12 Jobs AI Can Do for a Small Business Today (and 4 It Shouldn't)

Short version: AI is good at reading and drafting: receipts into line items, a rambling email into a reply, a month of numbers into a first-draft summary. It's bad at being the last word on money, people, or anything that goes out under your name. The pattern that works: AI reads the paperwork and drafts the reply; a person approves it; the software does the math.

Here are twelve jobs where that pays off today and four AI shouldn't do alone. For each: what it looks like, what to try this week with a general assistant you may already pay for (ChatGPT, Claude, Gemini, or Copilot), and when to build it into your own software. The bigger picture is in our practical guide to AI for owner-run shops; before pasting customer information anywhere, read what's safe to paste.

Front desk and inbox

1. Drafting replies to routine messages

Most of an inbox is the same dozen questions in different words: hours, prices, availability, "did you get my deposit?" AI drafts a reply in seconds; you fix it and send it yourself.

Try it now: paste the message and two or three of your past replies, and ask for a draft in the same tone.

Worth building in when the answer depends on your own records, like order status or open appointment times, so the draft arrives with the right facts filled in. Our AI customer service guide covers where that line sits.

2. Sorting and summarizing what comes in

Voicemail transcripts, web inquiries, service requests. AI can sort each one (urgent, routine, new lead, complaint) and summarize it in two lines, so your morning starts with a list, not a pile.

Try it now: paste a day's inquiries, names and numbers removed, and ask for a sorted list with one-line summaries.

Worth building in when it happens every day. Our own client portal does this with bug reports: Claude assigns a severity and category and writes a summary; an AI agent can draft a fix overnight, and nothing deploys until a person approves it.

3. Reading the replies to your texts

Ask fifteen people if they can work Saturday and you get fifteen kinds of answer: "yes," "let me check," "can't do Tuesday but Wed works." AI turns those into yes, no, or maybe.

Try it now: paste the thread and ask who's in, who's out, and who needs a follow-up.

Worth building in when you book people every week. In the production-company software we built, AI reads off-script replies like that last one, records them as yes, no, or needs time, and texts a short reply on its own. It's one of the few steps we let AI take without a click, and it's logged. Anything unclear goes to a person, not onto the schedule.

Paperwork

4. Reading receipts and vendor invoices

Photograph a receipt and AI pulls out the vendor, date, line items, tax, and total. A person checks them against the paper and approves.

Try it now: attach or paste a receipt and ask for a table of vendor, date, items, and total. Check it against the original.

Worth building in when someone keys in more than a handful a week, or each receipt has to land on the right job or grant. In Umbrella HQ, built for a nonprofit bookkeeping firm, AI pulls line items from receipts and budgets for a person to check. AI reading is cheap per item: on our own site, WillIFit, reading photos of 1,050 parking garages cost $51.73 in AI and image fees, about 5 cents each (more on AI costs).

5. Checking certificates of insurance and W-9s

Before a subcontractor starts or a rental goes out, someone should confirm the paperwork is current. AI can read a certificate of insurance or a W-9, pull out expiry dates and coverage, and flag what looks wrong.

Try it now: attach the certificate, list your requirements, and ask what's missing or expired.

Worth building in when nobody should start without current coverage on file. In Mr. Camera HQ, the software that runs my production company, Claude reads uploaded certificates of insurance and W-9s and flags anything that needs attention for the office to check.

6. Pulling the key terms out of long documents

Award letters, leases, vendor contracts: twelve pages, and you need six facts. AI pulls the dates, amounts, deadlines, and obligations into a short list.

Try it now: attach the document, ask for that list, then check each item against the page. AI can miss a clause or misread a number.

Worth building in when those terms drive something: a budget, a reporting deadline, a reminder. Umbrella HQ extracts grant terms from award letters for a person to confirm. More on this pattern in AI reads the paperwork, you approve it.

Quotes and sales

7. Turning a plain request into estimate line items

In Mr. Camera HQ, the software that runs my production company, I type "add two camera ops for three days" and the AI assistant turns it into estimate line items. It reads uploaded vendor bids into the estimate too. The software, not the AI, does the totals.

Try it now: paste your price list, describe the job, and ask for draft line items. Redo the math yourself.

Worth building in when you quote often and your rates live in one place. Stickman HQ, built for another Las Vegas production company, uses AI to draft estimates too.

8. Drafting listings and proposals in your own voice

Show AI three of your past proposals or listings and it can draft the next one in the same style. It's a first draft, but it beats a blank page.

Try it now: paste your best past examples plus the facts of the new one, and ask for a matching draft.

Worth building in when several people write the same kind of document. Our working prototype for a commercial real estate brokerage drafts listings in each broker's style from their own past listings; a second AI pass flags typos, gaps, and inconsistencies, and the broker reviews and submits.

Marketing

9. Posts, newsletters, and website text

Tell it what happened this month (a new service, a finished project, holiday hours) and ask for a short post or newsletter. It's quick at first drafts and at trimming long ones.

Try it now: paste your notes and one past post you liked, and ask for three short versions.

Worth building in? Usually not. A general assistant does this well, and custom software won't improve your newsletter.

10. Drafting replies to reviews

A calm, specific reply to a one-star review is hard to write when you're annoyed. AI drafts it level-headed; you add the details only you know.

Try it now: paste the review and what happened, and ask for a short reply that addresses the problem and offers to talk directly. Add no customer details that weren't already public.

Worth building in? Rarely, unless you answer dozens a week across several locations.

Operations

11. The monthly numbers, in plain English

A board or a partner wants a few paragraphs, not a spreadsheet: what came in, what went out, what changed. For a performing-arts nonprofit, we made a financials app that's built to draft a 150–250-word monthly summary for the accountant to edit and approve, from that month's figures only, before the board sees it.

Try it now: paste a month's profit-and-loss summary and ask what changed from last month, in plain English. Check every number it quotes.

Worth building in when the numbers already live in your software, so nothing gets retyped.

12. Writing down how things are done

In many family businesses, the procedures live in one person's head. Talk through how you close the month or open the shop, and AI can turn the transcript into a checklist someone else can follow.

Try it now: record a voice memo walking through one task, paste the transcript, and ask for numbered steps plus the questions a new person would ask.

Worth building in when people follow the checklist daily; then it can become a screen in your software instead of a binder. More in getting operations out of one person's head.

Four jobs AI shouldn't do on its own

1. Math that moves money, unless code checks it

AI misreads numbers more often than you'd expect, and its arithmetic isn't reliable. On WillIFit, a $5 test using only the cheapest model got 4 of 11 sign readings wrong; it read a speed-limit sign as a clearance height. The fix was rules: the cheap model only decides whether a sign is there, a stronger model reads it, and code rejects any number that doesn't match the model's own transcription.

The rule in everything we build: the model transcribes the numbers; the server does the arithmetic. Totals, taxes, invoices, and payroll come from ordinary code that gives the same answer every time. One payroll-allocation tool we built for a nonprofit uses no AI at all. It won't proceed until the imported totals tie out, and its journal entry has to balance to the penny.

2. Final decisions about people

Hiring, firing, discipline, who gets credit or a payment plan. AI can summarize applications or tidy interview notes. It shouldn't rank, reject, or decide. Those calls need a person who can explain the reasoning to the applicant, the employee, and, if it comes to that, a lawyer.

3. Anything sent to a customer that nobody read

AI writes confidently even when it's wrong: the wrong date, a policy you don't have, a discount you never offered. Drafting is its job; sending isn't. That's how we build AI into paperwork and anything touching money: the broker submits the listing, the board summary is built to wait for the accountant's approval, and in our own client portal nothing deploys until a person approves it. The few steps we let run on their own are small and logged, like a short text back to a crew member who replied off-script, and anything unclear goes to a person.

4. Regulated data on a consumer account

Patient health information, card numbers, Social Security numbers, and anything a client contract says stays private don't belong in a personal chat account, and some of it doesn't belong in any general chat tool. Which plans train on your data, and which offer a HIPAA business associate agreement: Is it safe to put customer data into ChatGPT or Claude?

Where to start

Pick the one job that eats most of your week and try it with the assistant you already have for two weeks. If it helps and the volume is real, that's when building it into your software pays.

Frequently asked questions