Eric Kuvykin on AI Adoption by Small Businesses: Useful Automation Beats Hype

By Eric Kuvykin. Review and new notebook example: October 17, 2026. Original publication date retained.

An AI workflow can change without a new subscription or a new application. A staff member edits the instructions, adds a spreadsheet column or changes the examples, and the same task starts producing different results. My practical recommendation is to keep a small change ledger beside the workflow.

A change ledger for a payment-support summary

Use fictional tickets to test a tool that summarizes deposit questions for a human reviewer. Save the instruction version, date, model or application version when available, test-set identifier, reviewer and reason for the change. The output must identify missing evidence rather than invent a deposit date or promise a refund.

For an illustrative comparison, version A correctly identifies the missing transaction reference in 8 of 10 tickets. Version B improves that to 10 of 10 but invents a funding date in two cases. Calling version B better because its first score improved would miss the more consequential error. Keep separate columns for useful extraction, unsupported statements and review time.

Keep the difficult examples

Retain cases with conflicting dates, incomplete references and two payments for one invoice. Use invented or properly de-identified data; do not put live credentials or full card details into the notebook. When a previously failing case passes, keep it in the test set so the next change cannot silently reintroduce the error.

Before release, the reviewer signs off on the comparison and records how to return to the prior instructions. Restrict the pilot to drafting internal summaries. Customer promises, refund approvals and payment changes remain separate authorized actions. A useful change log records the decision and evidence, not merely that somebody clicked Save.

Use the same method across locations

A Florida seasonal business may use weekend deposit questions; a New York service firm may use invoice follow-up; a New Jersey distributor may use incomplete remittance messages. These are hypothetical applications. The acceptance standard stays the same: source evidence, visible uncertainty and a named reviewer.

NIST's voluntary AI Risk Management Framework core organizes risk work around Govern, Map, Measure and Manage. This notebook applies that discipline at a small scale without claiming certification. Continue through Eric's four-topic hub and the AMS regional reading paths for Florida, New York and New Jersey for related operating resources.