AI for Small Business study guide · Lesson 7 of 8
Privacy, Security, and Responsible Use
Protect Customer and Business Data · Build Human Oversight and Quality Controls
This is the lesson's study guide from the FlashFeed app, plus its key facts and a self-test.
What you'll be able to do
- Classify business information, minimize shared data, and check vendor retention and training controls before use.
- Match oversight to consequence using risk tiers and human review.
Data
Before sharing information with an AI service, name the purpose and classify the data. Public marketing copy is different from a customer list, payroll file, health record, trade secret, or contract under confidentiality.
Practice data minimization. Share only fields needed for the task. Remove direct identifiers and unnecessary narrative. De-identification is not magic; combinations such as location, dates, and transaction details can still identify someone.
Evaluate the exact product and account type. Training, retention, sharing, administrator access, integrations, and deletion controls can differ between personal and business offerings. For example, as of 2026, OpenAI states that its Business, Enterprise, and API offerings do not use customer inputs and outputs for model training by default. That fact applies to those offerings, not every AI service, and vendor policies can change, so reverify current terms before relying on this.
The FTC advises businesses to collect only what they need, protect it, honor privacy promises, and dispose of it securely. If your privacy policy says one thing and your AI workflow does another, the new technology does not excuse the mismatch.
Governance
Use consequences to set controls. Internal brainstorming may need ordinary review. Customer messages need factual and brand review. Payments, employment decisions, legal commitments, safety advice, and sensitive-data workflows need stronger testing, approval, and documentation.
NIST's voluntary AI Risk Management Framework uses four functions: Govern, Map, Measure, and Manage. For a small business, that can be simple. Assign an owner and policy; understand the use case and affected people; test performance and risks; then choose controls and monitor results.
A human in the loop must be placed where review can still change the outcome. Add approval gates before consequential actions, escalation rules for uncertainty, least-privilege access, and audit logs.
Plan for incidents. Know how to stop the workflow, preserve records, correct bad messages or data, notify the owner, and prevent recurrence. Reassess when the vendor, model, source data, or business process changes.
Create a simple data map
List the information used by the proposed workflow, where it comes from, where it is sent, who can access it, how long it is kept, and how it is deleted. Include connected apps: an AI service may be acceptable while a plug-in or automation sends data somewhere unexpected.
Classify data using a small set of labels such as public, internal, confidential, and restricted. Define examples for your business. Customer contact details, employee records, payment information, health information, credentials, and trade secrets deserve explicit treatment.
Ask vendors concrete questions. Is content used for training by default? How long is it retained? Can administrators access it? Which subprocessors receive it? What happens when an account is closed? Can connected tools take actions? Product names and plan names change, so record the date and link for each answer.
Scale controls with risk
Create three practical tiers. Low-risk work is internal, reversible, and uses nonsensitive data. Medium-risk work reaches customers or influences operations but remains reviewable. High-risk work affects money, rights, safety, legal obligations, employment, or sensitive data.
For each tier, define required review, testing, approval, logging, and incident response. Prohibit uses whose risk you cannot manage. A short policy that people understand is stronger than a long document no one follows.
Train staff to report surprising output, accidental data sharing, incorrect customer messages, and unauthorized actions without hiding mistakes. Fast reporting limits harm and improves the system.
Key terms
- Data retention
- How long a service or business keeps information before deletion or archival.
- Human in the loop
- A workflow design in which a person reviews, approves, corrects, or can stop AI-supported work at a meaningful point.
Key facts

Share the Minimum
Do not paste an entire customer file when the task only needs a product category and date. Data minimization reduces exposure and often improves focus.
Plan Labels Are Not Data Controls
Free, personal, business, and enterprise products can have different training, retention, sharing, and administration settings. Verify the current terms and controls for the exact product you use.
Business Products May Offer Stronger Defaults
For example, OpenAI states (as of 2026) that ChatGPT Business, Enterprise, and API data are not used to train models by default. Every vendor's terms should be verified separately and periodically before sensitive use.

Control by Consequence
The more an AI output can affect money, rights, safety, reputation, or sensitive data, the stronger the testing, approval, documentation, and monitoring should be.
Govern, Map, Measure, Manage
NIST organizes AI risk work into four functions: establish governance, understand the context, assess risks and performance, and prioritize responses.
Plan for Failure
An AI incident plan should say how to stop the workflow, preserve logs, correct affected records or messages, notify responsible people, and prevent recurrence.
Matched pairs
Each item on the left goes with the one on the right.
- Public data
- Information intentionally available to anyone
- Confidential data
- Restricted business, customer, employee, or partner information
- Reversible action
- An action that can be corrected with limited harm or cost
- Consequential action
- An action that can materially affect money, rights, safety, or trust
Step by step
Check data before using an AI tool
- Name the business purpose
- Classify the information involved
- Remove fields that are not necessary
- Verify the tool's current training, retention, access, and sharing controls
- Confirm customer, employee, contractual, and legal permissions
- Choose a safer method or stop if controls are insufficient
When in doubt, use synthetic or de-identified examples for experimentation.
Run an AI workflow preflight
- Define the intended use and owner
- Identify affected people, data, and decisions
- Rate the consequence of errors or misuse
- Set tests, approval gates, permissions, and logs
- Pilot with limited scope
- Monitor outcomes and document incidents and changes
Reassess the workflow when the model, vendor, data, or business process changes.
Self-test
Answer each question from memory, then check your work against the answer key.
1. You want AI to segment a customer list. What should happen first?
- AAsk the AI provider whether breaches are impossible
- BUpload the full list to a free tool
- CReplace customer names, but keep every other identifier
- DVerify authorization, tool controls, needed fields, and retention before sharing data
2. What does practicing data minimization mean when using an AI service?
- AKeep every record indefinitely in case it is useful later
- BShare all available data so the AI has full context
- CShare only the fields actually needed for the task
- DEncrypt data before pasting it into any tool
3. Your team upgrades from a free to a business AI plan. What should you assume about its training and retention settings?
- AThey are identical across all plan tiers
- BThey may differ, so verify the current terms for that exact product
- CBusiness plans never retain any data, regardless of settings
- DFree plans always offer stronger privacy
4. Using the labels public, internal, confidential, and restricted, which item is public data?
- AA marketing brochure published on your website
- BAn employee payroll file
- CA signed contract under confidentiality
- DA customer contact list
5. In vendor and data terms, what does 'data retention' refer to?
- AThe number of people who can access a dataset
- BThe process of encrypting stored records
- CWhether inputs are used to train a model
- DHow long a service or business keeps information before deletion or archival
6. What does data minimization mean when sharing information with an AI tool?
- AConvert all data into images before sharing
- BShare only the fields needed for the task
- CDelete the AI vendor's account after use
- DEncrypt every file before upload
7. You only need a product category and date for an AI task. What should you do before sharing customer data?
- AShare names but keep every other identifier
- BAsk the vendor to retain the data forever
- CRemove fields the task does not need
- DUpload the entire customer file for context
8. Before using an AI vendor for sensitive work, what should you check?
- ADefault training use and data retention period
- BHow many five-star reviews the tool has
- CWhether the tool offers a free trial period
- DWhether the vendor offers phone support
9. Which OpenAI product tiers are stated to skip customer inputs and outputs for model training by default?
- AFree, Plus, and Pro personal accounts
- BBusiness, Enterprise, and API
- CPlus, Pro, and Enterprise
- DEvery OpenAI plan, including Free
10. Which workflow needs the strongest review and controls?
- AGenerating internal brainstorming ideas
- BReformatting a meeting agenda
- CDrafting three subject-line options
- DAutomatically changing employee pay based on an AI score
11. In AI governance, which best describes a consequential action?
- AAny step a person happens to review before it runs
- BAn action that can be corrected later, with limited harm or cost
- CAn action that can materially affect money, rights, safety, or trust
- DA task that stays entirely inside internal brainstorming
12. Where should a human in the loop be placed to be effective?
- AAt a point where review can still change the outcome
- BAfter the action is complete and irreversible
- COnly during the initial software setup
- DWherever it slows the workflow the least
13. Which action is NOT part of a sound AI incident plan?
- ADeleting all logs to reduce liability
- BStopping the workflow
- CNotifying the responsible people
- DCorrecting affected records or messages
14. A cafe uses AI to brainstorm in-house promotion ideas that never leave the shop and are easy to redo. Which risk tier fits best?
- AHigh risk
- BMedium risk
- CProhibited use
- DLow risk
Go deeper
Answer key
- 1. D Verify authorization, tool controls, needed fields, and retention before sharing dataSensitive-data work starts with purpose, authorization, minimization, vendor controls, and retention.
- 2. C Share only the fields actually needed for the taskData minimization means sharing only the fields needed for the task, which reduces exposure and often improves focus. Sharing everything raises risk.
- 3. B They may differ, so verify the current terms for that exact productTraining, retention, and sharing can differ by product and account type, so you must verify the current controls for the exact product. Plan labels alone are not data controls.
- 4. A A marketing brochure published on your websitePublic data is information intentionally available to anyone, like published marketing copy. Customer lists, payroll files, and confidential contracts are restricted business information.
- 5. D How long a service or business keeps information before deletion or archivalData retention is how long information is kept before deletion or archival. Access counts and training use are separate vendor questions.
- 6. B Share only the fields needed for the taskData minimization means sharing only what a task requires, which reduces exposure of unnecessary information.
- 7. C Remove fields the task does not needMinimization means removing fields beyond what the task requires, not just swapping out direct identifiers.
- 8. A Default training use and data retention periodVendor evaluation should cover default training use and data retention period before sensitive data is shared.
- 9. B Business, Enterprise, and APIAs of 2026, OpenAI's Business, Enterprise, and API offerings do not use inputs and outputs for training by default — verify the current policy, since vendor defaults can change.
- 10. D Automatically changing employee pay based on an AI scorePay decisions affect rights and livelihoods, so they require strong controls and accountable human judgment.
- 11. C An action that can materially affect money, rights, safety, or trustA consequential action can materially affect money, rights, safety, or trust, so it needs stronger controls. An action correctable with limited harm is a reversible one, not a consequential one.
- 12. A At a point where review can still change the outcomeA human in the loop must sit where review can still change the outcome, with approval gates before consequential actions. Review after an irreversible action cannot prevent harm.
- 13. A Deleting all logs to reduce liabilityAn incident plan says to preserve records, not delete them. Stopping the workflow, correcting affected data, and notifying owners are all core steps.
- 14. D Low riskInternal, low-stakes, easily undone work is low risk. Medium reaches customers, high affects money or rights, and prohibited uses are those whose risk cannot be managed.
Sources
- Federal Trade Commission — Privacy and security guidance — www.ftc.gov/business-guidance/privacy-security
- NIST — AI Risk Management Framework — airc.nist.gov/airmf-resources/airmf/
- OpenAI — Business data privacy, security, and compliance — openai.com/business-data/
- OpenAI Help Center — Data controls in ChatGPT — help.openai.com/en/articles/7730893-data-controls-faq
- OpenAI — Enterprise privacy — openai.com/enterprise-privacy/
Image credits
- Share the Minimum: AI-generated (gpt-image-1.5)
- Control by Consequence: AI-generated (gpt-image-1.5)