How to use AI at work: a practical guide for small teams

The useful way for a small team to use AI at work is by job, not by tool: pick a routine job someone does every week, let AI do the first pass, and have a person check the result before it reaches anyone else. Five jobs work well for most 3–10 person teams: drafting replies, summarising long threads, turning messages into tasks, checking deadlines, and doing routine steps such as filling in a form. Decisions, anything sent in a person’s name, and payments stay with a person. Write those rules down in a few lines so everyone uses AI the same way.

This guide is vendor-neutral: it works with a general AI chat tool, AI features inside the tools you already use, or an AI chief of staff. How Opitor does it is in its own section near the end.

Which page to read

Use this guide to decide what your team uses AI for. When you need the rules in a formal document for staff, fill in the free AI acceptable use policy template. If you are choosing between kinds of AI help, read AI chief of staff vs AI assistant. If you are a founder deciding for yourself first, read AI chief of staff for founders. For the terms used here, such as AI coworker, see the glossary.

Five jobs AI does well, and the check a person makes

Job What AI does What a person checks Stop using AI for it when
Drafting replies Writes a first draft from the thread and your notes Facts, names, numbers, tone; the person sending changes at least one line Drafts need rewriting more often than not
Summarising threads Pulls the decisions, open questions and owners out of a long conversation Each decision against the message it came from The summary leaves out the one point that mattered
Turning messages into tasks Suggests a task with an owner and a date from an agreed message That the owner actually agreed and the date is theirs It keeps turning ideas into tasks nobody accepted
Checking deadlines Lists what is due, overdue or missing a date The dates against the list the team trusts It works from a copy that is out of date
Routine steps Fills in a form, prepares a file, looks something up Every step that sends, submits, changes or pays, before it happens It cannot show what it is about to do

Drafting replies

Give the AI the thread, what you want to say in one sentence, and anything it cannot know: the price you agreed, the date you can actually meet. Ask for a short draft. Then read it as if you were the person receiving it. Most of the value is in not starting from a blank page; most of the risk is a confident sentence about something that is not true. Drafts, not sends explains why the person whose name is on a message should be the one who sends it.

Summarising long threads

A summary is useful when someone was away or a decision is spread across forty messages. Ask for three things only: what was decided, what is still open, and who owns each open item. Check each decision against the message it came from before you repeat it to anyone. If a summary is the only record of a decision, the decision is not recorded; put it where the team keeps decisions.

Turning messages into tasks

AI is good at spotting “I’ll send it Thursday” in a busy thread. It is bad at knowing whether that was a promise or a guess. Let it suggest the task, owner and date, and have the owner confirm before it goes on the team’s list. The message-to-task checklist covers what a good task needs.

Checking deadlines

Ask once a day or once a week: what is due this week, what is overdue, and what has no date but someone is waiting on. The answer is only as good as the list it reads, so point it at the one list the team trusts, not a copy.

Routine steps in other tools

Some AI tools can now act in other websites and apps: fill in a supplier form, change a setting, download a report. Treat each of those as a permission you grant one at a time. A tool that can submit, send or pay should show you the step and wait for your confirmation first, and you should be able to see afterwards what it did. Handing over a whole task rather than one step is covered in how to use AI agents at work; for the safety questions specific to websites, such as where passwords live and what happens when a payment fails, read can AI fill out forms for you.

How to use AI agents at work

An AI agent takes steps in other tools to finish a task, rather than only answering. For a small team, using one well is mostly about the handoff: what you give it, what it may do alone and where it must stop. What separates an agent from an assistant or a chief of staff is explained in chief of staff and agents; this section is about putting one to work.

  1. Start with a task you can check from end to end. A bounded, repeated task with a visible result: put three supplier quotes into one table, fill in the same monthly form, prepare a file from a template. “Handle the supplier” is not a task an agent can finish or you can check.
  2. Write the brief you would give a new colleague. The goal, where to look, what done looks like, and what it must not do. If you cannot write it in five lines, the task is not ready to hand over.
  3. Give it the narrowest access that does the job. Only the tools and accounts the task needs. Keep payment details and other people’s data out of reach unless the task is about them.
  4. Fix the confirmation points before it starts. Anything that sends, submits, changes a record, deletes or pays waits for a person. Try it on a test account first and check that it really stops at those points.
  5. Check the result and the trail. Compare the result with the brief, and look at what it did and in what order. If you cannot see what it did, do not give it the task again.
  6. Give every agent an owner. One named person answers for its work, decides what it may do next, and switches it off.

When an agent takes requests from several people, which some products call an AI coworker, add one rule: agree who may ask it for what, so two people do not hand it conflicting work. AI coworkers for a small team covers what to give one, the decisions to make before adding it and a role card to copy. Which product to choose depends on where your team’s work already lives; judge each one by what it may do without a person confirming, not by the length of its feature list.

Put the first agent task through the two-week trial at the end of this guide, and keep rule 5 of the team agreement below for every agent you add.

What to keep human

What not to put into an AI tool

Before anyone pastes work into an AI tool, agree what never goes in: passwords and access codes, payment card and bank details, customer personal data, anything under a confidentiality agreement, and colleagues’ personal information. Then agree which tools are approved for everything else, and under which accounts. Read each tool’s own terms and data settings; they differ, and they change. The AI acceptable use policy template has a section for each of these.

A short team agreement on AI use

A small team usually does not need a long policy on day one. It needs a few lines everyone has read. Copy this, change it, and pin it where the team works:

How we use AI at work
1. Approved tools: [list]. Use your company account, not a personal one.
2. Never paste in: passwords, payment details, customer personal data,
   confidential client material, or colleagues' personal information.
3. AI drafts; people send. Whoever sends a message has read every word.
4. Check names, numbers, dates and quotes before anything leaves the team.
5. Nothing is paid, ordered, deleted or submitted without a person
   confirming that step.
6. Decisions about people, prices and priorities are made by people.
7. Questions or a mistake: tell [name] the same day. We review this on [date].

When the company grows, works with regulated data or is asked by a client for a written policy, turn these lines into a full policy with the template. It is a starting template, not legal advice.

Try it for two weeks

  1. Each person picks one job from the table above that they do at least weekly.
  2. Use AI for the first pass only on that job, and keep the human check every time.
  3. Note two things at the end of each week: what you had to correct, and whether the job took less of your attention. These are your own observations, not a benchmark.
  4. Keep, change or drop each use after two weeks. If the check takes longer than doing the job yourself, drop it.
  5. Share what worked in one message to the team, and add any new rule to the agreement.

Where Opitor fits

In Opitor, each person has a personal AI chief of staff in Messages to help with messages, todos, deadlines and drafts. Several of the jobs above are built around the human check: AI replies are drafts you choose to send, signed in purple, and your chief of staff proposes todos for you to confirm. In an organization, your chief of staff can hand work to built-in AI coworkers, such as HR and admin, and administrators can add outside AI agents as AI members. When you ask it to work on other websites on your own cloud computer, it asks you to confirm before every step that would change something, send something or spend money. The team agreement above still has to be written by the team. See the FAQ for what is included.

Start with one real project.

Choose a project your team is already working on. Open a channel, invite a teammate, and capture the next commitment as a to-do.

Join Opitor