AI agents for small business: 10 we put into production, what they replace and cost

Vendors call almost anything an agent now. Here is the plain version.

Most of the ten are one AI step inside ordinary code. Code fetches the call, the reply or the CV. The model reads it and returns an answer in a fixed format. Code saves it, shows it on a dashboard and alerts someone. The model decides one thing; the rules around it stay fixed.

Two are agents in the stricter sense: they pick their own next step and use tools. Tala's architect agent reads a brief, designs a video pipeline and wires the nodes itself. Our writing partner is a Claude agent that edits a post, its hooks and its visual through tools while Mark chats with it.

For a small business, the first kind is where the money is. It's predictable, cheap per item, and easy to check. Start there.

• NextEnergy, sales-call compliance. Partner recordings arrive on FTP; each call is transcribed and checked against 30+ yes/no compliance criteria, repeated for consistency. The old flow asked for one score per call, and scores jumped from 60 to 85 for no clear reason. Many narrow questions fixed it.
• UK funding broker, lead engine and reply sorting. Public-register data runs through a waterfall to 14,620 qualified leads, 72% with an email. Replies come back from outreach and an AI step sorts each into one of 8 intents.
• Baltoro, content and tenders. Posts in each client's voice, a weekly government-tender hunter for 6 clients, news-trigger leads, and a marketing chatbot used across 22 clients.
• Atlas People, talent reports. Psychometric tests, CVs and role notes become job-analysis and candidate reports. Each report section has its own prompt and output format; guardrails stop the model inventing talents or mixing languages.
• Tala, ad pipeline agents. A main Claude agent works with architect, editor, animator and playable agents. One ad concept turns into many variants in parallel. Customers reported a batch going from 3 days to one morning.
• Our writing partner. Claude drafts three hooks and a post for each idea in Mark's voice, and a chat agent edits it with him. Mark edits every draft before it is posted.

There are three costs: building it, running it, and checking it.

Running is usually the smallest. Where we measured it: about $0.138 per kitchen render, under $3 per 20-minute talent report, about £0.18 of data per qualified lead. Model prices per token are low and still falling. The table below shows what the models and two ready-made agents charge.

Building depends on how many systems it touches. From the case pages: 70 hours for NextEnergy's pipeline and app, under 3 weeks for the broker's two apps, about 8 weeks for Incorra's Pipedrive app. Our own offer is a working version in 1 week, from $1,990, running in your accounts.

Checking is the cost people forget. Someone has to read the misses in the first weeks and fix the prompt or the rules. At Atlas People, early tests invented talents and mixed Danish with English until guardrails went in. At NextEnergy, one big score per call had to become 30+ small questions. Budget a few weeks of that before you trust the numbers.

Every client system in our table was a build, because each one scored or wrote against the client's own criteria: the clinic's 18-point script, Dutch compliance rules, a recruiter's vacancies. No off-the-shelf agent ships with those. If yours is a common job with a common answer, buy.

• The rule is fixed. A US private investment firm paid thousands of investors by hand every month, and each run took 3 days. We cut it to about an hour with a step-by-step payout workflow and bank-connected payments. More than $9M went out in a year. There is no AI in it, and there shouldn't be: a payout is arithmetic plus approvals.
• The volume is tiny. The dental build handles 1,200 calls a month. At 20 calls a week, a manager can listen.
• Nobody will check the output. Every agent above shows its evidence: a quote, a reason, a matched contact. If nobody looks, don't hand the decision over.
• The data can't leave your systems. Check client contracts and your industry's rules before any record goes to a model provider.
• You want it to talk to customers unsupervised on day one. Start with a job where a person reviews the output, like scoring or drafting. Move to customer-facing work once you trust it.

1. 1.Pick one job a person does on every record: every call, every reply, every CV, every report.
2. 2.Write down the criteria a good employee uses. Split them into yes/no questions; one big score drifts.
3. 3.Collect 20 real examples with the right answer for each.
4. 4.Run them through a model in a chat window. Count the hits and read the misses.
5. 5.If it holds up, build it as one AI step inside code, with the evidence saved next to each answer and a screen where someone checks the edge cases.
6. 6.Measure the cost per item from the first week. You'll want that number later.

Which model? We pick per job. Of the ten above, the reading and scoring mostly runs on OpenAI models, the renders on Gemini, Tala's agents and our writing partner on Claude. More on Claude for this work in Claude for small business, and on when a workflow tool is enough in n8n vs Claude Code.