Job to be done
Define the first visitor task in plain language: answer approved FAQs, guide service selection, collect an inquiry, or prepare a handoff. Separate required work from “nice to have” ideas.
Small-business buyer checklist
How to Choose an AI Chatbot for a Small Business
A practical checklist for evaluating the job, approved knowledge, boundaries, human handoff, data, accessibility, tests, ownership, maintenance, and the evidence behind vendor claims.
Write the job, limits, and ownership down first.
A small business should not choose a chatbot from a polished conversation alone. Start with the job visitors need done, the business material the assistant may use, the situations it must not handle, and the person responsible after launch. Then ask each vendor to demonstrate those requirements with evidence.
Buyer principle: select for fit and governability—not for the longest feature list or an unsupported promise of savings, accuracy, leads, or replacement of staff.
Ten-part small-business buyer checklist
Source material
Inventory the pages, policies, FAQs, service descriptions, hours, locations, and contact rules the assistant may use. Ask how sources are updated and how conflicting information is resolved.
Guardrails
List topics the assistant must decline, qualify, or escalate. Include unapproved prices, availability, guarantees, private records, safety advice, and regulated or high-impact decisions as relevant.
Human handoff
Name the triggers, actual recipient, coverage hours, transferred context, response expectation, and fallback. Ask the vendor to show the visitor and staff sides of the flow.
Data minimization
Collect only what the stated task and follow-up require. Identify where data goes, who can access it, how long it is retained, and how deletion or correction requests are handled.
Accessibility
Require keyboard operation, understandable labels and status updates, visible focus, readable contrast, mobile usability, and a non-chat contact option. Test with assistive technology relevant to your audience.
Testing
Use realistic questions, unsupported requests, adversarial wording, wrong information, mobile conditions, keyboard-only use, handoff failures, and correction scenarios. Record expected results before testing.
Ownership
Assign a business owner for source approvals, escalation rules, access decisions, conversation review, incident response, and vendor coordination.
Maintenance
Set a review schedule and event-based updates for changes to services, staff, pricing approach, locations, hours, policies, integrations, and risk.
Claim substantiation
For material vendor claims, request definitions, test conditions, dates, sample sizes, limitations, and customer references where appropriate. Put accepted commitments in writing.
1–2. Start with the job and approved knowledge.
The SBA notes that AI tools may help small businesses and specifically lists answering common questions as a potential chatbot use. It also recommends testing tools to see whether they add value and having another person review AI products for ethical, secure, and accurate business use.[4]
Turn that into a testable brief. “Help customers” is too broad. “Answer ten approved service questions and collect a callback request without confirming an appointment” is clearer. Attach the exact source material and identify the owner who can approve changes.
3–4. Make limits and handoff part of the purchase.
NIST describes the AI RMF as voluntary guidance intended to help organizations incorporate trustworthiness into the design, development, use, and evaluation of AI systems.[5] Its Generative AI Profile is a companion resource aimed at risks particular to generative AI.[6] A buyer can use that lifecycle view without turning procurement into a compliance claim.
Ask what happens when the assistant lacks a source, receives conflicting information, faces a prompt that tries to override its rules, or encounters a sensitive request. “It hands off” is incomplete unless the vendor can show the trigger, destination, context, timing, and fallback. Compare automated, staffed, and planned hybrid models in AI Chatbots vs. Live Chat.
5. Minimize data before discussing integrations.
The SBA cautions small businesses not to feed sensitive or proprietary information into AI tools.[4] For public-facing chat, map each requested field to a business purpose. If a city or ZIP code is enough to route an inquiry, do not collect a full address by default. Avoid requesting account, payment, medical, legal, identity, or other sensitive information unless the use is necessary, approved, secured, and supported by the full process.
- What information enters the model, transcript, email, CRM, analytics, and logs?
- Is customer content used to train any model, and under what terms?
- Who has access, and how is access removed?
- What are the retention, export, correction, and deletion processes?
- What happens when an integration or notification fails?
6. Test accessibility as behavior.
Do not accept “accessible” as a badge without a scoped test. Operate the launcher, panel, messages, controls, errors, and close action by keyboard. Check focus order and visibility, zoom, narrow screens, labels, dynamic announcements, contrast, motion preferences, and the ability to reach a conventional contact route. Ask which standards and assistive technologies were tested, by whom, when, and with what known limitations.
7. Run a realistic acceptance test.
Agree on pass and fail conditions before launch. A useful test set includes correct answers from approved sources, questions with no approved answer, attempts to obtain private information, prompt-injection language, corrections to captured contact details, repeat submissions, network interruption, and after-hours handoff. Test on the real page where the widget will run—not only in a vendor dashboard.
Record failures and retest them after changes. A successful happy-path demo does not establish accuracy, security, accessibility, or business value across real traffic.
8–9. Name the owner and maintenance cycle.
A chatbot is an operating responsibility, not a one-time content upload. Name one accountable business owner and backups. Set a recurring review cadence, but also update the assistant whenever a relevant business fact or risk changes. Keep an approval record for source changes, rules, access, integrations, and escalation routes.
Track measures tied to the original job, such as whether required intake fields are captured correctly or whether unsupported questions are escalated. Do not turn an easy-to-count metric into an unsupported claim of revenue or labor savings.
10. Require evidence for vendor claims.
CISA's Secure by Demand Guide gives software customers questions and resources for assessing a manufacturer's security approach and recommends bringing security into procurement discussions and, where appropriate, contract language.[7] Use that posture for security and other material chatbot claims.
Claim
“Accurate,” “secure,” “accessible,” “integrates with your CRM,” “24/7,” or “reduces workload.”
Evidence request
Definition, scope, architecture or test method, dated results, limitations, dependencies, and a demonstration in the intended environment.
Contract check
Ownership, support, security responsibilities, data terms, incident notice, export, termination, maintenance, and the exact included capabilities.
Do not assume a feature exists because it is common elsewhere. Distinguish what is available now, what requires configuration or integration, and what is only planned.
A compact evaluation scorecard
- Pass: the job is narrow and testable; sources and owners are named; guardrails and handoff work; data collection is justified; accessibility and failure cases are tested; material claims have evidence.
- Needs work: the demo is persuasive, but the source, retention, escalation, accessibility, or maintenance answer remains vague.
- Stop: the proposal depends on sensitive data without a clear need, promises unverified outcomes, hides limitations, or cannot identify who owns failures and updates.
The industries hub links this checklist and other practical planning guides.
Sources and further reading
- U.S. Small Business Administration: AI for small business — potential uses, testing, human review, sensitive-data cautions, and customer-trust considerations.
- NIST AI Risk Management Framework — voluntary guidance for managing AI risks and incorporating trustworthiness into design, use, and evaluation.
- NIST AI RMF: Generative AI Profile — a cross-sector companion resource addressing risks particular to generative AI.
- CISA Secure by Demand Guide — questions and procurement practices for assessing a software supplier's security approach.
Bring your sources, edge cases, and evidence requests.
A useful scoping conversation should make the assistant's job, boundaries, data, handoff, tests, and maintenance responsibilities concrete before launch.