AI Agent Use Cases for Small Businesses: 10 Worked Examples

AI agents can answer routine customer messages, reconcile bank payments, fill supplier baskets, pull details from referral letters, research leads and build reports, working through several steps alone. Running costs are usually tens of dollars a month per job, rising with volume, and a person still checks anything involving money, health or complaints.

The jobs that suit an agent share a shape: several steps, the same pattern most of the time, a few exceptions that a person can pick up, and mistakes that can be undone. Answering "what are your opening hours?" fits; deciding a refund doesn't. Matching a payment to an invoice fits; paying a supplier doesn't. Every example here keeps the irreversible step with a person.

Follow me on Instagram@sagnikteaches

The businesses are illustrations (a nursery, a dental group, a tutoring agency, a language school, a physiotherapy clinic and a veterinary practice), and the numbers are realistic but invented. Prices are list prices in USD as of September 2026, taken from each vendor's own pages, and they change often, so check before you buy. If the term itself is new to you, what an AI agent is, in plain English covers the basics.

Connect on LinkedInSagnik Bhattacharya

The ten at a glance

Two of those costs assume the business already pays for the underlying plan (HubSpot Professional in examples 2 and 10), which is a big assumption; the examples spell out the full price. Time saved is the gap between doing the job by hand and checking the agent's work, not the agent's running time.

Subscribe on YouTube@codingliquids

Customer conversations

1. A nursery's enquiries across three messaging apps

Before. About 120 enquiries a month arrived on Instagram, Facebook and WhatsApp, mostly from parents asking about places, age ranges, session times and fees. The manager answered in the evenings. Most parents waited until the next morning, and a few booked visits elsewhere in the meantime.

After. With the three accounts connected to Meta Business Suite's inbox, Meta Business Agent answers the routine questions from the nursery's information and offers the link to book a visit. It hands over anything about funding, additional needs, a specific child or a complaint. The manager reads the hand-overs twice a day.

Tool and cost. Meta has charged for its Business Agent per token since 1 August 2026, which it puts at roughly 4-5 cents a message. At about three AI messages per enquiry, 360 messages a month cost $14-$18.

What a person checks. Every visit booking is confirmed by the manager, and ten conversations a week get read in full. The rule that matters most: the agent never discusses an individual child. In the first week the agent quoted last term's session fees, because the fee sheet it worked from hadn't been updated for September; the fix was a calendar reminder to replace that document the day fees change. The full set-up for this kind of inbox is in one AI inbox for WhatsApp, Instagram and Facebook.

2. A dental group's website chat

Before. A three-site dental group already running its patient communications through HubSpot Service Hub Professional had about 300 website chats a month, answered by receptionists between phone calls. Many were the same questions: prices from, parking, how to register, whether a site takes new patients.

After. HubSpot's Customer Agent answers from the group's help articles and hands anything clinical, urgent or account-specific to reception. In this illustration it resolves about half of the chats, 150 a month.

Tool and cost. The Customer Agent needs any Professional or Enterprise hub (usually Service Hub). It uses 50 credits (about $0.50) per resolved conversation, and credits cost $10 per 1,000. Professional includes about 3,000 credits a month, and unused credits don't roll over. 150 resolutions use 7,500 credits, so the extra 4,500 cost about $45 a month. For a business not already on HubSpot, add Service Hub Professional at $90 a seat a month on annual billing and a $1,500 onboarding fee, which changes the sum completely.

What a person checks. HubSpot counts a resolution when there's no hand-off within 72 hours, so a patient who gave up counts the same as one who was helped. Reception reads a weekly sample of "resolved" chats, and anything mentioning pain or swelling is on the hand-over list from day one. The first review found the agent confidently answering a question about one site's parking with another site's details, because the help article didn't say which site it covered. Splitting it into three site-specific articles fixed it.

3. A tutoring agency's parent emails

Before. An administrator spent about five hours a week on parent emails, most of them variations on five questions: how to move a session, where to find an invoice, what happens when a tutor is ill, term dates, and how to pay.

After. Help Scout's AI Answers replies from the agency's help articles on the questions it can answer and passes the rest to the administrator. In this illustration it resolves about 120 questions a month.

Tool and cost. AI Answers is a $0.75-per-resolution add-on on Help Scout's paid plans only; the free plan doesn't include it. With two users on Standard at $25 each, that's $50 plus 120 x $0.75 = $90, so $140 a month. Help Scout doesn't bill a resolution if the parent escalates, searches the help docs, asks again or says they need more help, which is a fairer definition than most.

What a person checks. Fee disputes and anything touching a child's wellbeing go straight to a person, flagged by keywords in the set-up. The administrator reviews five AI answers a week and updates the help article behind any answer that was out of date. The first month's lesson was that the help articles mattered more than the AI: two questions the agent kept passing on simply had no article, and writing them took 20 minutes.

4. A language school's online learner support

Before. The school's online courses generated about 600 support conversations a month: login problems, certificate requests, course access dates, how to change a live-class slot. Two part-time staff answered them, often late.

Learner: I finished level A2 last week but my certificate isn't
in my account. Also can I move Thursday's live class to Friday?

Fin (illustrative): Congratulations on finishing A2! Certificates
appear in My Account > Certificates within 3 working days of your
final assessment being marked. For live classes, you can switch to
another slot in the same week from My Classes > Change slot, as
long as it's at least 24 hours before the class starts...

After. Intercom's Fin handles questions like that one, with its answers drawn from the school's help centre articles, and hands over refunds and complaints. In this illustration Fin resolves about 350 conversations a month.

Tool and cost. Fin costs $0.99 per resolved outcome, and Intercom seats are $29, $85 or $132 a month billed annually depending on plan. At 350 resolutions and two seats on the lowest tier, the bill is about $347 plus $58, roughly $405 a month. That is the priciest example here, and it only makes sense at this volume: at 50 conversations a month, a person is cheaper.

What a person checks. Intercom counts an "assumed resolution" when the customer goes quiet for 24 hours after Fin's last answer. The school samples 20 "resolved" conversations a month to see whether silence meant satisfaction, and every answer about dates or deadlines is checked against the course calendar each term.

Money and admin

5. A tutoring agency's bank reconciliation

Before. About 180 parent payments a month arrived with references like "Maths Tues" or a child's first name, and matching them to invoices took the administrator about two hours a week.

After. Xero's JAX reconciles transactions automatically when it's highly confident, using four methods Xero calls Rule, Match, Memory and Prediction, and leaves the rest for a person. In this illustration it clears about 150 of the 180, and the administrator handles the other 30 in half an hour a week.

Tool and cost. Automatic reconciliation comes with Xero's Growing plan ($55 a month) and above; moving up from Early ($25) costs $30 a month more. Xero still labelled the feature beta in its August 2026 update.

What a person checks. Xero says you can challenge or reject JAX's matches, and the administrator does exactly that on a weekly skim of what it matched. The biggest improvement came from outside the software: invoices now tell parents to use the invoice number as the payment reference, which raised the share JAX could match with confidence. The one wrong match in the first month was instructive: two brothers taught by different tutors had invoices of the same amount, and a payment referenced only by surname was matched to the wrong one. The administrator rejected the match, and the invoice-number reference has stopped it recurring.

6. A nursery's weekly supplier reorder

Before. Every Friday the deputy manager spent about 45 minutes on a wholesaler's website, working through a paper list of nappies, wipes, gloves, cleaning products and craft supplies.

Using the stock sheet I've shared, add this week's reorder
quantities to the basket on [wholesaler website]. Only use items
already in our saved list. If an item is out of stock, don't pick
a substitute; list it for me. Stop at the basket - don't check out.

Illustrative result:
Added 14 of 16 items to basket (total shown: $[amount]).
Not added: nitrile gloves (M) out of stock; kitchen roll (pack of
24) price up 18% since last order - left for you to decide.
Stopped before checkout.

After. Claude in Chrome, running in "Manually approve" mode for the first month, works through the list in the deputy's own browser and asks before each step. The deputy reviews the basket, decides on the two flagged items and pays: about 10 minutes.

Tool and cost. Claude in Chrome is available on paid Claude plans; Pro costs $20 a month, or $17 a month billed annually.

What a person checks. Quantities, prices and any substitution. Anthropic's safety guidance steers Claude in Chrome away from financial accounts and from sites holding other people's personal information, and some actions such as purchases are blocked, so paying stays with a person and the nursery keeps it well away from its parent billing system. In week two the wholesaler redesigned its product pages and Claude stopped twice to ask for help; the deputy finished those steps herself. Browser agents depend on how websites are laid out, so budget for occasional hand-holding. Whether staff should use an AI browser agent covers the wider policy question.

Documents, data and reports

7. A physiotherapy clinic's referral letters

Before. About 60 referral letters a month arrived as PDFs. A receptionist retyped the referrer, the body area, the requested number of sessions and any urgency note into the clinic's tracker, at roughly ten minutes a letter.

Field agent instructions (illustrative):
From the attached referral letter, extract: referrer name and
practice; body area; number of sessions requested; any wording
that suggests urgency. If a value isn't in the letter, write
"Not stated". Don't infer a diagnosis.

After. The tracker lives in Airtable, where a field agent reads each attached PDF and fills the fields. The receptionist checks each record in about two minutes. With "Run automatically" switched on, the agent re-runs when a new letter is attached, and Airtable never overwrites a cell a person has edited.

Tool and cost. Field agents need a paid Airtable plan. Team costs $20 a user a month on annual billing, so $40 for the two people who edit the tracker, and includes 15,000 AI credits per billable user a month. Airtable's own example puts analysing a 10-page contract at about 200 credits, so two-page letters sit comfortably inside the allowance.

What a person checks. Every urgency flag is confirmed by a physiotherapist the same day, and "Not stated" fields are chased with the referrer rather than guessed. Early on, the agent flagged a letter as urgent because the patient had written "urgent, please" on the envelope scan; the instructions now say to flag only urgency stated by the referring clinician, and to note patient requests separately.

8. A dental practice's cancellation analysis

Before. The practice felt that missed appointments were rising but had never looked properly. Reminder timing was based on habit.

Attached: 12 months of appointments (date, weekday, time, type,
status: attended / cancelled late / no-show), no patient names.
Find patterns in late cancellations and no-shows by weekday, time
and appointment type. Show the counts behind every claim.

After. Microsoft's Analyst agent, one of the pre-built agents included with the paid Copilot licence, worked through the export. The illustrative headline: Monday 8am hygiene appointments had roughly twice the average no-show rate, and appointments booked more than eight weeks ahead were cancelled late far more often. The practice moved Monday reminders to the Sunday evening and added a second reminder for long-lead bookings.

Tool and cost. Copilot Business costs $21 a user a month on annual commitment ($18 for the first year on the promotion available until 31 December 2026); one seat for the practice manager.

What a person checks. Two of the counts were re-run by hand against the practice system's own report before anything changed, and the manager treated the patterns as prompts to test, not proof of cause. The export left out patient names, as every analysis like this should.

9. A veterinary practice's quarterly performance pack

Before. Each quarter the practice manager spent most of a day turning three exports (appointments by type, revenue by service, vaccination reminders sent and booked) into a spreadsheet and a short slide deck for the partners.

After. ChatGPT Work, OpenAI's agent for longer, multi-step work, takes the three exports and produces the spreadsheet with charts and a six-slide deck in the practice's usual order. The manager spends about 90 minutes checking and adding commentary.

Tool and cost. Work is included on paid ChatGPT plans such as Plus at $20 a month or a Business seat; long tasks use more of the plan's allowance.

What a person checks. Every total against the practice-management system's own reports (in the first run, one chart had counted cancelled appointments as completed), and client names removed from the exports before upload.

Sales and growth

10. A language school's corporate training leads

Before. The director wanted more corporate language classes but found prospects in spare evenings, a few a month, with no consistent follow-up.

After. HubSpot's Prospecting Agent researches companies that match the school's criteria and drafts outreach. In this illustration it recommends about 40 leads a month; the director edits and sends around 15 of the drafts and discards the rest.

Tool and cost. The Prospecting Agent is available from Starter upwards on most hubs (HubSpot's page of 16 September 2026) and uses 100 credits (about $1) per lead recommended for outreach. Forty leads use 4,000 credits; Professional includes about 3,000 a month, so the extra 1,000 cost $10. For a business not already on HubSpot, Sales Hub Professional lists at $90 a seat a month on annual billing plus a $1,500 onboarding fee, which is what really decides this example. Which HubSpot agents small teams should switch on goes further into the plan tiers.

What a person checks. Every email is read and edited before it goes, each company is checked as a real fit, and the rules on business marketing emails where the school sells are checked with an adviser before the first campaign. An agent that drafts is useful; one that sends unread outreach under your name is a reputation risk.

Two jobs that looked like agent work but weren't

Not every multi-step job suits an agent, and the failures are as instructive as the successes. Two from the same set of businesses:

Rescheduling appointments directly in the booking system. The physiotherapy clinic tried letting an agent move appointments when patients asked by message. It worked for simple cases, but twice in a fortnight it moved a patient into a slot that a receptionist had just offered to someone else on the phone, because the two changes happened within the same minute. The cost of a double booking, an annoyed patient and a physio with two people in the waiting room, outweighed the minutes saved. The clinic went back to the agent collecting preferred times and a person confirming, which is rung two on any sensible ladder of autonomy.

Sending termly reports straight to parents. The tutoring agency considered letting an agent draft and send progress reports automatically. Drafting was a clear win; sending wasn't. A report is a judgement about a child, parents read every word, and one wrong score or misplaced name damages trust in a way no time saving repays. The agent now drafts, tutors approve, and the administrator sends.

The pattern in both: the agent was capable, but the action was hard to undo and the cost of a mistake landed on a customer. Keep those steps with a person, however good the agent gets.

What the ten have in common

• A short leash. Each agent gathers, drafts, matches or answers the routine, and hands the rest to a person. None pays, sends unchecked outreach or makes a clinical judgement.
• Costs that follow volume. Per-message and per-resolution pricing looks trivial per unit. Work out a month at your own volume before switching on, as examples 2 and 4 show.
• Definitions that matter. "Resolved" means different things at HubSpot (no hand-off within 72 hours), Intercom (24 hours of silence after Fin's answer) and Help Scout (not billed if the customer asks again or escalates).
• Clean inputs. The reconciliation improved most when payment references improved, and the referral agent was only as good as the letters. Tidy the data before blaming the agent.
• A named owner and a weekly look. Ten minutes a week reading a sample of the agent's work catches drift before customers do.

Choosing your first agent from this list

Pick the job that is most repetitive, least risky and already costing someone several hours a month. For many small businesses that's either customer enquiries (examples 1 to 4) or reconciliation (example 5), because both are frequent, rule-based and easy to check. Then:

1. Price it at your volume, including any plan upgrade, using the vendor's current pricing page. What an AI agent costs a small business lays out the categories.
2. Write the hand-over rules first: what the agent must never answer or do.
3. Run a one-month pilot on a low setting (drafts, manual approval, out-of-hours only) with a weekly review. Piloting your first AI agent without risking customers sets out the steps.
4. Measure against the "before": hours spent, response times, errors caught. Keep the agent only if the numbers moved.

One agent that works, with its limits written down and someone checking it weekly, is worth more than five switched on in a burst of enthusiasm. Add the second when the first has become boring.

AI agent use cases: follow-up questions

Which of these agents is the cheapest to try?

The ones built into software on your current plan. If you're on Xero's Growing plan, automatic reconciliation costs nothing extra to switch on; if you already have ChatGPT Plus or a Business seat, ChatGPT Work is included. Meta's Business Agent costs only a few cents per message, so a month's trial on a small volume of enquiries usually costs less than $20.

Do I need all ten?

No. Most small businesses get the most from one or two, chosen where the same multi-step job eats hours every week. Start with the job that is most repetitive and least risky, run it for a month with a weekly review, and only then add a second agent. Several half-watched agents are riskier than one well-run one.

What happens if an agent makes a mistake with a customer?

You own it, just as you would a staff member's mistake, which is why every customer-facing agent here hands over complaints, money questions and anything personal. Keep the agent's instructions and a weekly sample of its conversations on file, correct errors with the customer directly, and fix the instruction or source document that caused the mistake.

Further reads

• How to Prepare Your Small Business for AI Agents — Getting data and processes ready before any agent starts.
• What Can Go Wrong When AI Agents Take Actions for You? — What can go wrong once agents take actions.
• AI Agent vs Chatbot vs Automation: Which Does Your Business Need? — Whether a job needs an agent or something simpler.
• How to Use Xero's JAX for Invoices and Cash-Flow Questions — More on Xero's JAX beyond reconciliation.
• Airtable vs Google Sheets for Business Data AI Can Use — Where to keep the data your agents read.
• Copilot Chat vs Microsoft 365 Copilot: What Do You Get Free? — What the paid Copilot licence adds, agents included.
• What Can AI Realistically Do for a Small Business? 15 Tasks — Fifteen jobs AI can realistically take on in a small business, rated by how much you can hand over, with tools, costs and the catch for each.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Meta for Developers WhatsApp pricing pages; HubSpot, Intercom, Help Scout and Airtable pricing and help pages; Xero's bank reconciliation page on JAX; Claude help 'Claude in Chrome permissions guide'; OpenAI help on ChatGPT Work; Microsoft 365 business plan comparison (all checked September 2026).