A practical operator ebook for deciding what AI may draft, what humans must approve, and what evidence to check before small-business work goes live.
This ebook helps tiny operators decide what AI may draft, organise, or automate — and what must stay under human approval. Use it before delegating customer, money, legal, public, or operational work to an AI tool.
Legal commitments, payment changes, final refunds, HR decisions, public incident statements.
AI prepares words or options. A human checks every fact and sends manually.
AI classifies or summarises. Human samples outputs and keeps rollback.
Narrow permissions, logs, tested rules, and stop buttons are required.
AI may draft calm refund replies from your policy; a human approves tone, amount, and exception handling before sending.
Policy URL, order ID, refund window, prior messages.
Would I stand behind this if the customer screenshots it?
AI may produce polite follow-up options; owner selects timing and any discount language.
Original quote, expiry date, customer constraint, margin floor.
Does this help the buyer decide without pressure?
AI may rewrite benefits and FAQs; human checks accuracy, promises, and support burden.
Current listing, actual files included, limitations, target customer.
Could a buyer reasonably expect more than we deliver?
AI may create short captions; human approves claims, hashtags, and cultural references.
Product link, launch note, audience, banned claims.
Is this invitation honest rather than manipulative?
AI may summarise public positioning; human avoids copying names, text, layout, or offers.
Public URLs, date accessed, comparison notes.
Are we learning the market, not cloning it?
AI may classify issues and suggest first reply; human handles anger, refunds, legal threats, and repeated failures.
Ticket text, purchase status, known bug list.
Is this a file-access problem, expectation gap, or relationship issue?
AI may draft reminders; human checks payment status and relationship context.
Invoice number, due date, amount, prior promise.
Would this still feel respectful if payment already arrived?
AI may outline educational posts; human supplies real examples and removes unsupported claims.
Topic, reader level, business stance, source links.
Does this teach one useful decision?
AI may convert observed steps into a checklist; operator tests it once before adopting.
Screen notes, file paths, owner approvals, rollback path.
Can a tired future operator follow this safely?
AI may summarise notes; human confirms decisions, owners, dates, and sensitive omissions.
Transcript or notes, attendee list, action items.
What would be dangerous if summarised wrong?
AI may propose packaging options; human sets price from cost, value, and positioning.
Time cost, support cost, market range, margin floor.
Does this price create a sustainable promise?
AI may flag unclear sections; human decides final copy and legal claims.
Page URL, offer, audience, constraints.
Where could a reasonable visitor misunderstand us?
AI may draft role criteria; owner removes bias and confirms legal compliance locally.
Role need, must-have skills, values, work conditions.
Are we judging the work, not stereotypes?
AI may build comparison tables; human verifies current pricing and terms.
Vendor pages, quote PDFs, usage needs, exit costs.
What happens if this vendor disappears?
AI may produce plain-language policy options; qualified human/legal review required for binding terms.
Jurisdiction, business model, risk areas.
Is this guidance, or are we pretending it is legal advice?
AI may suggest triggers; human approves every write action and failure alert.
System map, API permissions, rollback process.
What is the worst thing this automation can do unattended?
AI may draft structure; human checks segment, consent, links, and unsubscribe compliance.
Audience segment, send reason, links, product truth.
Did people ask to hear from us about this?
AI may synthesize patterns; human grounds personas in real observations, not stereotypes.
Sales notes, support themes, search terms.
What evidence would change this persona?
AI may draft FAQs; human answers from actual product limits.
Support tickets, product files, refund rules.
Does each answer reduce support honestly?
AI may list failure modes; owner ranks severity and chooses mitigations.
Launch plan, dependencies, public promises.
Which risk needs a stop sign, not a note?
AI may generate variants; human approves platform rules, claims, and budget before any ad spend.
Offer, audience, banned words, budget cap.
Could this create demand we cannot serve?
AI may suggest categories; human approves deletion, merge, and retention rules.
Export sample, field meanings, backup path.
Can we reverse this if the categorisation is wrong?
AI may create first-pass translation; fluent human checks nuance before publishing.
Source text, locale, tone, terms to preserve.
Would a native speaker trust this?
AI may summarise trends; human checks definitions and sample size.
Dashboard export, date range, campaign notes.
Is this signal big enough to act on?
AI may suggest ethical add-ons; human rejects pressure, scarcity tricks, and irrelevant bundles.
Purchase context, complementary products, support load.
Does this genuinely help the buyer next?
AI may turn messy reports into clear tickets; developer/owner confirms reproduction steps.
User report, browser/device, logs, screenshots.
Can someone reproduce this without guessing?
AI may structure an application; human supplies truthful evidence and verifies rules.
Criteria, deadlines, metrics, proof.
Are we overstating impact?
AI may create practice scripts; manager checks realism and psychological safety.
Team role, common cases, escalation policy.
Does this train judgement, not robotic lines?
AI may summarise stock patterns; human checks counts before buying.
Stock export, sales history, supplier lead times.
What cash gets trapped if this is wrong?
AI may draft status language; accountable human approves facts, apology, and next update time.
Incident timeline, known impact, fix status.
Are we being clear without speculating?
AI may summarise clauses for easier review; qualified human/legal review decides meaning.
Contract, questions, risk tolerance.
Which clause can cost us money or control?
AI may draft collaboration notes; human approves rights, payment, usage, and boundaries.
Deliverables, usage term, fee, brand limits.
Are expectations explicit enough to prevent resentment?
AI may draft retention replies; human removes guilt, pressure, and false promises.
Customer reason, account status, realistic options.
Are we respecting the customer’s decision?
AI may assemble evidence; owner makes final call from readiness, risk, and capacity.
QA results, support plan, rollback path, cash need.
Would delaying one day prevent a public mess?
You are helping with [task]. Use only the facts below. If a fact is missing, ask for it or mark UNKNOWN. Allowed AI role: [draft / organise / compare / summarise]. Human approval boundary: [what must not be finalised by AI]. Evidence provided: [paste facts]. Return: [format]. Before the final answer, list assumptions and risk flags.