Skip to content

Dappasol / Guides

By , Founder · Updated November 2026

Small Business Automation Examples: 10 We Built and Measured

Quick answer. The small business automations worth copying are the boring ones: an instant reply to every new lead, a phone alert, a follow-up that stops when they reply, invoices that wait for a yes, a morning digest. Ten of ours, from the logs: nine still run, one we retired. Kits go out in a median 3 seconds, $0 added running cost.

Not sure where to start? The Automation ROI Calculator ranks your own tasks so you can see what to automate first.

Search for automation ideas and nearly every page you find is written by a company that sells the automation tool. The numbers are estimates, “save 5 hours a week”, with no log behind them.

This list is different in one boring way. Nine of these ten automations run our own business right now. The tenth we switched off on 12 November, and why we switched it off is one of the most useful lessons here. Every number below comes from our logs or our scheduler output. We pulled them on 12 November 2026. Nothing here is clever. That is the point: the automations that pay are the ones you forget exist.

What they share: they run on the same server as this website, they report to a phone, and none of them posts anything public or moves money without a person saying yes. Only one uses AI, and we will get to why. For the longer story of how the rest of the business runs on automation, read the build log. This page is the parts list.

Free, and from a human

Which one would I build first for you?

Send me your list of tasks, or your calculator list, and I will tell you which one I would build first and what to leave alone. Free, from a human: me, within 24 hours. No call needed.

Rather message? WhatsApp me your list or email ishan@dappasol.com.

The 10 automations at a glance

Every row is from our own logs and scheduler output on 12 November 2026. Added running cost means what each one adds to bills we already pay. Row 9 is the one we retired; it stays because what it taught us is worth more than what it did.

Seven columns: scroll sideways on a phone to see them all.

AutomationTriggerLogged runsSpeedHuman step leftAdded running costWhat broke
1. Instant lead replyForm submit43 real opt-insMedian 3 s, max 20 sNone$0Mail failed DMARC until moved to authenticated SMTP
2. Lead alert to phoneSame submitEvery leadMedian 2 s, max 18 sThe reply$0Server move: permissions stopped mail and alerts
3. Follow-up that stops on reply09:00 and 20:00, one run a day110 rows: 95 sent, 13 withheldOnce a dayEvery real reply$01 failure in 110, the server-move night
4. Recurring invoices28th preview, 1st send2 sent on 1 Nov20 s, whole runOptional hold reply$0Server move left its config behind; it stopped, not guessed
5. Morning digestEvery minuteDaily digest08:30 ISTDoing the thing$0Design rule: must-do re-pings stop at 5
6. AI traffic reportEach AI request5,729 hits in 15 daysWeekly, MondayReading it$0Count is a floor; upstream blocks unseen
7. Buyer-intent monitor08:30 and 18:30977 seen, 27 queuedTwice a dayAnswer or not$0Keywords are blunt; it only queues
8. Draft, then approveBuyer post foundEvery draft gatedTap polled every 30 sThe tap$0Two readers on one bot collide
9. Keep-alive pings (retired)Every 12 hRetired 12 Nov 2026Every 12 hNone, the problem$0Laptop slept 13 to 22 Oct; then outlived its purpose
10. Price-intake formPrivate linkOne catalogueSaves about 1 s after typingTyping$0Twice, edits said Saved and never arrived

The Founding Five, five spots

We're building two automations at $1,000, half the regular $2,000 Install price, in exchange for a case study with your real before and after numbers. Same fixed scope, live in 21 days. See the Founding Five.

10 small business automation examples, taken apart

Each one is a small business workflow automation in the plain sense: one trigger, a few steps, and one step deliberately left to a person. Each entry has the trigger, the steps, the logged number, what broke and the guard, and the same pattern elsewhere. In the flow diagrams, the coral box is the step a person keeps. No estimates: every number comes from our logs or our scheduler output.

1. Instant reply to a new lead: the kit lands in 3 seconds

Trigger: someone submits a form · Logged: 43 real opt-ins, 13 Jul to 11 Nov 2026 (19 test rows excluded) · Speed: median 3 s from submit to the email being sent, slowest 20 s (35 matched) · Human step left: none · Added running cost: $0

TRIGGERForm submitsomeone asks for akitAUTOMATICSave the leadplus first-touchsourceAUTOMATICAnswer thebrowserbefore any mailRESULTSend the kitauthenticated SMTPTRIGGERForm submitsomeone asks for a kitAUTOMATICSave the leadplus first-touch sourceAUTOMATICAnswer the browserbefore any mailRESULTSend the kitauthenticated SMTP
Entry 1: the lead is saved and the browser answered before the mail server is touched.

Someone asks for one of our free kits. Before the thank-you message has finished drawing, the email with the PDF is on its way: written by Ishan, signed by Ishan, with the one thing to do first and an easy way to reply. Across the 35 sends we could match to their form submission, the median gap was 3 seconds. The slowest was 20.

Two details make that number. The script saves the lead and answers the browser before it talks to the mail server, because sending mail can take 10 seconds or more and nobody should watch a spinner while SMTP thinks. And every lead records where it first came from: Google 22, direct or hidden 12, ChatGPT 2, Gemini 1, meta.ai 1, other 5. Thirty of the 43 asked for the same kit, which tells you more about what people want than any survey we could run.

What broke: deliverability, quietly. The site used to send through the server’s own mail function, from a machine no SPF record authorised, with no DKIM signature, so every message failed DMARC. Fast is worthless if it lands in spam. The guard: all mail now goes through the mailbox provider’s authenticated SMTP, and the lead is written to disk before any mail is attempted, so a mail problem can never lose a lead.

Same pattern in your business: a contractor’s quote request gets the price guide and two open site-visit slots straight away (AI for contractors). A clinic enquiry gets the intake form before anyone picks up the phone.

2. New-lead alert on your phone: 2 seconds, no inbox to check

Trigger: the same form submit · Logged: every lead · Speed: owner email median 2 s, slowest 18 s; the Telegram push fires earlier in the same request · Human step left: the reply · Added running cost: $0 (Telegram bots are free)

TRIGGERSame submitthe lead from entry 1AUTOMATICTelegram pushfires firstAUTOMATICOwner emailthe record copyHUMANYou replythe human partTRIGGERSame submitthe lead from entry 1AUTOMATICTelegram pushfires firstAUTOMATICOwner emailthe record copyHUMANYou replythe human part
Entry 2: the push skips email deliverability entirely. The reply stays human.

The same submission pings Ishan’s phone: what they asked for, which page the form was on, and where they first came from. The email copy is for the record. The push is for speed, and it skips email deliverability entirely, which matters after the story in entry 1.

Why bother with seconds: Harvard Business Review audited 2,241 US companies and found an average response time of 42 hours. Firms that tried to reach a lead within an hour were nearly seven times as likely to have a real conversation with a decision maker as those that tried even an hour later.

Telegram

New lead (magnet:automation) Email: jane@example.com Details: page=/guides/what-to-automate-first-in-a-small-business/ wa= marketing=yes Via: Google -> /guides/what-to-automate-first-in-a-small-business/ Form on: /guides/what-to-automate-first-in-a-small-business/ IP: 203.0.113.24
Telegram message, real format, sample data.

What broke: the night we moved servers, 30 October, a file-permissions bug stopped outgoing mail and the alerts. Every test had passed, because the tests ran as the admin user, not as the user the web server runs as. The guard: test as the web server’s user, and send yourself a real lead after every server change.

Same pattern in your business: an HVAC company’s after-hours enquiry pings the on-call phone instead of waiting for the morning inbox (AI for HVAC businesses).

3. A follow-up sequence that stops when they reply

Trigger: two calls a day, 09:00 and 20:00 IST, one real run per 20 hours · Logged: 110 rows, 5 Aug to 12 Nov 2026: 95 sent, 13 withheld for consent, 1 gap flag, 1 failure · Reach: 39 got day 1; 23 came due for day 3 and 11 were sent; 8 came due for day 7 and 7 were sent. The rest were withheld for consent · Human step left: every real reply · Added running cost: $0

TRIGGERTwo dailycalls09:00 and 20:00ISTAUTOMATICThrottlelockone run per 20hAUTOMATICDue andallowed?consent,suppress listRESULTSend day 3or 7or log itwithheldHUMANThey replya human answersTRIGGERTwo daily calls09:00 and 20:00 ISTAUTOMATICThrottle lockone run per 20 hAUTOMATICDue and allowed?consent, suppress listRESULTSend day 3 or 7or log it withheldHUMANThey replya human answers
Entry 3: two schedulers, one lock, and two rules that decide whether anything is sent at all.

Day 1 is the kit email from entry 1. Day 3 and day 7 come from a script that runs once a day, reads the lead log and the send log, and sends whatever is due. Run it ten times and it sends once. Skip three days and it catches up.

Two rules live inside it. Day 3 and day 7 are marketing, so they only go to people who ticked the marketing box, and anything unclear counts as no. That rule withheld 13 sends, and the log calls them withheld, not failed, which is the only reason we can trust the rest of the log. Then the suppress list: one address per line, and an address on it never gets another sequence email. A reply, an unsubscribe or a finished sequence puts you there. 29 of our 42 distinct lead emails are on it.

Of the 39 who got day 1, 23 came due for day 3: 11 of those were sent and 12 were withheld for consent. Of the 8 who came due for day 7, 7 were sent and 1 was withheld. The rest is the calendar and the suppress list: some replied, some have not reached day 7 yet. One row is a gap flag, a lead from a tool with no sequence behind it, logged as a gap instead of ignored. Nine leads got a personal reply from Ishan, logged by hand, because an answer to a real question should never come from a script.

What broke: one real failure in 110 rows, a day 1 email on the 30 October server-move night. The guard that matters most is boring: two triggers call it, 09:00 and 20:00 IST. A lock file allows one real run per 20 hours, so whichever comes second logs “skipped (throttle)”. In the log that is 59 runs and 21 throttled. If either scheduler dies, the other covers the day.

Same pattern in your business: a contractor’s quote follow-up on day 2 and day 6 that stops the moment the customer replies or books. A dental practice’s recall reminders that stop when the appointment is made (AI for dentists).

Free tool

Which of your own tasks would pay back first?

The Automation ROI Calculator ranks your tasks by what automating them would pay back, so you can see what to automate first before you spend anything. Then send me the list and I will tell you which one I would build.

Open the calculator →

4. Recurring invoices that wait for a yes

Trigger: the calendar, preview on the 28th, send on the 1st · Logged: 1 Nov 2026, 2 invoices emailed at 09:00 IST, the whole run 20 seconds · Human step left: one reply to hold a month, otherwise none · Added running cost: $0

TRIGGERThe 28thbuild nextmonth’s PDFsAUTOMATICPreview tophoneone captioneachHUMANHold?reply hold plusclientRESULTThe 1st,09:00number, email,copyAUTOMATICLedgerone line perinvoiceTRIGGERThe 28thbuild next month’s PDFsAUTOMATICPreview to phoneone caption eachHUMANHold?reply hold plus clientRESULTThe 1st, 09:00number, email, copyAUTOMATICLedgerone line per invoice
Entry 4: the owner gets three or more days to say “not this one”. Silence means send.

On the 28th the engine builds next month’s invoices as PDFs and sends each one to Ishan on Telegram with a one-line caption. Reply “hold” plus the client’s short name and that client skips the month. Say nothing and they go out on the 1st at 09:00 IST: numbered from one shared counter, emailed, the accountant copied, a copy back to Ishan, and a line written to the ledger. On 1 November the whole run, two invoices, took 20 seconds.

Telegram

Preview for 1 December 2026: acme USD 1,200.00. Reply: hold acme to skip, or hold all.
Telegram caption on the preview PDF, real format, sample data.

A second job checks for replies every evening from the 28th to the 31st and saves them. It exists because Telegram keeps unread bot updates for 24 hours at most. Without it, a hold sent on the 28th would be gone by the 1st.

What broke: the server move. On 28 and 29 October the scheduled checks on the new server logged “ERROR config.json missing”: the settings file had not come across. The engine stopped instead of guessing, the preview was rerun by hand on the 30th, and both invoices still went out on the 1st. Stopping is the design. If it cannot read Telegram on send day, it refuses to send and says so: “STOPPED send for November 2026: could not read Telegram for hold replies. Nothing was sent.” A late invoice is an apology. An invoice the owner told it to hold is a lost client. The guard people forget: invoices made by hand must take their number from the same counter, or two invoices end up sharing one. The engine also reserves the number before it sends, so a crash halfway cannot reuse it.

Same pattern in your business: retainers, memberships, monthly service contracts: anything billed the same way every month, where the owner occasionally needs to say “not this one”.

5. A morning digest and reminders that nag, politely

Trigger: a server job every minute · Digest: 08:30 IST, grouped by project · Must-do items: re-ping up to 5 times · Human step left: doing the thing · Added running cost: $0

TRIGGEREvery minuteserver cronAUTOMATICDue, or08:30?reminders,digestAUTOMATICPush tophoneDone, +1h,TonightAUTOMATICMust-do?re-ping up to 5timesHUMANYou do itthe only manualpartTRIGGEREvery minuteserver cronAUTOMATICDue, or 08:30?reminders, digestAUTOMATICPush to phoneDone, +1h, TonightAUTOMATICMust-do?re-ping up to 5 timesHUMANYou do itthe only manual part
Entry 5: one list, four screens, a job that checks it every minute.

One to-do list, everywhere: an iPhone app with a lock-screen widget, a web app, and Telegram. A server job runs every minute. It sends reminders when they fall due, rolls unfinished items forward, and at 08:30 IST sends the morning digest, grouped by project. Every reminder carries buttons: Done, +1h, Tonight, Tomorrow.

Telegram

☀️ Thursday, 13 Nov Clinic ops 🔸 Call the autoclave supplier · 10:00 • Send the March rota • Approve the recall email Open To-do
Telegram digest, real format, sample data.

The design choice: one ping gets swiped away with the rest of the lock screen, so must-do items re-ping up to five times and then stop. Anything that rolls forward three days in a row gets a question in the digest: reschedule, someday, or drop. A list that only grows is a list nobody opens.

Same pattern in your business: a restaurant’s opening checklist and supplier calls, or a clinic’s daily recall and callback list (AI for restaurants).

6. A weekly report on what AI assistants read on your site

Trigger: every request from a known AI agent is logged · Report: Mondays 09:00 IST, to Telegram · Logged: 5,729 hits in 15 days, 27 Oct to 12 Nov 2026, about 382 a day · Human step left: reading it · Added running cost: $0

TRIGGERAI botrequestany pageAUTOMATICClassifysearch, user,trainingAUTOMATICLog one lineplus AIreferralsRESULTMonday 09:00digest toTelegramHUMANYou read itand actTRIGGERAI bot requestany pageAUTOMATICClassifysearch, user, trainingAUTOMATICLog one lineplus AI referralsRESULTMonday 09:00digest to TelegramHUMANYou read itand act
Entry 6: one log line per AI request, one digest a week.

Every request from a known AI agent gets one log line: which agent, which page, why it came. In 15 days that was 5,729 hits: 5,474 bot fetches and 255 human visits that clicked through from an AI assistant. The bot side splits three ways: search crawlers 2,862, user-triggered fetches 1,469, training crawlers 1,143.

The column that matters is user-triggered. ChatGPT-User or Claude-User fetching a page means a person just asked an assistant something, and the assistant opened your page to answer. It is the only sign you get that an assistant used your page to answer, and your analytics never sees it, because bots do not run analytics scripts.

Telegram

AI traffic, dappasol.com, last 7 days (Nov 3 to Nov 10 UTC) Bot fetches: 2410 (train 520, search 1260, user-fetch 630) By agent: 560 PerplexityBot 410 ChatGPT-User 240 Claude-User AI referral visits (humans clicking through): 120
Telegram digest, real format, sample numbers.

What broke: nothing, but the count is a floor and the report says so at the bottom: some bots get blocked upstream by the host’s bot filter before they ever reach the log. The guard: print the blind spot inside the report, so nobody mistakes a floor for a total.

Same pattern in your business: a clinic that wants to know if ChatGPT shows its page for “dentist open Saturday”. One log line per AI request answers that, and the log costs nothing.

Recognise one of these from your own week? Tell us which one, and what it costs you now. No call needed. Ishan replies within 24 hours.

7. A buyer-intent monitor that reads the forums for you

Trigger: twice a day, 08:30 and 18:30 · Logged: 977 threads seen, 27 queued for a human · Typical run: 0 or 1 new · Human step left: deciding whether to answer · Added running cost: $0

TRIGGER08:30 and18:30twice a dayAUTOMATICRead publicfeednewest forumtopicsAUTOMATICDedupe,filterbuyers in,sellers outRESULTQueue0 or 1, mostrunsHUMANHumandecidesanswer or notTRIGGER08:30 and 18:30twice a dayAUTOMATICRead public feednewest forum topicsAUTOMATICDedupe, filterbuyers in, sellers outRESULTQueue0 or 1, most runsHUMANHuman decidesanswer or not
Entry 7: the machine reads every thread; a person reads the 27 that survived.

It reads the newest topics from a public forum feed (the Shopify Community, where store owners ask for help), skips anything it has seen before, throws out people selling themselves (“I built”, “for hire”), and keeps buyer phrases: “looking for someone”, “need help with”, “store not converting”, “migrate from”. What survives goes into a queue for a human.

Most runs report “0 new buyer-shaped threads”. That is normal, and it is the whole value: a machine looked twice a day so nobody had to. Across 977 threads, 27 made the queue.

The limit: keyword matching is blunt. It cannot tell “I need a developer” from “I need a developer’s opinion on this theme”. The guard: it never replies. It queues, and a human reads the thread before anything gets said.

Same pattern in your business: a wedding photographer watching local Facebook-style public boards for “anyone know a photographer”, or a law firm watching for questions in its practice area (AI for law firms). Public feeds only; private groups have rules for a reason.

8. Draft-then-approve replies: nothing posts without a tap

Trigger: a buyer-shaped LinkedIn post · Approval: Telegram buttons, Post it or Skip · Poller: checks for the tap every 30 s · Posting: on the next run, within 30 min · Human step left: the tap, every time · Added running cost: $0

TRIGGERBuyer postfoundon LinkedInAUTOMATICAI drafts areplyin Ishan’svoiceAUTOMATICCard tophonePost it or SkipHUMANThe tapchecked every30 sRESULTPostednext run,within 30 minTRIGGERBuyer post foundon LinkedInAUTOMATICAI drafts a replyin Ishan’s voiceAUTOMATICCard to phonePost it or SkipHUMANThe tapchecked every 30 sRESULTPostednext run, within 30 min
Entry 8: the only one of the ten with AI in it, and the one with the strictest gate.

This is the one with AI in it. When a hunt finds someone asking for help we actually give, an AI drafts a comment in Ishan’s voice and sends it to his phone as a card: who posted, what they want, the draft, two buttons. Post it, and it goes out from his LinkedIn on the next run, within 30 minutes. Skip, and it is dropped. Or he replies with his own wording and that version goes instead.

Telegram

LinkedIn intent #12 Jordan Lee, Founder, Example Dental Austin · 3h old · 4 comments Wants: someone to automate patient intake Open the post Draft comment Intake is usually the right first job. Before any tool, list which fields the front desk actually reads, then cut the rest. Happy to share the field list we start from. Post it = it goes out from your LinkedIn on the next run (within 30 min). Skip = dropped. Or reply to this message with your own wording.
Telegram card, real format, sample data.

What broke: two of our jobs read the same Telegram bot, and Telegram allows one reader at a time, so now and then they collide. The guard: a collision delays a tap by one poll, 30 seconds, instead of losing it.

Same pattern in your business: a restaurant’s replies to Google reviews, drafted by AI from the review text and approved by the owner with one tap. The AI saves the typing. The owner keeps the voice.

9. Keep-alive pings: the one we switched off

Trigger: a server cron every 12 hours · Why it existed: Supabase pauses free projects after a week of inactivity · Status: retired 12 Nov 2026 · Human step left: none, which turned out to be the problem · Added running cost: $0

TRIGGEREvery 12 halways-on serverAUTOMATICRun a real queryagainst the databaseAUTOMATICLog the resultone line per pingHUMANOwner, off datethe step it never hadTRIGGEREvery 12 halways-on serverAUTOMATICRun a real queryagainst the databaseAUTOMATICLog the resultone line per pingHUMANOwner, off datethe step it never had
Entry 9, retired: the coral box is the step it never had. A job with no owner keeps running after its reason is gone.

This is the one we turned off. Supabase pauses free projects after one week of inactivity (its pricing page says so), and a scheduled ping that runs a real query keeps one awake. We ran one every 12 hours for a project on the free tier.

What broke, first: the original version ran on a laptop. Between 13 and 22 October the Mac was asleep, the job never ran, and the warning came from Supabase, not from us. We moved it to an always-on server cron, kept the laptop as a backup, and took a rule from it: if a job can fail, the failure has to reach a phone. A log file is a diary, not an alarm.

What broke, second: the project it kept awake moved to its own site on another platform, and nobody told the job. For two weeks after it stopped mattering, it went on pinging every 12 hours: 28 pings in a row with no answer, written to a log nobody read. Nothing depended on it any more, which is exactly why nobody noticed. We switched it off on 12 November.

The guard: every automation now gets an owner and an off date, written down where it is scheduled: who reads its failures, and when someone asks whether it should still exist. Without both, a job outlives its purpose and fails quietly.

Same pattern in your business: booking widgets, free-tier forms, a promotion set up for one season, anything built for one client or one campaign. Give each job a named owner and a review date, and when the reason goes, switch the job off.

10. A private price-intake form that feeds the catalogue

Trigger: staff open a private link · Login: none, a token in the link · Human step left: typing the prices · Added running cost: $0 beyond hosting

TRIGGERPrivate linktoken, no loginAUTOMATICForm loadscatalogue withblanksHUMANStaff typeprices andspecsAUTOMATICAutosaveretry, historykeptRESULTCatalogueupdatedTRIGGERPrivate linktoken, no loginAUTOMATICForm loadscatalogue with blanksHUMANStaff typeprices and specsAUTOMATICAutosaveretry, history keptRESULTCatalogueupdated
Entry 10: the human types; everything around the typing is automatic.

A jewellery brand had a catalogue with holes in it: missing prices, missing specs. The usual fix is a spreadsheet emailed back and forth until someone gives up. Instead the staff got one private link. No account, no password: a token in the link is the login, because every extra step is a reason to stop. The form shows every piece with its blanks and saves as they type.

Saves go out about a second after typing stops, flush when the tab closes, and retry on failure with a red “Not saved yet, retrying”. Snapshots of earlier versions are kept, so nothing typed is lost to a bad connection.

What broke: twice, silently. Edits said Saved and never reached the server. The guard: test the round trip, save, reload, compare, before anyone types real data.

Same pattern in your business: any business collecting data from people who will not log in: a clinic’s pre-visit intake (cutting admin time in a private clinic), a contractor’s site-survey checklist, supplier price updates.

What ten automations taught us

Four things the logs say, in the order we would tell a friend:

  1. Speed to lead is solved, and cheap. Median 3 seconds to deliver, 2 seconds to alert, $0 to run. If your business still answers enquiries the next morning, this is the first thing to fix, and it is a small build.
  2. The reply is the hard part, and it stays human. All 43 got their kit in seconds. What happens next depends on a person answering a real question, which is why the nine personal replies were logged by hand, not sent by a script.
  3. Humans stay on anything public or money-shaped. Invoices wait for a possible hold, LinkedIn waits for a tap, the forum monitor only queues. The machine does the part that is the same every time. The person keeps the part that is reputation.
  4. Only one of the ten uses AI. Rules handle reminders, follow-ups, invoices and alerts better than a model: cheaper, predictable, easy to fix. AI earns its place where the input is messy, like drafting a reply to a stranger’s post, and that is exactly the one with the strictest human gate.

What should a small business automate first?

Rank candidates by three things: how often the task happens, how many minutes it takes, and what a day’s delay costs. In most small businesses that puts the same few jobs at the top. Our order, and the one we would suggest to almost anyone:

  1. The reply to a new enquiry (entry 1). Highest frequency, most time-sensitive, and being late costs money, not patience.
  2. The alert to whoever answers (entry 2), so the human reply comes in minutes, not tomorrow.
  3. The follow-up that stops on reply (entry 3). Most enquiries that go quiet were never followed up, not refused.
  4. Recurring invoices and payment reminders (entry 4), with a way to hold one.
  5. Appointment reminders and recalls: the same engine as entry 3, pointed at your calendar.
  6. Reporting: the weekly numbers someone assembles by hand, sent to a phone instead (entry 6’s pattern).

And four to leave with a person, however automatable they look:

  • Anything public in your name without a person approving it. Draft it with AI if you like; a human taps the button.
  • The first real answer to a real question. The machine sends the kit. A person sends the reply.
  • Money decisions: refunds, discounts, write-offs, a client who needs a month off.
  • A process nobody has written down. Automating a mess gives you a faster mess.

The full ranking method, with the arithmetic for what an hour of your time is worth, is in what to automate first in a small business. If AI agents will run any step, read what breaks when you automate with AI agents first. Or put your own tasks through the Automation ROI Calculator and get the order for your business, not ours.

The same ten in other businesses

None of the ten is specific to an agency. The trigger changes and the shape stays. Two versions per business, and where to read more:

BusinessFirst automationSecond automationRead next
Dental practiceRecall reminders that stop when the appointment is bookedPre-visit intake form, no loginAI for dentists
Contractor or rooferInstant reply to a quote request with two site-visit slotsQuote follow-up on day 2 and day 6 that stops on replyAI for contractors
HVAC companyAfter-hours enquiry pinged to the on-call phoneService-due reminders on a scheduleAI for HVAC businesses
RestaurantReview replies drafted by AI, approved with one tapMorning prep and supplier digestAI for restaurants
Law firmIntake form that pre-fills the conflict checkConsultation remindersAI for law firms
Online storeOrder-problem alert to a phoneRestock follow-up that stops on purchaseAI for ecommerce
Private clinicIntake before the visitDaily callback listAdmin time in a private clinic

How to automate a small business, and where AI fits

Six steps, in order. Every entry above went through them, usually after breaking at least once.

  1. Write the steps down first. If you cannot write the task as a checklist a new hire could follow, it is not ready to automate. Every entry above started as a list.
  2. Pick the trigger. A form, a time, a reply, a new row. Everything else hangs off it.
  3. Build it with rules. Add AI only where the input is messy: free-text emails, call notes, replies to strangers.
  4. Put a human on anything public or money-shaped. A tap on a phone is enough.
  5. Give it a second trigger or an alarm, an owner and an off date. Our follow-ups survive a dead scheduler because two things call them. The keep-alive in entry 9 had one caller, missed nine days while a laptop slept, then outlived its project by two weeks. And make the log tell withheld from failed: 13 withheld and 1 failed, out of 110, reads very differently from 14 failures.
  6. Test it as the system that runs it, then send yourself a real one.

Where AI fits. To automate a small business with AI, use the model for the messy input and rules for everything else. In our ten, AI drafts one thing, a reply to a stranger’s post, and a person approves it with a tap. AI is a tool here, not the point.

Tools. Zapier, Make and n8n all handle a form trigger, an email and a phone alert without code, and they are the right start if you do not run a server. Ours are small PHP and Python scripts on the website’s own server, scheduled with cron and reporting through a Telegram bot, because we already had both. If you want to do it yourself, do AI automation yourself, without hiring a team walks through the set-up.

What it costs

Running
About $0 a month added for the nine still running. They live on the server this website already pays for, Telegram bots are free, and the mail goes through a mailbox plan we already had.
Building it yourself
A no-code tool’s monthly plan, usually priced by the number of tasks, plus your evenings. Market prices for having it built are in our business automation cost guide and AI automation agency pricing. Comparing firms? The best business automation agencies for small business lists who does what.
AI Game Plan
From $500. We map your week, rank what to automate and what to leave alone, and the fee is credited 100% if you go ahead with a build. We find you a day a week to claw back, or the Game Plan is free. How it works.
The Residency
From $8,000. Ishan embedded in your business for 3 to 5 weeks, build included. What the Residency covers.
Not sure yet
Show me your Monday: a free 20-minute walkthrough. You talk through a normal Monday and get a one-page map of what to automate within 24 hours.

The honest version: if your list is short, do not hire anyone. Copy entries 1, 2 and 4 with a no-code tool this weekend: an instant reply to your enquiry form, a phone alert, and a monthly invoice that previews before it sends. That is most of the value on this page. Come back when the list gets longer than your evenings.

Keep going

FAQ

What are some examples of small business automation?

The ones that pay first are boring: an instant reply to every new enquiry, an alert to whoever answers, a follow-up that stops when the customer replies, recurring invoices with a way to hold one, appointment reminders, and a morning digest of what is due. This guide takes apart ten of ours, nine still running and one retired, with numbers from their logs.

What should a small business automate first?

The reply to a new enquiry. It is the highest-frequency, most time-sensitive task in most small businesses, and being late costs money. Ours delivers in a median of 3 seconds and adds nothing to the running bill. Next: the alert to whoever answers, then a follow-up that stops on reply, then recurring invoices.

How fast should a business reply to a new lead?

Within the hour at the outside, and in seconds if a machine can send the first reply. Harvard Business Review audited 2,241 US companies and found that firms trying to reach a lead within an hour were nearly seven times as likely to qualify it as firms that tried even an hour later. The average response time was 42 hours.

How much does it cost to run small business automations?

Ours add about $0 a month beyond hosting we already pay for: they run on our existing server, Telegram bots are free, and mail goes through a plan we already had. No-code tools charge a monthly plan, usually priced by the number of tasks. The real cost is building them properly and fixing them when they break.

Do I need AI to automate my small business?

Usually not. Only one of our ten uses AI, the one that drafts replies to strangers’ posts, and it has the strictest human approval. Reminders, follow-ups, invoices and alerts are rules, and rules are cheaper, more predictable and easier to fix than a model. Use AI where the input is messy.

Should I use Zapier, Make or n8n, or custom code?

Start with a no-code tool if you do not already run a server. Ours are small scripts because we already had a server and the skills. Move to code when per-task pricing starts to bite, when you need a rule the tool cannot express, or when the data should not leave your systems.

What breaks in small business automation?

Mostly silent things: mail that lands in spam, a scheduler that stops because a laptop slept, a permissions change after a server move, edits that never reach the server, and jobs that outlive their reason. The fixes are boring: authenticated mail, a second trigger or an alert on every job, testing as the system that runs it, logs that separate withheld from failed, and an owner and an off date for every automation.

Is it safe to let automation send invoices or post for my business?

Yes, with a human gate. Our invoices preview to a phone on the 28th and wait for a possible hold, and they refuse to send if the hold replies cannot be read. Nothing posts on LinkedIn without a tap on Post it. Automate the repetition and keep a person on anything public or money-shaped.

Have a list of tasks, or just a question about this? You don't have to book a call. Message us. Ishan replies within 24 hours.