What Can AI Realistically Do for a Small Business? 15 Tasks

Realistically, AI can draft most routine writing, summarise meetings, answer common customer questions, pull details from invoices and forms, suggest bookkeeping categories, analyse spreadsheets and translate messages. It does these at first-draft quality: fast and usually right, but a person still checks anything that reaches a customer, a patient or your accounts.

The useful question isn't whether AI can do a task, but how much of it you can hand over. Some jobs can run with a weekly spot check; others will only ever produce drafts. That hand-over level matters more than which tool you pick, because it decides how many staff hours you actually get back. Every task below carries one of three levels:

Follow me on Instagram@sagnikteaches

• Draft: AI writes a first version; a person edits it and decides what happens next.
• Check: AI does the whole job; a person approves each item before it counts.
• Spot-check: AI runs on its own; a person reviews exceptions and a weekly sample.

How much of each task you can hand over

Notice how few tasks reach spot-check. That's the realistic picture for 2026: AI removes the blank page and the retyping, while people keep the final say. If you want quick set-ups rather than an overview, twelve AI quick wins you can set up this week is the practical companion to this list.

Connect on LinkedInSagnik Bhattacharya

Writing and communication

1. Draft emails and replies in your own voice

AI is at its most dependable turning a few rough notes into a clear, polite message. It gets the structure and tone right almost every time; what it can't know is the fact you didn't give it. An illustrative prompt from a dental practice's reception, and the draft it produced:

Subscribe on YouTube@codingliquids

Prompt: Reply to a patient who asked if we can move her hygiene
appointment from Tuesday 9am to later that week. We have Thursday
2:40pm or Friday 11:10am. Warm, short, no more than 80 words.

Draft: Hi [first name], thanks for letting us know. We can move your
hygiene appointment to Thursday at 2:40pm or Friday at 11:10am. Just
reply with the one that suits and we'll confirm it. If neither works,
tell us a couple of times that would, and we'll do our best to fit
you in. Kind regards, [practice name] reception

What to fix: nothing in the wording, but the receptionist still checks the diary before sending, because the slots were typed from memory. Tools: Gmail's "Help me write" can match the tone of your past emails on Business Starter, Standard and Plus and on Google AI plans; Microsoft 365 Copilot Chat, included with business plans, now works on Outlook mail. Watch-out: prices, dates and promises in drafts, which the AI will state confidently whether or not they're true. Hand-over level: draft.

2. Write marketing posts and newsletters

AI produces serviceable social posts, email newsletters and web copy quickly, and it's good at variations: five subject lines, three lengths, a version for each platform. Left alone, it writes like everyone else. A before-and-after from a physiotherapy clinic shows the difference a few specifics make:

• Before (generic prompt): "Discover a better you this spring! Our expert team is here to help you move better and feel your best."
• After (prompt with details): "Running your first 10K this spring? Book a 30-minute running assessment before you increase your mileage. Wednesday evenings, from 5pm."

The second version came from giving the AI the service, the audience, the time slot and a rule against slogans. Tools: any business chat plan; Canva's AI features draw on a monthly allowance (the Free plan gets 200 Standard or 20 Premium uses); Klaviyo's marketing agent, Composer, drafts email and SMS campaigns for review. Instagram now caps hashtags at five per post, so tell the AI that too. Watch-out: health and results claims, which clinics in particular must check against their professional body's advertising rules. Hand-over level: draft.

3. Translate messages and documents

For everyday messages, machine translation is now good enough that a customer can read your reply comfortably. For anything with legal, medical or financial consequences, it's a draft that a fluent person should check. A veterinary practice might translate its post-operative care sheet into the two other languages its clients most often speak, then pay a professional to review the result once, because an ambiguous instruction about medication is costly. Tools: DeepL's plans now run Free, Individual ($8.74), Team ($28.74) and Business ($57.49 per user a month on annual billing); for human review, Gengo lists rates of $0.09 and $0.18 a word. At $0.09 a word, checking a 600-word care sheet costs $54. Watch-out: dosages, dates and negatives ("do not"), which are the details that go wrong and matter most. Hand-over level: check.

4. Draft policies, procedures and job adverts

AI is a fast first drafter for the documents small businesses put off: a lone-working procedure, a social media policy, a job advert, a complaints process. Give it your real practice and it produces something usable in minutes. A nursery manager might ask for a drop-off and collection procedure, then fill in the parts only the nursery knows:

Collection procedure (draft, filled in)
Who may collect: named adults on the registration form only
Password: set by parents at registration; asked every time
  a non-named adult collects
Late collection: key person calls parents at 15 minutes late,
  then emergency contacts at 30 minutes
Record: time, collector's name and staff initials in the daily log
Reviewed by: nursery manager, every September

Tools: a shared Project in ChatGPT or Claude (Claude's Free plan allows up to five projects; custom instructions need a paid plan) keeps your house style and past policies to hand; Copilot in Word can produce first drafts on Microsoft 365 business plans, with Microsoft labelling access as Basic or Premium depending on the licence. Watch-out: legal and employment content. Have an adviser check anything contractual, and if AI ever screens job applicants, recruitment counts as a high-risk use under the EU AI Act for businesses hiring in the EU. Hand-over level: draft.

Customers and appointments

5. Answer routine customer questions

A chatbot or messaging agent can answer the questions you get every day (opening hours, prices, parking, what to bring) and hand anything else to a person. The realistic ceiling depends on how good your written answers are. Most failures come from gaps and outdated pages in the knowledge you gave it, not from the AI inventing things from nothing.

An illustrative exchange on a nursery's website:

Parent: Do you have a place for my 14-month-old from January?
Bot: I can't see live availability, so I don't want to guess. I've
passed your question to the nursery manager, who replies within
one working day. Could you share your child's start date and the
days you need?

That's a good answer, because the owner told it never to confirm availability. Tools and costs: Intercom's Fin charges $0.99 per resolved outcome; Help Scout's AI Answers is $0.75 per resolution on paid plans only; HubSpot's Customer Agent needs a Professional or Enterprise hub and uses 50 credits (about $0.50) per resolved conversation; Shopify Inbox includes a free AI agent. A quick sum: 300 resolved questions a month costs about $297 on Fin or $225 on Help Scout, before seat fees. Watch-out: anything the bot states as fact about availability, prices or medical matters, and a hand-off route customers can actually find. If you sell to customers in the EU, tell them they're talking to AI. Hand-over level: spot-check, once transcripts show it answers well. The wider picture is in what to automate first in AI customer service.

6. Send reminders and handle rebooking messages

The reminders themselves are usually a practice-software feature, not AI. Where AI helps is in timing and wording them, and in answering the replies. A podiatrist whose patients often reply "can I move it?" to a reminder text can have an AI phone or message service take those requests out of hours and queue them for the morning. Tools: Smith.ai lists its AI Receptionist as free for 25 calls, then $150 for 75; Jobber's AI Receptionist, aimed at trades, costs $29 for 30 conversations and $0.79 each after that. Watch-out: an AI that changes the diary without checking clashes, and texts that show clinical details on a lock screen. Hand-over level: spot-check for fixed-wording reminders, check for rebooking changes.

An illustrative overnight queue shows the split. The AI logged three replies to Tuesday's reminder texts: "Can I move Thursday to next week?" (tagged rebooking request), "Running 10 minutes late tomorrow, sorry" (tagged note for the diary) and "Is the car park open during the roadworks?" (tagged question for reception). Reception cleared all three by 9:15 the next morning. The AI didn't change a single booking; it saved the time spent reading, sorting and finding the right record.

Paperwork and money

7. Pull details out of invoices, forms and PDFs

Reading a supplier invoice or a typed form and pulling out the supplier, date, total and reference is now routine for AI, and it's reliable on clean, printed documents. Handwriting, photos taken at an angle and multi-page scans are where it slips. A veterinary practice sending 40 supplier invoices a week through its accounting software's capture tool might find that 36 arrive correct and four need a fix, usually a credit note read as an invoice. Tools: Xero's Smart Document Capture, free on all Xero plans; Dext's Business plan at $302.50 a year for 250 documents and five users; Claude reads PDFs up to 1,000 pages for one-off jobs. Watch-out: totals and tax lines, which you should reconcile rather than trust. Hand-over level: check at first, spot-check once accuracy holds for a month.

8. Suggest bookkeeping categories and bank matches

Accounting software now suggests how to categorise transactions and which bank lines match which invoices. The suggestions improve as it learns your patterns, but they're suggestions. Here's the kind of mistake to expect, illustrated with an osteopathy clinic: the software learned that payments to a particular supplier were "clinic consumables", then kept applying that to a one-off equipment purchase from the same supplier, which should have been treated differently. Nobody noticed until the year-end review. Tools: Xero's JAX automatic reconciliation comes on Growing plans and above (Xero still labelled it beta in its August 2026 update); QuickBooks presents its AI as Intuit Intelligence, with a $10 a month add-on for extra prompts, and Intuit says its AI features can't currently be switched off one by one. Watch-out: unusual transactions from familiar payees. Hand-over level: check, with your bookkeeper or accountant reviewing monthly.

9. Sort a shared inbox

AI can label incoming email by type (enquiry, booking change, invoice, complaint), summarise long threads and route messages to the right person. It's one of the few tasks that reaches spot-check, because a misrouted email is usually caught by the person who receives it. Tools: Copilot Chat in Outlook and Gemini in Gmail for individual inboxes; for rules across a shared inbox, an automation tool with an AI step. Zapier's AI step uses 1, 3 or 5 tasks per run depending on the model tier, while its filter steps use none, so route the obvious messages by rule and send only the ambiguous ones through AI. Watch-out: urgent messages labelled as routine. Keep an "unsure" label that a person clears twice a day. Hand-over level: spot-check.

A label set that works for a small clinic has six options and no more: new enquiry, booking change, invoice or bill, complaint, clinical question (to a clinician, never answered by AI) and unsure. More labels than that and the AI starts splitting hairs, the unsure pile grows and staff stop trusting it. The misroute to watch for is a complaint phrased politely ("just wanted to mention that the last invoice seemed high"), which tends to land under billing; add that exact example to the instructions once you've seen it happen.

Meetings, records and research

10. Turn meetings and calls into notes and actions

Meeting summaries are good now: who said what, the decisions and a list of actions. The common error is attribution, with an action assigned to the wrong person or a "we should" recorded as a decision. A physiotherapy clinic's weekly team meeting might get a summary within minutes; the practice manager spends two minutes correcting owners and deadlines before it goes out. Tools: Zoom retired its "AI Companion" brand in June 2026, and features such as meeting summary remain included on eligible paid Zoom Workplace plans; Google Meet's "Take notes for me" needs Business Standard or above; Teams intelligent recap needs Teams Premium or a Microsoft 365 Copilot licence; Otter Pro costs $8.33 a user a month billed annually ($16.99 monthly). Watch-out: recording people without telling them, and confidential discussions sitting in a vendor's cloud. Hand-over level: check.

The corrections are usually small but they matter. An illustrative action list from that clinic meeting read: "Order new treatment couch covers ([first name]); update the website opening hours (team); review the cancellation policy (practice manager, next week)". Two fixes before sending: the couch covers were agreed to be the practice manager's job once prices were in, and "team" is nobody, so the website change got a named owner and a date. A summary that nobody corrects quietly turns into a record of decisions that were never made.

11. Turn dictation into records and reports

Speaking notes and letting AI structure them into a record is one of the biggest time savers for clinicians and field workers. An osteopath might dictate for 90 seconds after a session and get a structured note to review, rather than typing for five minutes. Tools: SimplePractice's Note Taker costs $35 per clinician; Heidi says it doesn't keep audio; Nabla discards audio and keeps transcripts for 14 days by default. Vendors change these settings: since 16 June 2026, new SimplePractice Note Taker users are opted in by default to keeping de-identified transcripts. Watch-out: consent from the patient or customer, and a clinician reading every note before it's signed. Hand-over level: check, always.

12. Research suppliers, competitors and rules

AI with web search can pull together a comparison of suppliers, summarise what competitors charge publicly, or explain a new rule in plain English. It's fast and often right, and it's wrong in a particular way: it cites a page that says something slightly different, or an old version of it. An illustrative slip: a nursery asked for the current food-hygiene guidance for early-years settings and got a tidy summary built partly from a withdrawn document. The fix is to open every source it cites before acting. Tools: Perplexity Pro at $20 a month; the web search built into ChatGPT, and into Claude's Free plan. Watch-out: dates on sources, and summaries of legal requirements, which need checking against the official text or an adviser. Hand-over level: draft. How often each kind of task goes wrong is covered in how accurate ChatGPT is by task.

Numbers, images and multi-step work

13. Analyse a spreadsheet and explain the numbers

AI can read an export, spot patterns, build a pivot table and explain what changed in plain English. It's also capable of stating a wrong total with complete confidence, so ask it to show the formula or the rows behind every number. An illustrative exchange from a dental practice:

Prompt: This sheet lists last quarter's appointments with columns
date, type, clinician, status. Which appointment types had the
highest cancellation rate? Show the count behind each rate.

Reply: Emergency slots had the highest rate, 22% (11 of 50),
then hygiene at 14% (38 of 271). Check-ups were lowest at 6%
(19 of 317). Note: 12 rows have a blank status and I left them out.

The last line is the useful one: 12 blank rows is a data problem to fix before trusting any rate. Tools: Copilot in Excel works with a business subscription that's eligible for Copilot Chat; Excel's =COPILOT() worksheet function was retired on 14 September 2026, so use the side pane; Google Sheets has an =AI() function that fills 350 cells at a time and doesn't refresh on its own. The fuller method is in analysing data with Copilot. Watch-out: totals, averages of averages and silently dropped rows. Hand-over level: check.

14. Make images and simple graphics

Social graphics, website banners, simple illustrations and product mock-ups are well within reach. Photographs of your actual premises, staff and results are not, and passing off generated images as real ones misleads customers. Tools: Canva's AI allowance on its plans; Midjourney, whose default model has been V8.2 since 24 July 2026; Adobe Firefly, where Adobe's own models are trained on licensed content but partner models in the app aren't covered, and IP indemnity is enterprise-only. If you list products on Google Shopping through Google's merchant tools, AI-generated product images must keep their IPTC "TrainedAlgorithmicMedia" metadata. Watch-out: hands, text inside images and anything that looks like a real before-and-after. Hand-over level: draft.

A prompt that gets usable results for a nursery's autumn open-day post is specific about what to leave out: "Flat illustration of a bright nursery room with wooden toys, an easel and autumn leaves on a window, soft colours, no people, no text, no logos, square format." Leaving out people avoids generated children's faces, which parents can find unsettling, and leaving out text avoids garbled lettering. Add the date and wording yourself in the design tool, where you can spell-check it.

15. Run multi-step tasks across websites and apps

Agents can now carry out a sequence of steps: open several websites, collect information, fill a spreadsheet, draft a document. They're slower and less predictable than a person, and they need watching when they act rather than just read. A podiatrist might ask an agent to collect opening hours, delivery charges and minimum orders from five dressing suppliers into a table: a low-risk, read-only job that saves an hour. Tools: ChatGPT Work, which replaced ChatGPT agent on 9 July 2026 and is on Plus, Pro, Business and Enterprise, asks for approval before important actions by default; Claude in Chrome became generally available on 26 August 2026, and Anthropic's own guidance says to avoid using it on financial accounts, health information and sites holding other people's personal information. For recurring jobs, ChatGPT's scheduled tasks allow 5 active tasks on Plus and 10 on Business. Watch-out: an agent logged into accounts where one wrong click costs money. Hand-over level: check every action.

Picking your first two from the fifteen

Don't start with the most impressive task. Start where your hours go. For one ordinary week, have everyone note the time they spend on each of these fifteen jobs, then rank them by hours multiplied by hand-over level: a draft-level task saves perhaps half the time, a check-level task more, a spot-check task most.

A 12-person nursery's tally (illustrative) looked like this: parent emails 9 hours, policy and procedure updates 4 hours, newsletters 3 hours, supplier invoices 2 hours, staff meeting notes 1.5 hours. Parent emails won easily, and the second pick was invoices rather than newsletters, because invoice capture reaches check level and was free in the accounting software the nursery already used. Policies stayed on the list for a quiet month.

Two rules keep the choice sensible. Leave out any task where a mistake could hurt someone until you've run something lower-risk for a month, and check the list of tasks to keep human before automating anything that involves judgement about people. Then test your first task properly with the method in choosing your first AI project.

Questions about what AI can take on

Which of these tasks is safe to leave running without checking?

Very few, and only after a trial. Sorting a shared inbox and sending fixed-wording reminders can run with a weekly spot check once they've proved accurate. Anything that states a fact to a customer, touches money or records health information should keep a person approving each item, at least until you have weeks of evidence that errors are rare and caught.

Do I need a separate AI tool for each task?

Usually not. A business chat plan covers drafting, research, translation and spreadsheet questions, and the AI built into your email, accounting and booking software covers several more. Separate tools earn their place for high-volume jobs such as customer chat or document capture, where the specialist product's integrations save more time than they cost.

What can't AI do for a small business yet?

It can't know your unwritten rules, take responsibility for a decision, or be trusted with arithmetic it hasn't shown. It struggles with messy handwriting, judgement calls about people, and anything where the right answer depends on context it wasn't given. Keep legal, disciplinary and clinical decisions with a person, and use AI to prepare the paperwork around them.

Further reads

• AI Workflow Automation Examples: 20 Processes to Automate First — Twenty processes to automate once you've picked your tasks.
• The Limits of AI: What It Still Gets Wrong in a Small Business — A closer look at what AI still gets wrong.
• AI Agent Use Cases for Small Businesses: 10 Worked Examples — Ten worked examples of the multi-step agent tasks in number 15.
• 10 Admin Tasks a Small Clinic Can Hand to AI This Month — The clinic-specific version of this list.
• Small Business AI Statistics 2026: Adoption, Spend, and Results — How many small businesses use AI for these tasks, and what they spend.
• Is My Business Too Small for AI? What Works for Teams of 1 to 10 — Which of these tasks pay off for a team of one to ten.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: facts verified on vendor pages for this series in September 2026, including Google Workspace (Gmail Help me write, Meet notes, Sheets AI function), Microsoft Learn and Support (Copilot Chat, Teams recap, Copilot in Excel, =COPILOT() retirement, scheduled prompts), Zoom's AI feature naming, Otter, Intercom, Help Scout and HubSpot pricing and help pages, Smith.ai and Jobber pricing, Xero and Dext pricing, Canva's AI allowance page, Midjourney release notes, DeepL and Gengo pricing, SimplePractice, Heidi and Nabla documentation, OpenAI help pages (Projects, scheduled tasks, ChatGPT Work), Anthropic's Claude in Chrome safety page, and Zapier's task rules.