What should a local business automate?

1. Begin with the business problem, affected people, intended outcome, and accountable owner

Start with a repeated problem, not with an AI or automation product. Describe what happens now, who performs the work, who could be affected by an error, what a useful result would look like, and which real person owns the process. A checklist, clearer responsibility, or an existing software feature may be the better answer when the work is rare, unstable, or low value.

Good first candidates are often assistive and reversible: preparing an internal summary, formatting an approved record, drafting material for review, routing a task, flagging an exception, or preparing an approved follow-up without sending it. A carefully controlled operational workflow might later create a task, update a low-risk status, or move approved information between systems, but only after its authority, exceptions, recovery path, and production gate are documented.

NorthWorx uses the following R0–R3 labels as an editorial and scoping framework. They are not a NIST classification, government standard, legal determination, security certification, or universal risk rating. Classify the workflow at the highest level that applies; do not average a serious risk down because other steps are simple.

• R0 — Prohibited: deceptive, unauthorized, credential-seeking, consent-evading, or uncontrollable behavior that must be declined or redesigned
• R1 — Assistive: reversible internal support that prepares information or suggestions for a person
• R2 — Controlled operational: a recoverable action with documented authority, safeguards, exceptions, and separate production approval
• R3 — Consequential or restricted: commitments, external publication, security or financial changes, material deletion, or high-impact decisions that default to propose-only or no automation pending qualified review

Sources for this section:National Institute of Standards and Technology: Artificial Intelligence Risk Management Framework (AI RMF 1.0)National Institute of Standards and Technology: NIST AI RMF PlaybookNational Institute of Standards and Technology: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)

2. Classify the information, access, consent, and authority involved

List every system, field, message, file, account, provider, and person the workflow would touch. Classify the information as public, confidential business information, personal information, sensitive or regulated information, a reusable credential, or information the project is not authorized to use. Record why each item is needed, who controls it, who may access it, where it may travel, how long it should remain, and how access can be removed.

Privacy and security are related but different. Security controls can reduce unauthorized access, but they do not by themselves make collection, use, sharing, retention, or automated decision-making appropriate. Use the least information and least access needed for the approved purpose, and obtain case-specific privacy, legal, contractual, or industry review when the context requires it.

Do not collect passwords, 2FA codes, recovery codes, API secrets, private keys, payment-card information, or other reusable credentials through ordinary intake or workflow fields. Prefer official invitations, delegated roles, OAuth connections, individual accounts, least privilege, and separately reviewed secret handling.

Permission to use information does not automatically authorize a communication. Email, SMS, calls, reminders, confirmations, review requests, and other outward messages remain behind consent, sender-identity, privacy, suppression, quiet-hour, provider, A2P, template, configuration-review, testing, and separate activation gates.

• The purpose and affected people are documented
• Information and retention are minimized
• Every account uses approved individual or delegated access
• Consent and channel authority are verified separately
• Access can be reviewed, removed, and recovered without sharing reusable credentials

Sources for this section:National Institute of Standards and Technology: NIST Privacy FrameworkNational Institute of Standards and Technology: NIST SP 800-63 Digital Identity GuidelinesCybersecurity and Infrastructure Security Agency: Require Multifactor AuthenticationFederal Communications Commission: FCC 24-24: Rules governing revocation of consent for robocalls and robotexts

3. Define rules, human decisions, exceptions, and uncertainty stops

Write the normal rule in plain language: the approved inputs, trusted sources, allowed actions, prohibited actions, output, owner, and evidence required. Then list the exceptions. Missing, ambiguous, stale, conflicting, sensitive, or unauthorized information must stop or route to a named person; the workflow must not guess, silently continue, or invent a fact, identity, price, promise, consent state, or completed action.

A human approval step is meaningful only when a real person has enough context, time, authority, and practical control to approve, reject, correct, pause, or shut down the action. A ceremonial click placed after the decision has effectively been made is not responsible human control.

Drafting is not sending. Preparing a reply, quote, reminder, confirmation, review request, or follow-up can be R1 assistance, while sending it changes the risk and requires the applicable R2 or R3 controls and separate activation. Prices, guarantees, refunds, contracts, offers, public claims, account or security changes, payments, and other business commitments require authorized human review and retained support.

Employment, credit, legal, medical, safety, insurance, tax, eligibility, and similarly consequential decisions should be propose-only or not automated by default. Qualified domain review, specific authority, explainable source evidence, appropriate accommodation or appeal paths, and a genuine human decision are required before considering any different treatment.

• Allowed and prohibited actions are specific
• Uncertain or unauthorized inputs stop or escalate
• The approver can practically reject, pause, and correct
• External communications and commitments remain separately gated
• Consequential decisions default to propose-only or no automation

Sources for this section:National Institute of Standards and Technology: Artificial Intelligence Risk Management Framework (AI RMF 1.0)Federal Trade Commission: Advertising FAQs: A Guide for Small BusinessFTC, DOJ, CFPB, and EEOC: Joint Statement on Enforcement Efforts Against Discrimination and Bias in Automated SystemsU.S. Equal Employment Opportunity Commission: Artificial Intelligence and the ADAConsumer Financial Protection Bureau: CFPB Circular 2022-03: Adverse action notification requirements for complex algorithmsFederal Communications Commission: FCC 24-24: Rules governing revocation of consent for robocalls and robotexts

4. Test normal operation and realistic failure paths

A demonstration that works once is not a readiness decision. Use privacy-safe synthetic data to test an ordinary case and realistic failures before considering real information or production access. Check missing and stale data, conflicting instructions, sensitive content, duplicates, partial writes, timeouts, rate limits, unavailable providers, unavailable approvers, changed APIs or models, expired access, and retries after an uncertain result.

The workflow should fail visibly and recoverably. It should not send twice, overwrite a good record, claim success from a provider acknowledgment alone, retry without a limit, or switch to a higher-impact action when the intended path fails. Define duplicate protection, validation, a manual fallback, recovery steps, and the evidence needed to confirm the actual business result.

AI-assisted output needs the same failure discipline plus testing for unsupported or inconsistent output. Provider, model, data, and business-rule changes can alter behavior after a prototype is accepted. Passing a prototype test confirms only that the tested version met the documented test at that time; it does not establish accuracy, safety, compliance, security, availability, or authorization for production use.

• Normal, duplicate, partial, delayed, unavailable, and unauthorized paths are tested
• Retries are bounded and uncertain results do not become duplicate actions
• A provider response is verified against the intended business result
• Manual fallback and recovery steps are usable
• Prototype acceptance and production authorization remain separate decisions

Sources for this section:National Institute of Standards and Technology: Artificial Intelligence Risk Management Framework (AI RMF 1.0)National Institute of Standards and Technology: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)National Institute of Standards and Technology: The NIST Cybersecurity Framework (CSF) 2.0

5. Retain evidence and establish pause, shutdown, recovery, maintenance, and retirement responsibilities

Keep enough evidence to reconstruct what happened: the workflow and rule version, relevant input source, action attempted, provider result, exception, human approval, and final verified state. Protect the logs, limit access and retention, and do not copy secrets or unnecessary personal information into them. Logging supports investigation; it does not guarantee constant monitoring or detection.

Assign a named escalation owner, manual fallback, pause or disable control, shutdown conditions, recovery steps, review cadence, change-control process, access review, regression tests, incident record, and retirement plan. The responsible person must know how to use those controls before production activation. A switch that nobody is authorized or prepared to operate is not a meaningful safeguard.

Measure the operating process before claiming a business result. Useful checks may include whether approved inputs were used, exceptions reached the right person, duplicate protection worked, communications remained gated, and the final state was verified. Do not promise saved time, fewer errors, lower costs, increased bookings, revenue, accuracy, availability, security, or another outcome without retained evidence, a defined method, relevant dates, and appropriate limitations.

This framework is practical planning guidance for a local business. It is not legal, privacy, cybersecurity, employment, credit, medical, tax, insurance, or regulatory advice; it does not certify a workflow or make every recommendation appropriate for every business. Production still requires the specific provider, data, security, consent, A2P, configuration, test, and separate activation approvals that apply to that use case.

• Evidence can reconstruct the action without retaining unnecessary secrets or personal data
• Escalation, manual fallback, pause, shutdown, and recovery have named owners
• Provider and business changes trigger review and regression testing
• Retirement removes or transfers access and preserves required records
• No operational or business outcome is guaranteed without supporting evidence

Sources for this section:Cybersecurity and Infrastructure Security Agency: Use Logging on Business SystemsNational Institute of Standards and Technology: Artificial Intelligence Risk Management Framework (AI RMF 1.0)National Institute of Standards and Technology: The NIST Cybersecurity Framework (CSF) 2.0Federal Trade Commission: Advertising FAQs: A Guide for Small Business