Skip to content

Why every agent draft cites its sources

A draft without sources asks you to trust it. A draft with sources lets you check it in seconds. Why citations are on every Mokanata draft, and how to make an agent say "I don't know".

Product team3 min read

Every draft a Mokanata agent writes carries the sources it used. In the approval queue they sit under the draft. In the inbox, Why this draft? opens the steps the agent took and the documents it relied on. In the agent's runs, the run drawer lists the sources cited next to the rules that were checked.

This post explains why we made citations a default rather than a setting, and how to tune them.

The reviewer's problem

Approval mode only works if reviewing a draft is faster than writing it. Without sources, a reviewer has two options: trust the draft, or go and check every claim by hand. The first is risky and the second is slow, so in practice people do the first and hope.

A cited draft changes the question from "is this right?" to "does this source say this?". That's a question you can answer in a few seconds, by glancing at the snippet the agent quoted.

What counts as a source

An agent can cite anything in its knowledge: help-center articles, pricing sheets, policy documents, product specs, and the answers you write to fill coverage gaps. Each source shows its document count and whether its sync is fresh or stale, and you can sync it on demand.

Sources are configured per agent, so a billing agent can be grounded in invoices and refund policy while a closer works from pricing and case material.

Note

Stale sources are a common cause of wrong answers that look right. If a price changed last week and the pricing sheet hasn't synced since, the draft will cite it faithfully and still be out of date. Check the sync state when a citation surprises you.

Making citations mandatory

There are three places to turn "should cite" into "must cite":

  1. Knowledge settings. Require a citation for every claim. A draft that makes a factual claim without a source can't go out unreviewed.
  2. Citation policy in the wizard. Show source links to the reader, and tell the agent to say "I don't know" rather than guess.
  3. A guardrail step in a flow. The citation required guardrail, or a guardrail gate grounded in knowledge sources with an inline-citation requirement, stops a flow that would send an unsupported claim.

The second one deserves a closer look.

Teaching an agent to say "I don't know"

The most expensive thing an agent can do is answer confidently with something it made up. The "I don't know" policy gives it a better option. When nothing in its knowledge covers the question, the agent says so and hands off, instead of filling the gap with a plausible guess.

That handoff is useful data. Every question the agent couldn't answer is a candidate for the coverage gaps list, where you can write the missing answer yourself or let AI draft it for you to approve. Over a few weeks the gaps close and the handoffs drop, for the right reason.

How citations interact with confidence

Confidence and citations are separate signals, and the useful cases are where they disagree:

ConfidenceCitationsWhat it usually means
HighStrongRoutine answer, a good candidate for auto-approve
HighWeak or noneThe agent is sure but can't show why: review it
LowStrongThe facts are there but the situation is unusual
LowNoneA coverage gap: write the answer once

Drafts below 0.80 confidence land in the Low confidence tab of the approval queue regardless of their sources, so the bottom two rows always reach a person.

Citations after the fact

Sources don't disappear once a message is sent. Every action is written to the audit log with the sources it used, so a month later you can answer "why did we tell this customer that?" with the exact document, not a reconstruction.

That matters most when something goes wrong. A wrong answer with a citation points straight at the document to fix. A wrong answer without one leaves you guessing whether the problem was the knowledge, the prompt or the model.

In short

  • Every draft shows the sources it used, in the queue, the inbox and the agent's runs.
  • Require a citation for every claim, and prefer "I don't know" to a guess.
  • Treat handoffs as a list of gaps to fill, not as failures.
  • Keep sources fresh, because a faithful citation of an old document is still wrong.

Keep reading

All posts
  • Product

    Introducing the auto-send countdown

    Drafts released by an auto-approve rule with a review window now show "Auto-sends in mm:ss" until they go. Stop one, edit it, or switch auto-send off for the account, from wherever you are.

    3 min read

  • Playbooks

    Running WhatsApp sales inside the 24-hour window

    WhatsApp lets a business reply freely for 24 hours after a customer's last message. After that, only approved templates. A playbook for selling within that rhythm instead of fighting it.

    3 min read