Skip to content
KEIKI

Slack agents

Your agent, on Slack

AI agents for Slack

The answer is already in the workspace.

Every team answers the same internal questions forever — expenses, access, deploys, policy. Your agent answers them in the channel from what your team already wrote, links the source, files the request when one is needed, and says "I do not know" instead of inventing.

TL;DR

What you get

  • The questions your team asks weekly get answered in-channel, with the source link.
  • When the workspace does not contain the answer, it says so and routes to an owner.
  • Access requests and tickets get filed in the thread, not linked to a form.
  • Replies stay threaded, so channels do not fill with bot messages.
  • Setup is about 20 minutes: Keiki generates the Slack app manifest for you.

What it handles

Six outcomes, in the channel where the question got asked.

01

Verifiable answers

Every answer carries the doc or thread it came from, so the person asking can check it rather than trust a bot.

02

No confident wrong answers

When the workspace does not contain it, the agent says so and routes the question to an owner instead of returning the nearest document.

03

Requests actually filed

Access requests, laptop swaps and vendor forms are collected in the thread and filed, not linked to a form nobody opens.

04

Readable channels

Replies stay threaded on the question, and the agent only speaks when mentioned or asked something it can answer.

05

New hires unblocked in week one

The twenty questions every joiner asks get answered immediately, privately, without them feeling like an interruption.

06

One answer, every channel

The same reply in #eng, #support and a DM, because it comes from one source instead of five people remembering differently.

A real thread

It answers where the question was asked.

Nobody opens a knowledge base voluntarily. The question gets asked in the channel it was always going to be asked in, and the answer arrives there with its source attached.

Slack · #ask-it
  1. Customer:do we reimburse a coworking day if the office is closed?
  2. Agent:Yes — up to €35/day when the local office is closed, expensed under "Workplace". Source: Expense Policy v4, §3.2 (updated Aug 2026).
  3. Customer:and for contractors?
  4. Agent:The policy only covers employees. I could not find anything about contractors in the workspace, so I have not guessed — I have flagged this for @people-ops in this thread.
  5. Customer:thanks. also i need access to the billing dashboard
  6. Agent:Filed IT-2841 for billing-dashboard read access with your manager as approver. You will get the approval prompt here; nothing else needed from you.

A Slack agent vs a search-the-wiki bot

Search returns ten documents. A colleague answers.

The usual internal bot is a search box in a Slack skin: it returns links ranked by keyword and leaves the reading to you. That is why people ask a person instead.

Your Slack agent compared with Wiki search bot
CapabilityYour Slack agentWiki search bot
What comes backThe answer, plus the source link.Ten results, sorted by keyword.
When it does not knowSays so, and routes the question to an owner.Returns the closest-matching document anyway.
Follow-upsKeeps the thread and answers in context.Each query is independent.
Doing the thingFiles the request in the same thread.Links to a form.
Stale contentCites the document and its date, so you can see the age.Ranks the 2021 page first.
Where it livesIn the channel, threaded on the question.In a modal you had to open.

Start from a complete template

Company Brain

The complete template for internal questions: answers from what the team already wrote, with the link, and the ticket filed while the agent is already in the thread.

Use this template

What comes with it

  • Answers grounded in your own docs and past threads, with citations
  • An explicit "I do not know", with routing to the owner
  • Ticket and request filing from inside the thread
  • Threaded replies that keep channels readable
  • Conversation memory for follow-up questions
  • Handover to a human with the sources and the promise attached

Evidence

Why answering in-channel is worth measuring.

Published research on AI assistance at work and on the cost of interruption. None of these studies tested Keiki — they measure the behaviour an internal agent is built on.

  1. AI assistance raises throughput most for the least experienced.

    Support staff using a generative-AI assistant resolved nearly 14% more issues per hour, with the largest gains among novice and low-skill workers.

    Generative AI at WorkBrynjolfsson, Li & Raymond, NBER Working Paper 31161, 2023

  2. Assistance also cuts the time a task takes.

    In a preregistered experiment with 453 professionals, ChatGPT reduced time spent on mid-level writing tasks and raised output quality as rated by blinded evaluators.

    Experimental evidence on the productivity effects of generative artificial intelligenceNoy & Zhang, Science, 2023

  3. The cost of asking a colleague is the interruption.

    A study of information workers found that time spent handling messages tracks closely with self-reported stress and fragmented attention.

    Email Duration, Batching and Self-interruptionMark et al., CHI 2016

Further reading

Worth reading on Slack apps.

Questions

What people ask about running an agent in Slack.

Where do its answers come from?
From the sources you connect — docs, handbooks, past threads. Every answer cites what it used, so the person asking can verify it instead of trusting a bot.
What does it do when the answer is not there?
It says so plainly and routes the question to an owner in the same thread. It does not return the nearest-looking document as if it were an answer.
Will it clutter our channels?
It replies in threads on the message that asked, and only when it is addressed or asked something it can answer. Channels stay readable.
Can it actually do things, not just answer?
Yes — it collects what a request needs in the thread and files it, so access requests and internal tickets do not depend on someone opening a form later.
How does it handle information people should not see?
It only answers from the sources you connect it to, and it is meant to be scoped to what the channel it lives in should know. Sensitive requests are routed rather than answered.
How long does setup take?
Start from the Company Brain template, connect Slack and your first source, and put it in one channel. Expand once it has earned it.

Stop answering the same question.

Request access at onkeiki.com, open the Company Brain template, install the Slack app Keiki generates. About 20 minutes to a channel that answers itself.