This is the fastest to adopt — no automation, no risk, just a faster search box.
• On-trigger (proactive): set a Ticket Updated trigger so that when an agent starts responding to a ticket, the agent quietly drops a ranked list of relevant articles into an internal note. The agent opens the ticket and the suggestions are already there.

What it costs

Macha is credit-based, and credits are spent per AI action, not per ticket or per "deflection." Surfacing articles is a cheap action: a single search-and-rank on the default GPT-5.4 Mini model runs at 1 credit (models range 0.5–9 credits depending on which you pick). An agent who triggers ten lookups a day is spending pennies of your allowance, and you can put the whole thing on a 7-day free trial, no credit card required to watch the real consumption before committing. See the pricing page for current plans and credit allowances — we don't hard-code numbers in blog posts because they change.

The thing to weigh isn't the credit cost; it's whether you'd rather pay a credit per lookup or pay an agent 45 seconds per lookup. For most teams the math is not close.

Watch-outs — when not to reach for this

A few honest caveats:

• If you only use Zendesk Help Center and live entirely in the Zendesk agent workspace, use the native Knowledge panel. It's free with your plan and tightly integrated. Adding Macha here buys you nothing unless you have knowledge outside Zendesk or want the agent to take actions.
• Garbage in, garbage out. A retrieval agent is only as good as the articles it searches. If your Help Center is thin, stale, or contradictory, the agent will faithfully surface stale, contradictory answers. Fix the knowledge first; a good Studies pass over resolved tickets can show you which articles you're missing.
• Don't let it auto-send. This use case is agent-assist, not auto-reply — a human still writes the response. If you want the AI to answer the customer directly from the knowledge base, that's a different, more autonomous workflow (auto-reply using knowledge base articles), and it deserves its own guardrails and review.
• Relevance scores are a guide, not gospel. A 98% match is usually the right article, but the agent should still read it before pasting. Treat the score as triage, not proof.

Where this fits in a bigger setup

"Surface articles during chat" is usually the first, safest agent a team turns on, because it can't do any harm — it only ever suggests to a human. Once your agents trust the retrieval, the natural next steps are letting the same knowledge power autonomous first replies for the truly repetitive questions, routing the rest, and running analysis over your ticket history to find the articles you should have written but never did. It all runs on the same connected knowledge and the same Zendesk integration — the article-finder is just the gentlest place to start.

FAQ

Does this replace Zendesk's native article suggestions? No — and you shouldn't make it. If your knowledge is all in one Zendesk Help Center and your agents work inside the Zendesk agent workspace, the built-in Knowledge context panel is the right tool. Macha adds value when you need to search across Zendesk Help Center plus Notion, Confluence, Google Drive and internal docs in one ranked answer, or when you want the agent to draft the reply rather than just list links.

Which knowledge sources can the agent search? Your Zendesk Help Center via the Get Articles action, plus any connected source — Notion, Confluence, Google Drive, and uploaded files. All of it is searched together and returned as a single ranked list.

Will it send replies to customers automatically? Not in this workflow. This is agent-assist: a human always writes the response. Autonomous, customer-facing replies are a separate, more guarded use case.

What does the match percentage mean? It's a relevance score from the search ranking — a quick signal of how closely an article fits the question. Higher is a stronger match; it's there to triage, not to replace the agent reading the article.

How much does it cost to run? Credits per AI action — a single search-and-rank is 1 credit on the default model (0.5–9 by model). Start a 7-day free trial, no credit card required to see real usage; see pricing for plan details.

Try it

If your agents are still alt-tabbing to your Help Center mid-conversation, this is the fastest, lowest-risk agent to put in front of them. Start a 7-day free trial, no credit card required, connect Zendesk and your other knowledge sources, and build a @knowledgeFinder agent in a few minutes — or read the docs for the full setup.

Written by Abbas (Customer Support & AI, Macha) · Reviewed by Ankeet Guha (Co-founder & CTO) · Published 2026-06-24 · Last updated 2026-06-24.