LLM agents for small teams: workflows that start themselves

A seven-person company buys the same software a seven-hundred-person company buys. What it lacks is someone whose main job is keeping records current. When a customer emails new billing details, the salesperson on the account is supposed to update the CRM, but they’re busy selling, and the change slips through the cracks. A bigger company has operations staff for that, so someone catches it before the next invoice run. At the seven-person company nobody does, so the update happens late, or after the invoice bounces.

Advice for small teams usually arrives as a list of tools. Small teams already have tools. What they don’t have is an hour to wire them together. The bigger problem comes earlier: nobody is watching what arrives before anyone has built an automation for it, or even noticed that one is needed.

Attention is the constraint

Microsoft’s study of the infinite workday found the average worker is interrupted once every two minutes, 275 times a day. In a large company, requests from customers and partners mostly reach people in reactive roles, such as operations or support, whose day is built around responding. In a small one they reach the people who are also selling and shipping.

Buying more AI doesn’t fix that by itself. Atlassian’s State of Teams 2026 found that 89% of executives say AI increases speed, while only 6% are sure they have clear examples of AI returns across the organization. Speed at one desk is easy to feel. It doesn’t add up across the company while somebody still has to read each message and decide what it means.

What an LLM agent adds to workflow automation

An LLM agent is a program that reads unstructured input, decides what to do about it, and calls tools to do it. Rule-based automation, by contrast, needs its input in fields. Somebody submits a form, a CRM record changes, a webhook fires, and the rules run. Messages don’t arrive that way. A renewal question and a change of billing details are both emails from somebody you know, and a rule that only sees the sender can’t tell them apart.

AI reads the message as written, so a step that used to need a rule now needs a sentence. In this+that, a workflow can ask Claude to classify or decide mid-run, branch on the answer, loop over a list, call a tool on any connected MCP server, or hand one step to a sub-agent with a narrow set of tools. The AI decides during the run, so you don’t have to anticipate every case in advance.

An invoice arrives, a workflow classifies it, and it reaches finance without anyone opening it. A quote request arrives, and a workflow reads your current pricing from the brain and drafts the reply for a person to send.

Work that starts when a message arrives

A chat assistant waits to be asked. You open a chat window, you type, and something happens. That covers work you remembered. The work that hurts a small team is the work nobody remembered, like the reply that needed sending before the weekend, or the follow-up somebody promised in a thread.

So ask what starts an LLM agent. In this+that, a workflow can start from triggers such as a message on a connected channel (Gmail, Outlook, Slack, Microsoft Teams, Google Chat, and WhatsApp Business), a change to a task in DoBox, a new contact, a schedule, or a person pressing run.

Where non-technical setup is easy

To connect a mailbox, you sign in to it, and this+that starts extracting tasks from the messages already there without anyone writing a rule. A small team can do that between meetings, and tasks start arriving in DoBox before any workflow exists.

From there:

• Prebuilt workflows. Our catalog has ready-made workflows and agents you install and edit rather than build from scratch.
• Plain language. You describe what you want and get a workflow you can open, read, and change.
• An approval gate. A workflow can draft and hold instead of sending, so a bad response shows up as a draft you delete.
• Corrections that stick. Reject a task the AI suggested, say why, and that becomes a standing rule that shapes which tasks it suggests later.

Connecting a mailbox gives us access to your mail. We know it’s confidential, and that where it goes matters. Our AI runs on Amazon Bedrock inside AWS. Bedrock doesn’t retain prompts or use them to train models, and we hold SOC 2 Type I as of July 31, 2026, with Type II underway.

Where setup isn’t easy

The prebuilt workflows, the approval gate and your corrections all work without needing to know much about your company. The moment a workflow needs your pricing or your account history, the brain has to hold them, and getting them there takes setup.

Setting up a company brain takes four steps, and most of them need a person today. You decide which kinds of thing your company tracks, such as customers, products or competitors. You list what belongs to each kind, which for a company keeping its customers in HubSpot or a spreadsheet requires an import somebody builds. You add required fields where a kind needs them. Then this+that reads your messages to draft what each page says.

After that, workflows do most of the upkeep. Contacts keep themselves current from your messages with no workflow at all. The prebuilt Company Brain workflow keeps pages current from public Slack channels, email and your codebase, and holds updates from private sources for approval. It’s an early version of what we’re building toward: one workflow designed to maintain most of the brain.

Most of the setup needs no engineer, though an import from a CRM or a spreadsheet may. What it always needs is somebody who knows how the company actually works and has time to answer questions about it. On a seven-person team that’s usually the founder, and the time they put in up front comes back each time someone gets an answer from the brain instead of from them.

What to check before you pick an agent

• Does anything happen before somebody asks? An agent you have to open is a faster way to do work you already remembered.
• Can the person who noticed the problem make the change? If every fix routes to whoever maintains the automations, your automation list becomes somebody’s backlog.
• What does it know about your company, and who has to tell it? Ask for the setup path rather than the demo.
• What happens when it’s wrong? Look for a way to review its work before it goes out. That protects you more than any claim about how accurate it is.
• Does what it learns stay with the company when somebody leaves? Panopto’s 2018 survey put 42% of institutional knowledge with one person and nobody else. A small team feels that the day somebody resigns.

Where the general platforms win

Zapier, Make and n8n beat us on a large class of problems. When an automation spans many systems, needs a developer’s control over retries and errors, or runs on infrastructure you host yourself, go there. n8n’s node editor and self-hosting suit a team with an engineer who wants to own the deployment, and Zapier’s app coverage is far wider than ours. We compare each in more detail on our Zapier, Make and n8n pages.

Where we differ is the starting point. DoBox reads every message without a rule and turns the requests in them into tasks, so work nobody anticipated still shows up as a task, even on a team with nobody spare to notice.

If most of your work arrives as a message from outside the company, connect one mailbox and see what tasks come out of it and what you can automate before you build anything. The first 30 days are free.