AI agents for business: what actually works for a 2 to 20 person company

The short answer: In a company of 2 to 20 people, the AI agents that pay off do one repeatable job, work from your own data, and hand a draft to a person before anything goes out. Inbox triage, lead follow-up, meeting notes turned into tasks, and first-draft replies to routine questions all work today. "An agent that runs your sales" does not. Budget roughly $20 to $100 a month per agent in software, plus a few hours of setup.

Prices from each vendor's official pricing page, verified between 7 August and 27 August 2026. See the linked tool pages for current figures.

What people mean by "agent"

A chatbot answers when you ask. An agent acts: it watches for something (a new email, a form submission, a calendar event), decides what to do, uses tools (your CRM, your inbox, a spreadsheet), and produces a result. The word gets stretched to cover everything from a Zapier step with an AI prompt in it to a model that browses the web on its own.

For a small business, the useful question is not "is it a real agent?" It is: what job does it do, and who checks its work?

Five agents that work at small-team scale

1. Inbox triage

Reads incoming email, labels it (sales lead, support, invoice, spam), drafts a reply for routine ones, and flags the rest. The person still hits send. This is the most common starting point because the input is already digital and the cost of a wrong label is low.

2. Lead follow-up

When a form or enquiry comes in, the agent looks up what it can, drafts a personal first reply, logs the lead in your CRM, and reminds you if nobody has answered in a day. See our guide to AI customer follow-up.

3. Meeting notes to action items

Records the call, writes the summary, pulls out decisions and owners, and posts tasks to your project tool. The recording tools do most of the work; the agent part is routing the output. More in AI meeting notes that lead to action.

4. First-draft answers from your own documents

Staff or customers ask a question, and the agent answers from your policies, price list or help pages, with a link to the source. It works when the documents are current. It fails when they are not.

5. Routine data movement

New order to spreadsheet, invoice to accounting, form to CRM, with an AI step to clean up or classify messy text on the way. Less exciting, and often the biggest time saver.

What does not work yet

• Unsupervised customer-facing agents that send, refund or promise things without review. One wrong promise costs more than months of saved time.
• "Do my marketing" agents with no clear inputs or success measure.
• Agents on top of messy data. If your CRM is half-empty, an agent will confidently act on the half that is wrong.
• Buying an agent platform before you know the job. Pick the job first, then the tool.

What it costs

Three common ways to build these, with official prices:

• Assistant-based agents inside tools you already pay for. ChatGPT Business is $25 per user per month ($20 billed annually, minimum 2 users). Claude Team is $25 per seat per month ($20 annually). Good for drafting and document answers; limited for acting in other apps.
• Automation platforms with AI steps. Zapier Professional starts at $19.99 a month billed annually ($29.99 monthly), Make Core at $12 a month billed annually, and n8n Starter at €20 a month billed annually, with a free self-hosted edition. Best for inbox triage, lead routing and data movement. Compared in Zapier vs n8n vs Make, priced.
• Dedicated agent products. Lindy Plus is $49.99 a month. Faster to set up for email and calendar work, less flexible for custom data.

If you build on a model API instead, per-token costs for routine jobs can be very low. DeepSeek V4.1 Flash charges $0.15 per million input tokens off-peak. Someone still has to build and maintain it.

The hidden cost is setup and upkeep: writing the instructions, connecting accounts, testing on real examples, and fixing it when a vendor changes something. Plan a few hours to set up each agent and an hour a month to check on it.

How to start

1. Pick the one job that eats the most time and has a clear "done" state.
2. Write down the steps a person follows today. That becomes the agent's instructions.
3. Build it with a human approval step. Run it in parallel with the person for two weeks.
4. Only remove the approval step for outputs that have been right every time, and keep spot checks.
5. Add a second agent after the first one is boring.

Getting help

A Business AI Plan ($99 one-time, optional $29/month) names the first agents worth building for your business and the tools to build them with. If you want someone to build and connect them for you, see custom integration, from $1,000. Not sure yet? Try the free 2-minute AI plan finder.

Sources: ChatGPT pricing, Claude pricing, Zapier pricing, Make pricing, n8n pricing, Lindy pricing, DeepSeek pricing

See also: AI agents for business, which maps agent jobs to roles and the approval gates that stay with you, and the Dapols AI consultant.