How do you actually put AI to work in a small company?

Most small companies I talk to have already “tried AI.” Someone has a ChatGPT login, someone else uses it to rewrite emails, and nothing about how the team actually works has changed. The gap isn't the technology. It's that nobody wired AI into the places where the work happens: the meeting that ends with five action items nobody writes down, the Slack thread where a bug gets reported and never makes it to the tracker, the “I'll send that over” that slips by Friday.

In 25 years of running operations, most of it as a COO, I learned that small teams rarely drop the ball on the hard work. They drop it in the handoffs. That's exactly where AI is now good enough to help. Below are the five setups I'd start with, written for Claude because it connects to the tools small teams use most, though the same ideas work in other assistants.

If you're still deciding which tasks to hand over, start with which tasks a small business should hand to AI agents first. If you're planning a company-wide rollout, read how a company should start implementing AI. This article is the hands-on part: what to switch on and what to type.

What does every AI setup need before it works?

Every setup below is built from the same three pieces. Once you see them, you can design your own.

1. A connection. Claude can only act in tools it's connected to. These are called connectors: you add them under Customize > Connectors in Claude and sign in to each tool once. There are ready-made ones for Gmail, Google Calendar, Google Drive, Slack, Asana, Linear, Jira, Notion, Fathom, Granola and many more. For tools without one, you can add a custom connector by pasting the tool's MCP server address (MCP is the standard way AI assistants plug into apps).
2. Instructions. A short, specific brief, the same one you'd give a new assistant: what to look at, what to produce, what format, and a never list (never send to a client, never delete, never discuss money). The never list is what turns a clever demo into something you trust.
3. A trigger. Either a schedule (“every weekday at 5pm”) or a mention (“@Claude, file this”). In the Claude app, scheduled tasks run hourly, daily, on weekdays or weekly, and they run on Anthropic's servers, so they keep going when your laptop is closed. They're included on every paid plan.

Setup 1: How do you turn meeting transcripts into action steps automatically?

This is the one I'd build first in almost any company. Meetings end, everyone agrees on next steps, and by Thursday half of them are forgotten. A meeting agent fixes that without anyone having to take notes.

What you need:

• A meeting recorder that already makes transcripts. Fathom and Granola both have Claude connectors, and several others do too. The Fathom connector is read-only, which is exactly what you want: Claude can read transcripts, summaries and action items but can't change anything.
• Your email connector (Gmail, for example) so Claude can write recap emails as drafts in your account.
• Your task tracker's connector (Asana, Linear, Jira, or Shortcut as a custom connector) if you want the action items filed as tasks too.
• A scheduled task that runs every weekday at the end of the day.

How to set it up: In Claude, open a new conversation and say “create a scheduled task that runs every weekday at 5pm,” then paste the instructions below. (In Cowork you can also go to Scheduled > New task.) Run it once by hand on a day with real meetings before you let the schedule take over.

Your end-of-day job shrinks to five minutes: open your drafts, fix any OWNER? lines, and press send. The recaps go out the same day, which clients notice, and nothing agreed in a meeting depends on someone's memory again.

Setup 2: How do you create tickets by tagging @Claude in Slack?

Here's the setup people get most excited about. Someone reports a problem in Slack, the team talks it through, and instead of somebody copying it into the tracker later (or never), you type:

“@Claude file this thread as a ticket in Shortcut, assign it to the owner we discussed above, and post the link here.”

Claude reads the thread, writes the ticket with a clear title, the details and the agreed owner, creates it in your tracker and posts the link back in the thread. The same trick turns a thread into a decision doc, a customer reply or a planning outline.

What you need:

• Claude Tag, the version of Claude that lives in your Slack workspace as one shared teammate. It needs a Claude Team or Enterprise plan and is in public beta as of October 2026. Your Slack users don't each need a Claude seat. Channel work is billed by usage against a monthly spend limit the owner sets.
• A one-time setup by an owner: go to the Claude Tag page in your organization settings, send @Claude connect in Slack to get a pairing code, enter it, set the spend limit, and add Claude to the channels you want.
• A connection to your tracker. Asana, Linear and Jira are in the standard list. For Shortcut, first add Shortcut's official hosted MCP server, https://mcp.shortcut.com/mcp, as a custom connector in Claude; it can create and update stories, add comments and create epics, and Claude Tag can then use it. Basecamp is the one that takes more work today: its official MCP server comes inside the Basecamp command-line tool and runs on your own computer with Claude Code or the Claude desktop app, so it's great for one person but not yet a one-click option for a shared Slack channel.

One important detail: it's an @mention, not a spoken wake word. Typing “Claude” in a message doesn't call it; tagging @Claude does. To make it automatic for a whole channel, save the rule to the channel's memory once:

Use a dedicated account for Claude in your tracker (a seat named “Claude”) rather than your own login. Anyone in the channel can ask Claude to act, so it should have its own, narrower access, and every ticket it creates is easy to spot in the history.

Setup 3: Can AI send Slack messages on your behalf?

Yes, and this one saves more time than it sounds like. Think of all the small nudges you send in a week: “Can you send me the Q3 numbers by Thursday?” “Reminder, the client call moved to 2pm.” “Great work on the launch.” You can hand those to Claude in one sentence each.

What you need: the Slack connector on your own Claude account (available on Team and Enterprise plans). Once connected, Claude can search your workspace for context and send a message to a person or a channel as you, within the permissions you grant. If you use Claude Tag, a one-to-one DM with @Claude runs on your own account and your own connectors too.

Two rules keep this safe. First, keep sending on “ask first,” so Claude shows you every message before it goes. It takes one click and you'll catch the one time it misread the situation. Second, know the difference between the two Claudes: a message from your own connector goes out under your name, while Claude Tag in a shared channel posts as Claude, and its scheduled routines never DM people directly. They nudge in the channel thread instead, which is usually what you want for team reminders anyway.

Setup 4: How do you get a daily project update without asking for one?

Every small team has someone who spends Friday afternoon asking “where are we on this?” in five different places. With Claude Tag, one message in the project channel sets up a routine, a standing job that posts on a schedule:

Three details make this work. Name the timezone, since routines run on a fixed clock (and recheck the time after daylight saving changes). Name the format so every digest looks the same and stays scannable. And use numbers like “three working days” instead of “stale,” so stalled work surfaces itself. Anyone in the channel can ask “what routines do you have set up here?” to see, change or stop them.

Setup 5: How do you stop forgetting what you promised people?

The last setup is for owners and managers personally. Over a week you say “I'll send that,” “let me check and get back to you” and “I'll intro you” dozens of times, across email, Slack and meetings. A weekly sweep catches every one of them. Set it up as a scheduled task for Friday morning:

The first time you run this, expect a longer list than you'd like. That's the point. Clients and staff remember the promises we forget, and closing those loops is one of the cheapest ways to look organized.

How do you keep AI setups safe in a small company?

Treat each setup like a new hire on probation. The rules I use with clients:

• Give AI its own accounts in shared tools. A dedicated “Claude” seat in your tracker or Google Workspace keeps its work auditable and lets you switch it off without touching anyone else's access.
• Read-only wherever possible. If a setup only needs to look things up (meeting transcripts, docs, reports), connect it read-only.
• Drafts before sends. Anything that reaches a client or goes out under someone's name gets reviewed by a person until the setup has a clean record.
• A never list in every set of instructions. Money, HR topics, deletions and anything legal stay with people.
• Watch the bill. Set the spend limit on Claude Tag before you launch, then check which channels use the most after the first month.
• Write it down. A one-page AI use policy (what's allowed, what's not, what data never goes in) saves a lot of awkward conversations later. The rollout article covers what to put in it.

Which setup should you start with?

Pick the one that fixes your loudest problem, and give each setup two weeks before adding the next:

• Weeks 1 and 2: the meeting agent. It works for one person on any paid plan, shows results on day one, and teaches everyone what good AI instructions look like.
• Weeks 3 and 4: @Claude in Slack for ticket filing in one channel, if you're on a Team plan. Pick the channel where requests get lost most.
• Month 2: the daily digest and the promises sweep. By now you know where AI is reliable and where it needs a human check.

Time each task by hand before you build, then time your review after two weeks. That difference is your real savings, and it's the number that tells you what to automate next.

Frequently asked questions

Still choosing what to automate? Read which tasks to hand to AI agents first for the four-question test and the savings math, or how a company should start implementing AI for the full 90-day rollout plan.