AI agents for a small team: where to start, and what to leave alone

Every owner I talk to has heard that AI agents will change their business. Very few have been told where to start. So they either do nothing, or they buy a tool that promises to do everything and ends up doing one demo well.

I’ve built apps, and now I build AI agents and automations. The useful ones are rarely impressive to watch. They take a boring, repeated job off someone’s plate and do it the same way every time. That’s the bar I’d use for your first one.

What an “agent” is, in plain terms

For a small business, an AI agent is software that can read something, decide what to do with it, and take an action, often by using other tools you already have. It might read an incoming email, work out whether it’s a sales enquiry, a support question or spam, draft a reply, and put it in front of a person to send.

An automation is the simpler cousin: when X happens, do Y. A form is submitted, so add the person to your CRM and send a confirmation. No judgement involved.

Plenty of what gets sold as “AI” is really automation. That’s fine. Automation is cheaper, more predictable and easier to fix. Use the AI part only where something actually needs reading or judgement.

The test for a good first project

I look for jobs that pass four checks:

1. It happens often. Daily or weekly, not twice a year. Otherwise you’ll spend more time building it than it saves.
2. It follows a pattern. If you can explain how to do it to a new hire in a page, a machine can probably do most of it.
3. Mistakes are cheap and visible. Early on, a person should be able to check the output before it reaches a customer. Start where an error is an annoyance, not a disaster.
4. Someone hates doing it. This one’s not technical, but it matters. Automations that take away a job people dread get used and looked after. The ones nobody asked for get quietly ignored.

Good places to start

These are the kinds of jobs that usually pass:

• Sorting the inbox. Reading incoming messages, tagging them by type, and routing them to the right person or folder. Drafting replies for common questions, with a human pressing send.
• Getting leads into one place. Taking enquiries from forms, DMs and email and putting them in one list with the details filled in, so nobody has to copy and paste.
• First replies to new leads. A fast, useful first message while a person gets to the full reply. Speed matters here, and it’s the kind of thing that slips on a busy day.
• Turning calls and meetings into notes. Transcribing, summarising, and pulling out the follow-up actions, with permission from everyone on the call.
• Weekly numbers. Pulling the same figures from your ad account, store and bank every Monday and putting them in one short summary.

None of these replace a person. They give a person back an hour or two a week, and they make sure the boring steps happen even when everyone is busy.

What I’d leave alone for now

Some things look tempting and tend to go badly at a small company:

• Fully automatic replies to customers on anything sensitive. Refunds, complaints, medical or legal questions. Keep a human in the loop.
• Anything where you can’t check the work. If nobody on the team can tell whether the output is right, you won’t notice when it isn’t.
• Replacing your judgement on strategy. AI is good at drafts, summaries and first passes. Deciding what your business should do next is still your job.
• A big platform before a small win. Buying a large system before you’ve automated one real job usually leads to a lot of setup and not much use.

How to start without a big budget

Pick one job that passes the four checks. Write down, step by step, how a person does it today. Be specific: where does the information come from, what decisions get made, where does the result go?

Then build the smallest version that handles the common case, and have a person check every output for the first couple of weeks. Keep a short list of what it got wrong. Most of the fixes will be about the instructions, not the technology.

Only once it’s reliable should you let it act on its own, and even then, keep a way to see what it did. A simple log that someone glances at once a week is enough.

Costs for this kind of thing are usually modest: a few software subscriptions and some usage-based charges from the AI provider. Watch the usage costs in the first month so there are no surprises, and set a spending limit with the provider if it offers one.

An example, step by step

Here’s how a first project might look for a small service business that gets enquiries by email and through a website form. This is an illustration, not a client story.

1. Write down today’s process. Someone checks the inbox a few times a day, works out which messages are new enquiries, copies the details into a spreadsheet, and writes a reply. On busy days, some enquiries wait until the evening.
2. Automate the plumbing first. Every form submission and every email to the enquiries address lands in one list automatically, with the time it arrived. No AI yet.
3. Add one AI step. For each new message, the agent decides whether it’s a new enquiry, an existing customer or spam, pulls out the name, what they want and any deadline, and drafts a first reply in the business’s tone.
4. Keep a person in charge. The draft goes to a person, who edits and sends it. Nothing reaches a customer without that check.
5. Measure. Track how long first replies take and how often the drafts need heavy edits. After a few weeks, you know whether it’s working and what to fix.

The result is modest on paper: faster first replies, fewer forgotten enquiries, and less copying and pasting. For a small team, that’s often worth more than an impressive demo.

Privacy, briefly

Whatever you automate, you’re moving customer information through more tools. Check where each tool stores data and whether it uses your data to train its models. Tell customers what you use, in your privacy policy, in plain language. If a job involves sensitive personal information, think hard about whether it needs AI at all.

Where I fit

If you’d like help picking the first job and building it properly, that’s something I do. Manyhats Automations is the company I set up for that work, and I own it, so if I suggest it, you’ll know. If the answer is that you don’t need AI for this yet, I’ll tell you that too. Book a call and tell me which job you’d most like to stop doing.