Slack
Integration · Messaging and voice
Putting an AI agent to work inside Slack
Slack works best as the place your team talks to an agent: asking it questions, approving what it has prepared and receiving alerts when something needs a person. The agent runs as a Slack app with only the permissions you grant, and it does the real work in your other systems.
Slack is the front desk, not the filing cabinet
Your customers, invoices and bookings live elsewhere. What Slack offers is the place your team already looks all day. That makes it a good interface for an agent: someone types a question or clicks a button in a channel, and the agent does the work in your CRM, accounts or database and reports back.
Three patterns cover most useful Slack agents:
- Ask: “What did we quote the Murphy job?” The agent searches the approved sources and answers in a thread, with links. See internal knowledge assistant.
- Approve: the agent posts a drafted reply, refund or invoice with Approve and Edit buttons. Nothing goes out until a person clicks. See human-in-the-loop AI.
- Alert: the agent posts when something needs attention, such as an urgent email, a failed payment or a booking clash, with the context already summarised. See email triage.
An example approval flow
For example, a hypothetical accountancy practice receives supplier invoices by email. The agent reads each one, prepares the entry and posts a message in #invoices: supplier, amount, VAT, suggested code and a link to the PDF. A bookkeeper clicks Approve, and the agent records it in the accounts system and replies in the thread to confirm. If the bookkeeper clicks Edit, a short form opens, and the correction is saved so the agent can learn which codes the practice prefers.
The value here is less about AI and more about putting the decision in front of the right person, at the right moment, with everything they need on one screen. See invoice processing for the rest of that workflow.
What does setting it up involve?
- Create a Slack app for your workspace and choose its permission scopes. A bot token, which belongs to the app rather than a person, is usually the right choice, so the agent keeps working if the person who installed it leaves.
- Decide which channels the agent joins. It only sees channels it has been added to and messages its scopes allow.
- Set up how Slack tells the agent something happened, such as a mention, a button click or a message in a channel, and how the agent responds.
- Have a workspace admin approve the installation, in line with your workspace’s app policy.
The agent itself runs on your own hosting, not inside Slack. Slack passes it events and displays its replies; the agent holds the connections to your other systems and keeps its own log of every request, decision and approval. That log matters, because Slack messages can be edited or deleted and your workspace’s retention settings may remove history you later need. If a request from Slack fails part-way, for example the accounts system is down, the agent should say so in the thread rather than leave the team guessing.
Before you connect Slack: a checklist.
Connecting Slack safely
0 of 8 done
Limits and gotchas
- Slack history is not a reliable knowledge base. Old threads contradict each other. Point the agent at maintained documents for answers and use Slack for conversation.
- Reading history is rate limited, and Slack has tightened limits for some distributed apps that are not listed in its Marketplace. Apps built internally for your own workspace were not affected by that change, but bulk history reads still need pacing.
- Private channels and direct messages are sensitive. Do not grant scopes to read them unless a specific task requires it.
- Noise: an agent that posts too often gets muted. Agree what deserves a message and batch the rest into a daily summary.
When is a simple Slack notification enough?
If all you need is “post a message when a form is submitted”, an incoming webhook or a Zapier step does it in minutes and needs no AI at all. Slack also has its own built-in AI features on some plans, which may cover summarising channels. A custom agent is worth building when a message needs to lead to an action in another system, with permissions and an audit trail, or when the same agent also serves customers on WhatsApp or email.
The app and its tokens should be registered to your workspace and stored in your own secrets manager, not pasted into a script. Revoking the app in Slack’s admin settings should stop the agent immediately.
Common questions
Can the agent read all our Slack messages?
Only if you give it the permissions to and add it to those channels, and you should not. Grant the narrowest scopes that support the task, keep it out of private channels and direct messages unless needed, and tell your team where it is listening. Workspace admins can see and remove installed apps at any time.
Does this work with Microsoft Teams instead?
The same patterns (ask, approve, alert) work in Teams, but the setup is different and runs through Microsoft’s platform and your Microsoft 365 tenant. If your business runs on Microsoft 365, it usually makes more sense to put the agent there than to add Slack just for it.
Who is responsible when someone clicks Approve?
The person who approved it, which is why the approval message must show everything they need to decide: the amount, the recipient, the source document and what will happen next. The agent should record who approved what and when, so you have an audit trail if a decision is questioned later.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.