An AI colleague that lives in Slack shows up in your member list with a name, answers inside the thread where you tagged it, and still knows next month what your team decided this week.
The good ones also do the task. They pull the numbers, open the ticket, or draft the reply, then post the result where everyone can check it, and eight tools fit that description in 2026.
Short version first:
If you want an AI colleague that lives in Slack, hire Lindy. It's the best one in 2026 because it answers under your message, keeps know-how in memory files you can correct, and acts in 1,500+ connected apps or any MCP server, asking the shared thread before it writes.
Viktor is the cross-tool generalist, Atlas by WorkOS the promising newcomer, and Dust a workshop for specialist agents. Claude Tag acts as a channel-wide analyst, while ChatGPT workspace agents are custom-built specialists.
Slackbot is the built-in colleague that comes with Business+, and Salesforce Agentforce in Slack brings department agents to Salesforce shops.
Four abilities separate a colleague from a clever search box, and every tool below handles them a little differently.
Show up where the work is. It answers in the thread where someone asked, so the context stays attached to the conversation and nobody has to open a separate app to find the reply.
Remember. The second time you explain your discount rules should be the last time, which means memory that persists and that someone on your team can read and correct.
Act. A colleague opens the ticket, updates the CRM record, and drafts the email using the tools your team already pays for.
Ask before acting. Anything that writes to another system or reaches a customer should wait for a human click until you trust it (you wouldn't hand a new hire the customer inbox on day two, either).
These three take a request from a single message and come back with finished work, which suits teams that want fewer tabs open.
This is the colleague to hire when everyone should share one assistant, from the ops lead who wants a weekly report to the account manager who needs tomorrow's call prepped.
What kind of colleague it is: The teammate everyone shares, and onboarding it takes one admin a few minutes in Slack, without API keys or a ticket to IT. Coworkers connect their own accounts with a single tap on their first DM or @mention, and their first message replays after linking.
What it's good at: In a channel it replies inside the thread, and after that first reply it can keep up with the conversation without another @mention. Because it writes down what it learns in plain, editable memory files, your renewal process stops living in one person's head.
Recorded calls end up in a meeting library the team shares, so a single attendee is enough and everyone else can catch up when it suits them.
It also ships with 40+ built-in skills (research, data analysis, slide decks, and dashboards among them), and Watch and Learn saves anything you show it once for the whole team.
Its reach covers 1,500+ integrations and any MCP server, and for an app with no connector it can click through a browser itself, like you would.
You set one of three rules per integration: Don't offer, Require approval, or Always allow. When it wants to write something from a shared thread, it posts Approve and Deny buttons, and the person who asked or a workspace admin gets to press them.
Cost: On Free you pay $0, get $50 in credits, and have 7 days to spend them, card not required. Team bills $49.99/month for every active user, folds each person's 5,000 credits into a team-wide pool, and sells more at $10 per 1,000 credits.
Enterprise is custom, with SSO and SCIM.
Where it struggles: Approval buttons only appear in shared threads. A DM with it runs without that check, so agree as a team on what people may ask it to do in private.
Shared-channel answers stay public-safe, too, which narrows its tools there to Notion (if connected), public Slack search, and the web. Anything involving your calendar or inbox moves to a DM or a thread you started, and that split can feel odd at first.
Viktor is the strongest pick for teams that want one shared AI employee turning out reports, dashboards, code, and campaigns from a single Slack message.
What kind of colleague it is: Viktor bills itself as "the AI employee in Slack and Teams," which matters if part of your company works in Microsoft Teams. You @mention it in a channel or DM, it runs the task on its own cloud computer, and it replies with finished work.
What it's good at: Cross-tool jobs. A single request can send it hopping between Linear, Stripe, Google Drive, HubSpot, Notion, and GitHub, and Viktor counts more than 9,800 integrations in all.
It remembers decisions and preferences, runs work on a schedule, and, in the vendor's words, "asks before anything he cannot undo." On security, Viktor says it's SOC 2 compliant and has passed Slack's security review.
Cost: As of October 2026, Viktor's Free plan asks for $0 and no card, and its credits (up to $100) never expire. On Team, the whole workspace splits 40,000 credits for a flat $100/month, nobody pays per seat, and anything left at month's end carries over for one more month.
For a heavy month you can buy top-ups or switch on auto-top-ups capped at a monthly spend limit. Enterprise is custom, and nonprofits get 15% off.
Where it struggles: Spend is hard to predict because tasks vary so much. Viktor's credit guide says a deep report can eat 500 to 2,000 credits and a fast Slack question 50 to 100, so 40,000 credits stretch anywhere from about 80 to 400 tasks.
Every recurring job it runs for you bills the same pool, so a daily report nobody reads anymore still costs you. A fair trial runs your real weekly workload for two weeks while someone watches the meter.
What kind of colleague it is: WorkOS launched Atlas on August 4, 2026, as "your AI coworker," and opened it to anyone from launch day. It's ready in any DM, channel, or thread, and it does its work out in the open in shared channels.
What it's good at: Atlas picks up your team's shorthand from conversations, and in WorkOS's words, "Correct it once, and the correction sticks." It's useful with Slack alone, so you can wait to connect other tools from a catalog that runs into the hundreds.
You can also set up custom team agents, each with its own @-name, instructions, skills, memory, and integrations, plus shared skills and recurring work. Permissions are scoped per agent across files, memory, integrations, and workflows.
Because it's built on WorkOS identity infrastructure, "Every request respects the permissions your organization already enforces." For a security team that already trusts its identity setup, that's a reassuring starting point.
Cost: WorkOS hasn't published pricing on its Atlas pages as of October 2026, so ask for numbers before it spreads past one team.
Where it struggles: It's the newest product here, launched in August 2026, so you're adopting an early version with no public price to budget against. WorkOS describes it as a Slack product with no Teams version mentioned, which rules it out for a mixed workspace.
In a trial, I'd correct it on something small and check a week later whether the fix held, then look at how per-agent permissions appear to an admin.
These three shine when someone needs a knowledgeable reply or a specialist in the thread, and each leans on a platform your company may already pay for.
Dust is built for companies that want a bench of specialist agents for support, sales, and ops, each grounded in their own knowledge and tools.
What kind of colleague it is: A team of specialists behind one name, pitched as "Multiplayer AI for human-agent collaboration." Once an admin enables the Slack Bot (Admin > Integrations), anyone can DM it with no mention or tag @Dust inside a thread or channel.
What it's good at: Routing you to the right expert. Typing +AgentName summons one specialist per message, a channel can have a default agent that answers unprompted if an admin sets it, and Slack Workflow Builder can trigger agents as well.
Each specialist can run on one of 20+ models (DeepSeek, Mistral, Gemini, Claude, and GPT among them), pull from dozens of connectors and MCP servers, and team up with others in multi-agent workflows that fire on triggers or schedules.
Pods hand a group its own shared workspace, and Frames can turn an answer into an interactive app or dashboard.
Security reviewers will find SOC 2 Type II, single sign-on via Jumpcloud, Entra ID, or Okta, and a choice between EU and US data residency.
Cost: As of October 2026, Dust prices in euros, excluding VAT. Max runs €150/seat/month for 40,000 credits, Pro costs €30/seat/month (or €24 on annual billing) for 8,000, and the €0 Free tier hands out 500 lifetime credits.
Enterprise is custom and adds SCIM and audit logs.
Where it struggles: Every specialist has to be designed and kept up to date by someone on your side, so expect a stretch of build work before the bench produces anything useful.
Unused credits vanish at the end of each month, and because Dust bills in euros only, a US finance team has to plan around exchange rates.
What kind of colleague it is: Claude Tag arrived on June 23, 2026, in beta for Claude Team and Enterprise customers, replacing the company's earlier Slack app. Admins pick its channels and connect tools, data, and codebases, and anyone there can tag @Claude to hand off a task.
What it's good at: Working in public. A channel gets a single Claude that talks with everybody, so progress is visible to the room, and tagging it there needs no Claude account by default.
It remembers relevant information from its channels, though it doesn't report from private ones. Its memory is walled off per channel, so lessons from the sales room never bleed into engineering's.
Turn on ambient mode and it'll speak up with useful information unprompted, nudge threads that stalled, and put its own tasks on a schedule. A DM runs privately on your own claude.ai account with your personal tools and connectors.
For oversight, admins can set token-spend ceilings per channel or across the org, break spend down by channel, and read a log showing each thing @Claude did along with who requested it.
Cost: There's no separate price as of October 2026. Your org needs Enterprise or a Claude Team plan, where Premium seats are $125 and Standard seats $25/seat/month ($20 billed annually) for groups of 2 to 150, and whatever it does in channels is charged to the organization as usage.
Where it struggles: Beta status keeps it gated to Team and Enterprise plans. Usage lands on top of seats, so the bill grows with how often people tag it, which makes per-channel limits worth setting on the first day.
What kind of colleague it is: You build an agent in ChatGPT, then deploy it to Slack channels through the ChatGPT Agents app. Slack gives each one a handle of its own, tied to a Slack user group, so coworkers can tag it by name like anybody else.
What it's good at: Reading the room. Per channel, you decide between "Respond to all relevant messages" and "Only reply when @mentioned," and the agents keep following up in-thread and can join private channels after the app is added.
They can post results to a channel on a schedule and "send emails, update documents, and more without ever leaving Slack." They also stick to Slack permissions, so only messages and files you can already reach are searchable.
Behind the scenes, ChatGPT Business links up with Microsoft 365, GitHub, Slack, and Google Workspace, gives admins spend controls and usage analytics, and by default keeps business data out of training.
Cost: As of October 2026, a Standard seat on ChatGPT Business costs $20/user/month billed annually or $25 billed monthly, Premium costs $125 ($100 billed annually), and you need at least two users. Slack itself has to be on a paid plan, too.
Where it struggles: Onboarding has three moving parts, installing the ChatGPT app and the Slack app and then allowing members to create user groups, and Enterprise Grid adds org-level steps on top. Workspace agents aren't available on personal accounts.
OpenAI's own help page warns that "Responses from ChatGPT Agents may be inaccurate," which is a good reason to start in one low-stakes channel.
If your company already pays for Slack Business+ or Salesforce, these two are on staff before anyone signs a new contract.
Slackbot is the easiest AI colleague to bring on, because it's already part of Slack, with no install and no new vendor for security to review.
What kind of colleague it is: The personal AI agent Slack builds into its own pricing grid, which workspace or org owners and admins can enable or disable. You tag it into a channel to pull a thread together and share the summary back.
What it's good at: Research with receipts. Ask it something and it digs through the web, your connected apps, and Slack history, then hands back a cited report, and through MCP it can take action in Docusign, Salesforce, Figma, Jira, and more.
It builds repeatable workflows when you ask, creates dashboards, polls, and calculators right in the conversation (pinnable to a channel), and shares skills across the team. With Slack CRM it can log calls, update records, and open new accounts, and on mobile you can talk to it by voice.
Slack's privacy line is clear: it only surfaces what you already have access to, your interactions stay private, and your data never trains AI.
Cost: As of October 2026, a workspace on Free or Pro ($8.75/user/month) can only borrow Slackbot through a trial or preview. Business+ costs $18/user/month ($15 billed annually) and gives each member 15 Slackbot messages a week.
Enterprise+ is a sales conversation, with usage limited only by Slack's overall policies, and Salesforce Flex Credits can buy extra usage.
Where it struggles: Fifteen messages a week works out to about three per workday, which a heavy user can burn through well before the reset. The full agent also lives at the plan level, so getting it for one power user means Business+ for the whole workspace.
What kind of colleague it is: A purpose-built agent that your company creates in Salesforce and puts to work in Slack. Slack's help center describes agents "designed to fulfill a specific purpose, like responding to IT help requests or providing information about benefits."
What it's good at: Being easy to find. Agents sit under Tools > Agents in the Agents tab, where you can DM one using prompt starters and past threads, or bring it into a channel and @mention it "just like you would any other teammate."
They answer inside the limits of your own Salesforce and Slack permissions, so nothing comes back that you couldn't already see, and admins control who's assigned to which agent.
A link to an agent's prompt can go into a workflow, canvas, or channel, and the building happens in Agent Script, Agentforce Builder, and Prompt Builder.
Cost: As of October 2026, its meter runs on Flex Credits, sold at $500 per 100,000, and each customer-facing conversation is $2. The Agentforce User License adds $5/user/month on top of those credits.
Max Edition begins at $550/user/month, with add-ons at $125/user/month (or $150 on Industries).
Any Slack plan can host it once you hold an Agentforce license, though detailed numbers come from a sales rep per the rate card, with "Try for free" listed next to the builder tools.
Where it struggles: This colleague arrives without a job description, because each agent only exists once someone in-house builds and configures it.
Its value also rides on an Agentforce license and the Salesforce data behind it, which makes it a weaker fit for teams that keep customer records elsewhere.
Write the job description before you watch a single demo, because these tools sort themselves fast once you know whether you need someone who answers or someone who does.
If the job is "take requests from anyone and finish them across your tools," start with the first group and compare how each bills: per active user, a shared credit pool, or a price you'll have to ask for.
If the job is expert answers from company knowledge, the tag-in experts fit, and the tiebreaker is usually which AI platform your company already pays for.
If budget approval is the hard part, look at who's already on payroll: Slackbot comes with Business+, and Agentforce makes sense once Salesforce holds your customer data.
Then give your top two the same three tasks for a week. Watch memory closely: correct each one on a small detail and see whether the correction holds the next time someone else asks.
Last, settle the admin question before rollout. Decide who can approve actions, which channels the colleague joins, and what it's allowed to do in DMs.
AI colleagues are heading for the org chart. Expect them to appear in onboarding docs next to the people they work with, each with an owner, a scope, and a short list of tools it's allowed to touch.
The teams that get the most from them will be the ones that wrote down how they work first: where pricing exceptions live, who signs off on refunds, and which channel gets the weekly numbers.
That document becomes the colleague's onboarding packet, and it helps your next human hire too (who will be equally relieved that nobody made them guess).
Yes, and Lindy is one that does: it sits in on Zoom and Google Meet calls, writes up notes, and adds them to a meeting library your whole team can search. So one person attends, and anyone who missed it can ask what got decided.
It depends on the tool, since most scope memory to where it was learned: Claude Tag keeps memories per channel, and Slackbot only surfaces what you can already access. Before rollout, ask whether you can open and edit the memory directly, which is the simplest way to audit what it knows.
Most work for both, and the place you ask decides which: tag one in a shared channel and the whole thread benefits, or DM it for something only you need. Check the billing model too, since some charge per person who uses it and others draw on one shared pool.