AI for small business: what it can actually do today, and what stays yours

What AI can actually do for a small business today

Strip the vendor lists down to what survives contact with a real business and five jobs remain. Follow-ups: the quotes, invoices, reminders, and check-ins that leak money when they go unsent. Marketing drafts: the posts, emails, and page copy that never get written because writing is nobody's job. Bookkeeping prep: the sorting, matching, and month-end tidying that stands between you and your accountant. Research: the competitor, supplier, and how-does-this-rule-work questions you never get to. Scheduling: the confirmations, reminders, and reschedules that repeat every week. Notice what these five share. Each one is made of words and records, repeats on a rhythm, and is small on its own, and the pile of them is why owners answer email at nine at night. That shape is exactly what AI handles well, and using it for this is now ordinary: the Census Bureau tracks AI use in its business surveys and the SBA publishes plain-language guidance on it. What AI still cannot do is also specific: carry responsibility, hold a relationship, or make a judgment call it has never seen before. The rest of this guide takes the five jobs one at a time, then draws the line around what stays yours.

• SBA AI for small business
• Census Bureau Business Trends and Outlook Survey

Customer follow-ups: the first job to hand off

Every small business leaks the same way: the quote sent Tuesday and never chased, the regular who went quiet in March, the invoice sitting at forty days, the happy customer nobody asked for a review, the appointment that a reminder would have saved. None of these are hard. Each one is a two-minute message that arrives at the wrong time, in the middle of the actual work, which is why they do not get sent. This is the single best first handoff to AI, because the job is pure follow-through: draft the chase in your voice, remember the timing, send the polite third nudge without feeling awkward about it, and surface the reply when one comes back. You still set the tone and the cadence, and anything involving a sensitive history gets read by you before it goes. The reason to start here rather than anywhere clever is that the result is visible fast: within a week or two, replies start arriving to messages that would never have existed, and unlike almost everything else in the AI conversation, nobody has to take the value on faith.

Marketing drafts: AI writes the first pass, you stay the voice

The marketing most small businesses need is not a campaign. It is volume they never get to: the social post about this week's work, the monthly email to the customer list, the seasonal offer, the page that still describes what the business sold two years ago. AI collapses the cost of that volume, a month of posts and a decent email draft now take an evening of editing instead of a winter of guilt. Two honest warnings keep this useful. First, generic AI writing is now the baseline everywhere, and readers can smell it, so the value is in your specifics: the real job you finished, the street name, the thing a customer said on Thursday. AI drafts the frame; you supply the proof. Second, every message that goes out in your business's name is your claim. The FTC's truth-in-advertising rules apply to a two-person shop the same as a national brand, and a model will cheerfully overpromise if you let it. So the working rule is simple: AI drafts, you edit for truth and voice, and nothing ships with a promise you could not repeat to the customer's face.

• FTC advertising and marketing guidance

Bookkeeping prep, not bookkeeping judgment

The books are where AI's limits are clearest, and the split is worth stating plainly: AI is good at bookkeeping preparation and bad at bookkeeping judgment. Preparation is the pile: expenses sorted into categories for review, invoices matched against orders, a running list of who owes what and for how long, a drafted note to the vendor whose bill does not match the purchase order, and a month-end folder your accountant can open without sighing. That work is pattern matching over records, which is exactly what the models do well, and it is the difference between tax season as a weekend and tax season as archaeology. Judgment is everything else: what a payment actually was, how an edge case should be treated, anything you would sign. That stays with you and your accountant, who gets more useful, not less, when the prep arrives clean. One grounding fact keeps this section honest: the IRS expects records behind your numbers regardless of who or what prepared them, so the win to aim for is not fewer records but organized ones, kept current weekly instead of reconstructed in April.

• IRS recordkeeping for small businesses

Research: a fast first pass on the questions you never get to

Owners carry a list of questions that never make it to the top of a working day. What did the competitor across town change their prices to? Is the supplier's new contract normal or worse than what everyone else offers? What does that new local rule actually require? What have customers been saying in reviews and messages, read as three months at once instead of one sting at a time? AI is a genuinely strong first-pass researcher: it reads a lot quickly, returns a draft answer with its reasoning visible, and turns a vague worry into a one-page brief you can act on. The classic recurring version is a weekly competitor watch, one paragraph covering anything that moved, which is the kind of thing no owner sustains by hand past week three. The known failure mode is confidence at the edges: AI is most persuasive exactly where it is guessing. So treat every answer as a briefing, not a verdict, and check anything touching money, law, or safety against the primary source before acting. A briefing you got is still worth more than the verdict you never had time to look up.

Scheduling, reminders, and the weekly admin loop

The least glamorous family is the most reliable: confirmations sent when someone books, reminders the day before, a drafted reschedule offer when someone cancels, the unpaid-invoice nudge on day thirty, and a Monday morning summary of the week ahead, what is booked, what is owed, what went quiet. The rule of thumb underneath it is frequency: the more often a task repeats without changing shape, the better AI is at it and the worse a use of an owner's attention it is. A reminder that goes out every single time beats a clever one that goes out usually, so the goal here is boring reliability, not intelligence. This is also where the pieces compound: the reminder cuts the no-show, the summary catches the invoice, the invoice nudge feeds the follow-up, and the loop runs whether or not you had a hard week. You still make the structural calls once, in settings rather than per appointment: what hours are bookable, who can claim them, and what never gets scheduled without a human yes.

What stays the owner's judgment

Here is the section the tool lists skip. Some work should not be delegated to AI at any price, and knowing the list is what makes the rest of the handoff safe. Pricing, because it encodes what you know about your costs, your market, and your nerve. Hiring and firing. The quality bar, meaning what is allowed to leave the shop with your name on it. Any customer conversation with heat or history in it, because an upset regular can tell the difference between you and a system, and the difference is the relationship. Anything legal or financial you would sign. And the meta-decision itself: which work should never be automated. The reason is structural, not temporary. AI predicts likely words; it does not carry responsibility, and your customers trust a person, not a stack. Two practical habits follow. Review everything customer-facing until a track record is earned, then loosen deliberately, job by job. And be careful what you feed free tools: customer records and anything sensitive stay out of products with unclear data terms, a risk the SBA's small business cybersecurity guidance covers in plain language. A good test for the whole section: hand off what you would give a competent new hire in their first month, and keep what you would not give them in their first year.

• SBA strengthen your cybersecurity

The honest math, and how Foundable runs it

The tool lists price each product and never total the stack. Add up the usual recommendations, one subscription each for writing, phones, meetings, books, and marketing, and a small business is looking at a few hundred dollars a month, plus the setup hours each product demands, plus the quiet cost that none of them talk to each other, which makes you the integration. So price any approach honestly: setup time up front, review time until trust is earned, and a monthly bill, against a win counted first in hours, because hours are what an owner is actually out of. When it works, the shape is unmistakable: follow-ups stop leaking, the books stay current, the admin loop runs on schedule, and the evening email session shrinks. Foundable's version is one operator instead of a stack. You tell Ted, the AI operator, how the business runs, connect the services you already use through 150+ verified integrations, and hand off workflows one at a time: customer follow-ups, the Monday update, inbox triage, the vendor contract scan. The work stays visible, you can pause any of it, and anything that reaches a customer waits for your yes until you decide otherwise. Foundable has an ongoing Free plan with no credit card required. The optional paid plan starts at $49 monthly, or $37 monthly equivalent billed $441 annually, with 2,000 credits each month, and owners can choose a higher recurring allowance. The Foundable for running businesses page covers the product side in detail.