AI for small business: uses that pay back, and one worked on a sales file

AI for small business pays back fastest in text work: drafting quotes and emails, answering routine customer questions, pulling data off invoices for the books, and asking questions of your own sales file. Pick one task, time it before and after, and check anything with a figure in it before it reaches a customer or a tax return.

Where AI pays back in a small business

A small company has no analyst and no spare week for a project. The uses that work are the ones that save someone an hour this week on a job they already do, and whose output the same person can check in a minute. The SBA's guide to AI for small business puts the case as AI helping small businesses "do more with less," with the advice to have a person review what it produces.

One task at a time. Choose one job, time it for two weeks without AI and two weeks with it, and keep the tool only if the saving is real after checking time.

Eight uses, with what to measure

The last one is different from the other seven: its output is a number, not text. A wrong word in a draft email is visible; a wrong total is not. It is the one use worked through below.

Worked example: which regular customers have gone quiet?

A 20-person food wholesaler, as of September 30, 2026. The question: "Which regular customers have gone quiet?" The rule is that each customer's own rhythm sets the alarm:

Days since last order = As-of date − Last order date

Typical gap = 365 / Orders in the last 12 months, rounded down to whole days

Flag if Days since last order > 2 × Typical gap

The tool computes the dates and gaps; the external AI model explains the list.

In Excel, with the as-of date in J1, orders in B, last order in C:

D2: =INT(365/B2)
E2: =$J$1-C2
F2: =2*D2
G2: =IF(E2>F2,"Yes","No")
Flagged revenue: =SUMIF(G2:G7,"Yes",H2:H7)

The result: three customers flagged, Corner Deli Co, Greenway Grocers and Juniper Kitchens, with $49,300 of twelve-month revenue between them, 39.7% of the $124,200 total. Greenway orders about every 60 days, so a 60-day silence would be normal; at 133 days it is not. A fixed cut-off of "no order in 60 days" gives the same three today, but it would have flagged Greenway on July 20, at 61 days, when it was ordering on its usual rhythm, and it would leave Fenwick, which orders every 15 days, unflagged for two months of silence. The same method in Excel, step by step, is in how to find dormant customers in Excel, and what to do with the list is in the reactivation list.

The check

Flagged revenue + Unflagged revenue = Twelve-month total

18,400 + 9,600 + 21,300 = 49,300 flagged; 24,000 + 38,200 + 12,700 = 74,900 not flagged; 49,300 + 74,900 = $124,200, the twelve-month total from the sales file. Then recompute each flag from its dates: Corner Deli, July 14 to September 30, is 78 days, more than 2 × 33 = 66. If a general chatbot gives you a different list from the same file, the dates were not computed; can ChatGPT analyze sales data? shows how that happens.

What it costs and how to judge payback

Judge it on your own hours, not the vendor's claims:

Net monthly saving = Hours saved per month × Hourly cost − Checking time cost − Monthly subscription

Payback (months) = Setup cost / Net monthly saving

A worked case with round illustrative figures: quote drafting saves 12 hours a month at a loaded cost of $40 an hour, $480. Checking the drafts takes 2 hours, $80. A tool costing $100 a month leaves a net saving of 480 − 80 − 100 = $300 a month. Eight hours of setup at $40 is $320, so payback is 320 / 300 = 1.1 months. If the net saving comes out at zero or below, stop; a spreadsheet may be all you need, as Excel or an analytics tool sets out.

Risks for a small company

• Customer data. Read the terms for whether your inputs are used to train the provider's systems, how long they are kept and who can see them. The SBA's advice is to avoid feeding the tool "any sensitive data or proprietary information."
• Wrong figures. Totals and averages typed by a chatbot go into a quote or a sales tax return unchecked.
• Over-reliance. Staff stop checking output that has been right ten times in a row.
• Rights and accuracy in generated content. The SBA warns to make sure anything you produce does not infringe patents, copyrights or trademarks, and the FTC's artificial intelligence guidance shows it acting against AI-generated content, such as fake reviews, that misleads customers.

For a structure to think these through, the NIST AI Risk Management Framework is voluntary and written for organizations of any size.

Where it goes wrong

• Customer lists in a free chatbot whose terms allow training on what you paste.
• Unchecked totals from a chatbot carried into a quote or a sales tax return.
• Five tools at once, so no single one is measured.
• Automated replies to complaints, where a person is needed.
• Generated images or copy used without checking rights and accuracy.

Your own sales file, answered

With Covirage you upload the sales file and ask; its tools compute each customer's gap and flag, cite the rows, and check the flagged and unflagged revenue add to the total, while the external AI model explains the list. See AI analytics to ask your own sales file which customers have gone quiet. For the wider picture, see artificial intelligence in business, AI data analyst and AI data analysis.