AI for Small Business: A Practical Guide for Owner-Run Shops

Short version: AI is useful today for the reading and writing chores of a small business — answering the same customer questions, drafting replies, pulling dates and totals off paperwork, and turning your notes into a first draft. It isn't ready to be left alone with anything that matters. The rule we build by: AI reads and drafts, a person approves, and the software does the math. To start this month, pick one repetitive job, use the AI you already pay for, and have a person check every result for 30 days.

I run Mr. Camera, the Las Vegas production company my father founded in 1981, and I've built AI into its software and into tools for other small businesses and nonprofits. This is what I'd tell a friend who owns a shop and keeps hearing they should be "doing AI" — including where the honest answer is "don't bother yet."

Where AI helps a small business today

For each job below: what the AI actually does, the cheapest way to try it, and when custom software starts to make sense. The longer version, including the jobs AI shouldn't touch, is 12 Jobs AI Can Do for a Small Business Today (and 4 It Shouldn't).

• Answering common customer questions
  • What AI does: answers the hours, prices, parking, and "do you take walk-ins" questions from a short document of your real policies, and hands anything unusual to a person.
  • Start with: paste your policies into a chat assistant and have it draft answers you copy into replies. See what it gets right before it talks to anyone directly.
  • When custom makes sense: when the answer lives in your records — open slots, order status, a customer's balance — and the assistant should look it up, not guess. More in AI customer service for small business.
• Drafting replies to email and DMs
  • What AI does: reads the message and drafts a reply in your tone. You edit and send.
  • Start with: the AI built into your email, if your plan has it, or a chat assistant given a few of your past replies as examples.
  • When custom makes sense: when every reply needs facts from your records — a quote, a booking, a job status — and you're retyping them into each draft.
• Responding to reviews
  • What AI does: drafts a thank-you for good reviews and a calm first draft for bad ones.
  • Start with: paste the review into a chat assistant, ask for three short options, and rewrite the best one in your voice. Read any reply to a complaint twice.
  • When custom makes sense: almost never. Review volume rarely justifies more than a chat assistant.
• Reading receipts and invoices
  • What AI does: reads the vendor, date, line items, and total from a photo or PDF into fields a person checks.
  • Start with: see whether your bookkeeping software already offers receipt capture before you buy anything.
  • When custom makes sense: when receipts belong to a job, a grant budget, or a client invoice your bookkeeping app doesn't track, and someone retypes them every week.
• Pulling dates from insurance certificates and contracts
  • What AI does: reads expiration dates, coverage amounts, and named parties off a certificate of insurance or contract, and flags what's missing or lapsing.
  • Start with: upload one document to a business chat plan, ask for the dates and limits as a list, and check the list against the paper.
  • When custom makes sense: when you track dozens of vendors, subcontractors, or renters and need a warning before coverage lapses.
• Drafting quotes from notes
  • What AI does: turns "two techs, three days, plus the lift rental" into draft line items.
  • Start with: give a chat assistant your price list and your notes and have it lay out the line items. Do the totals yourself or in your spreadsheet, not in the chat.
  • When custom makes sense: when quotes become jobs and invoices, and the line items should use your real rates with software doing the arithmetic.
• Monthly financial summaries
  • What AI does: writes a short, plain-English summary of the month's numbers for an owner or a board.
  • Start with: give it the finished report your bookkeeper already produced, not raw bank data. Check every number it mentions against the report.
  • When custom makes sense: when the summary should be drafted from your system's own figures each month and approved by your accountant before anyone sees it.
• Turning walkthrough recordings into procedures
  • What AI does: turns a recorded walkthrough — you narrating how you close out the register or open the shop — into numbered written steps.
  • Start with: record yourself doing the job, get a transcript, and have a chat assistant turn it into a checklist. Then watch the person who'll use it follow it once.
  • When custom makes sense: rarely. This is about getting knowledge out of one person's head, which is where taking over a family business starts.
• Product and service descriptions
  • What AI does: drafts descriptions for your menu, website, or listings from your notes and a few examples you like.
  • Start with: give it three descriptions you're happy with plus the facts for the new one. Check every claim; AI will invent features if the facts are thin.
  • When custom makes sense: when you write many in a consistent style. We built a working prototype for a commercial real estate brokerage that drafts listings in each broker's style, using their past listings as examples; an AI reviewer flags typos and gaps, and the broker reviews and submits.

Notice what isn't on the list: letting AI make calls about money or customers with nobody checking. Where we do let AI act on its own in software we've built, the step is small, it's logged, and anything unclear goes to a person. More on that below.

Start with the AI you already pay for

Before you buy anything new, check what you already have. Many email, office, and bookkeeping tools have been adding AI features, sometimes inside plans you're already on, with no new account or bill.

Two examples, as of October 2026:

• Google Workspace's pricing page lists the Gemini assistant in Gmail on its Business Starter plan, and in Gmail, Docs, Meet, and more on Business Standard and Business Plus.
• Intuit's QuickBooks Online AI page lists AI features — it calls them AI agents — across its paid plans, with different ones on different tiers.

These change often, so check the vendor's own page for your plan before counting on a feature.

If nothing you have fits, the next step is a business plan from one major chat assistant. For scale, Anthropic lists a standard seat on Claude's Team plan at $25 per seat per month billed monthly, or $20 billed annually, as of October 2026. Pick one assistant, not three; one tool used daily teaches you more than three tried once.

The rule: AI reads and drafts, a person approves, the software does the math

This is the rule we build by, and it works even if you never hire anyone:

• AI reads and drafts. It's good at pulling the date off a certificate, the total off a receipt, and the point out of a rambling email, and at turning notes into a first draft.
• A person approves. In paperwork and anything touching money — a quote, an invoice, a summary for the board — AI drafts and a person approves before it counts or goes out.
• The software does the math. Totals, taxes, splits, and anything that moves money are handled by ordinary code that gives the same answer every time. The model transcribes the numbers; the server does the arithmetic.

A few steps run on their own and are logged — sorting bug reports, and in our production-company software, reading a crew member's off-script text as yes, no, or needs time and sending a short reply — with anything unclear going to a person. Those are deliberate exceptions: small jobs with only a few possible answers, where a quick response matters and a person can check the log.

Here's why I'm strict about it. WillIFit, my own site that lists clearance heights at parking garages for RV and truck drivers, uses AI to read clearance signs in Google Street View images. In an early $5 test that used only the cheapest model, it wrote 4 of 11 readings wrong. One was a speed-limit sign it read as a clearance height. For a driver deciding whether a rig fits under a garage entrance, that's not a typo. It's a hazard.

The fix wasn't a cleverer prompt; it was structure. The cheap model now only detects whether a sign is visible. A stronger model reads it, and the number must match the model's own transcription or it's thrown out. Later, when a "high-confidence" reading disagreed with an earlier reading of the same sign, about half turned out wrong — so conflicting readings now wait for a person. Readings that pass the checks go live on their own.

The same pattern works in any small business: let AI read, check it with simple rules where you can, and put a person on anything that matters. How we build that into client software is in AI Reads the Paperwork, You Approve It.

What it costs

There are three ways to pay for AI; most small businesses only need the first two:

1. The AI in tools you already pay for. Sometimes included, sometimes an add-on. Check your plan.
2. A chat assistant subscription. A monthly fee per person, like the Team plan price above.
3. AI inside software, billed by use. When software calls an AI model directly, it pays per token — roughly, per word — in and out.

For that third kind, Anthropic's published prices per million tokens, as of October 2026, are: Claude Haiku 4.5 at $1 in and $5 out, Claude Sonnet 5.5 at $2 in and $10 out, and Claude Opus 5.5 at $4 in and $20 out.

What that means for one receipt, using the same assumptions as our cost post: on Claude Haiku 4.5, one photo counts as at most 1,568 tokens. Add 500 tokens of instructions and 400 tokens of answer, and it's 2,068 × $1 ÷ 1,000,000 + 400 × $5 ÷ 1,000,000 ≈ $0.004: about 0.4 cents a receipt, or about $4 per thousand. If testing says you need Claude Sonnet 5.5, which looks at images in finer detail and charges twice as much per token, the worst case is about 1.6 cents a receipt. So reading a typical receipt or one-page document costs a fraction of a cent to a few cents.

Real numbers from my own project: WillIFit's full verification sweep of 1,050 parking garages across 87 cities cost $51.73 in AI and image fees, about 5 cents a garage. An earlier audit of 1,545 scans ($103) found 61% of the spend bought "no sign found." Splitting the job — cheap model detects, stronger model reads only when there's a sign — cut the cost from about 18 cents a garage to about 7 in a pilot. The savings came from how the software split the work between models — not from switching everything to the cheapest one, which in an earlier test misread a speed-limit sign as a clearance height.

If you hire us to build AI into your software, these are our standard rates, adjusted to each project and confirmed in a written quote before work starts:

• Platform fee: $650/month.
• License fee: $49/month per staff login. The monthly fees cover hosting, monitoring, backups, security updates, bug fixes, and email support.
• Changes and new features after launch: $150/hour, done only after you approve the work in writing (email is fine). Hours are tracked and reported.
• Larger builds or new modules may carry a one-time build fee, quoted as a fixed price or hourly before work begins.
• Third-party services — hosting, email, AI usage, storage — are included at ordinary usage levels. If usage materially exceeds that, as with large-scale AI processing, we tell you first, then either pass the extra through at cost or agree on new pricing in writing. SMS/text messaging is billed separately.

A five-person shop is $650 + 5 × $49 = $895/month. A two-person shop is $748/month. Billing is monthly in advance, and invoices are due Net-15. We own and maintain the software and license it to you while you're a client, which is why hosting, fixes, and security updates are included. Your data is yours: if you leave, it's made available for export in a common format, and you have 30 days after the agreement ends to take it. The full breakdown is in What AI Actually Costs a Small Business in 2026.

Is it safe to put customer data into AI?

It can be, if you set a few rules first. Vendor terms differ by plan and change often, so these are principles, not promises about any one product:

• Use a business plan for business data. A free personal account is not the place for customer records. Read the data terms of the plan you choose.
• Turn training off. Find out whether your conversations can be used to train the vendor's models, and turn that off if your plan gives you the choice.
• Never paste Social Security numbers, full card or bank account numbers, passwords, or anything a customer told you in confidence. And never paste patient health information into a general AI tool unless your practice has a Business Associate Agreement that covers that specific tool.
• Leave names out when you don't need them. "The customer" works as well as a real name for drafting a reply.
• Write it down. Put these rules on one page and give it to everyone who uses the tool. (That's week two of the plan below.)

If you're a medical practice, a law firm, or anyone else with professional confidentiality rules, those rules come first. The longer version is Is It Safe to Put Customer Data Into ChatGPT or Claude?

Getting found when customers ask AI

Some customers now ask an AI assistant instead of a search box: "who fixes transmissions near me," "is there a groomer open Sunday." Showing up there takes the same basics as showing up anywhere: a complete, accurate Google Business Profile; a website that says in plain text what you do, where, and for whom; the same name, address, and hours everywhere you're listed; and real reviews. Nobody can promise you'll be the answer, so be wary of anyone who does. The longer version is How Customers Find Local Businesses in ChatGPT and Google's AI Answers.

When custom AI software makes sense

Custom makes sense when the AI job is tied to your own records and happens often enough that copy-and-paste becomes its own chore. What we've actually built, with honest status labels:

• Mr. Camera HQ's estimate assistant (In daily use). At my own company, an AI assistant in the estimate builder turns plain requests like "add two camera ops for three days" into line items and reads uploaded vendor bids into the estimate. The software does the totals; a person sends the estimate. More in the Mr. Camera HQ case study.
• Receipt reading in Umbrella HQ (Live). For a nonprofit bookkeeping firm, it extracts grant terms from award letters, pulls line items from receipts, budgets, and funder reports, and flags anomalies for a person to check. See the Umbrella HQ case study.
• Certificate reading in Mr. Camera HQ (In daily use). Claude reads uploaded certificates of insurance and W-9s — expiry, coverage, carrier, tax details — and flags anything that needs attention for the office to check.
• A board financials app (Live). For a performing-arts nonprofit, the app turns the accountant's monthly figures into a board packet. It'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.
• Bug triage in the Revenant Portal (In daily use). When a client reports a bug from their site, Claude sorts it by severity and category and writes a summary on its own; that step is logged. An AI agent drafts the fix; nothing deploys until a person approves it. The portal meters and caps the AI it runs for each client.
• WillIFit's verification pipeline (Live). Google Street View images and Claude vision: a cheap model detects whether a clearance sign is visible, a stronger model reads it, exact-match guards reject misreads, readings that pass are published on their own, and conflicting readings wait for a person.

Everything else we've built, AI or not, is on our work page; the industry pages show how this looks in different lines of business, including nonprofits.

When it doesn't make sense

• The AI you already pay for does the job. Keep using it.
• The job happens a few times a month. Copy and paste is cheaper than software.
• Nobody can check the output. If no one can tell a right answer from a wrong one, you won't catch the mistakes.
• The process isn't settled. If everyone does it differently, write the process down first. Our automation guide has a rule for this: don't automate anything you haven't done by hand at least 20 times.
• The job is judgment or a relationship. Complaints, exceptions, condolences, and hard conversations belong to a person.
• The job is exact arithmetic. Sometimes the right answer is no AI. The payroll allocation tool we built for a grant-funded social-services nonprofit, now Live (beta), has none: it splits wages across funding sources and won't proceed until totals tie out. See the payroll allocation case study.

A 30-day starting plan

Week 1: Pick one job

Choose one job from the list above that happens often, bores everyone, and is easy to check. Time it for a week the way you do it now; you'll need that number in week four.

Week 2: Write a one-page AI policy

Which tool and plan you're using. What's allowed in it and what never goes in. Who reviews what before it goes to a customer. Who to ask when you're not sure. Keep it to one page so people read it, and give it to everyone who uses the tool.

Week 3: Try it on real work, with review

Use the AI on the real job, and have a person read every result before it goes anywhere. Tally each one — used as-is, edited, or thrown out — and write down every mistake.

Week 4: Measure the time saved yourself

Time the job again and compare it with week one, counting the time spent checking and fixing drafts. Don't take a vendor's word for the savings, or ours.

Day 30: Decide

Keep it, drop it, or add a second job. If the job is worth doing but copy-and-paste has become the bottleneck, that's when custom software is worth a conversation.

Frequently asked questions