The AI-Powered AI Chatbot Installation Side Hustle

What is inside, section by section — all 45 of them, in order: 1. A dental practice bot that told a patient their appointment was Thursday at ten past three when it was Tuesday — an invented time, stated with complete confidence, one of eleven similar inventions over three weeks that nobody caught because nobody was reading; and a plumbing company bot at eleven on a Sunday night cycling a customer with water coming through a ceiling through FAQ options with no route to a human anywhere in it. 2. Putting the widget on the page takes an afternoon; the trade is everything that decides whether what it says is true. 3. Six triggers, and the strongest is the same eight questions arriving forever at a business with nobody spare to answer them. 4. Six services, and the recurring one is reading transcripts — the only part nobody has automated and nobody wants to sell. 5. They think it knows things, that it will not make things up, that it replaces answering, and that launch is the end. 6. Rule one: every sentence is a statement by that business, made without anybody reading it first — which is why the refusal wording is approved by name. 7. Rule three: a bot that cannot hand off is a wall with a text box, and the specification is a precondition to launch rather than a feature. 8. Rule four: nothing anywhere flags a wrong answer — no alert fires, no counter increments, and the customer who was misled does not come back to say so. 9. Rule two: the failure mode is invention rather than silence — a fluent, specific, plausible sentence that is simply not true. 10. Not the client's spokesperson, not their compliance function, not their customer service — and no launch without a tested handoff. 11. Five variables decide margin, and one screening question outranks them all: who picks it up when it cannot answer, and when are they there? 12. Count the answers that must be written from scratch — that number sets the price, not the question count. 13. Almost every provider sells deflection percentages and almost none sell the reading, which is where the whole service is. 14. Four lines and four numbers, and the reading ceiling is the one that decides how big this can get. 15. Short list, low floor — and the usage cost you absorb is the line that turns a profitable bot unprofitable quietly. 16. Questions rather than answers, and the first is who is liable when a system you configured states something untrue. 17. Choose on transcript access and refusal behaviour, and re-test both after every platform or model change. 18. You price five things and the client sees one; the answers written from scratch get their own stage because their timeline is the client's. 19. You can never show a transcript and never demonstrate a live client bot — so the question audit is the portfolio. 20. Nine steps, with the handoff specification agreed before any design — the step clients most want to defer and cannot. 21. Two failed attempts then an unconditional handoff, an explicit request for a person honoured first time, and never a persona. 22. Nothing enters without a source, a named owner and an approval date — and anything with a price gets an expiry rather than a review. 23. Ask for what the business will actually act on, in the customer's order rather than the CRM's. 24. Tested during hours, outside hours and at a weekend before launch, because eleven on a Sunday is when it matters most. 25. Every integration's failure behaviour must read "refuses and hands off" — a stale read stated confidently is worse than no read. 26. A test pass that found nothing was not adversarial enough; rewrite it until the bot invents something. 27. Bots get switched off after one screenshot, and the knowledge base is what survives — hand it over as a document they own. 28. The agreement buys the only supervision that exists, and there is no tier below weekly reading. 29. Sign it before the build invoice closes, because afterwards it reads as an upsell rather than as the service. 30. Reading is delegable; deciding what a finding means is not, and the flag comparison is the only real readiness test. 31. Five records, and the incident record and the review log cannot be reconstructed afterwards. 32. Built on measured review minutes rather than estimates, and never averaging a standardised bot with a non-standard one. 33. Platform concentration is the risk that one pricing change hits every bot you support at once. 34. Book the reading first and build the year around it, because it is the commitment that silently disappears. 35. In a trade where nothing flags a failure, procedure is not process discipline — it is the sensor. 36. Grounded answer rate is the master variable; supported bots unread in the last week is the only number whose target is zero at every moment. 37. Break your own bot before you launch anybody else's, and compute a real grounded answer rate before you sell the reading. 38. A model asked to write knowledge base answers writes confident ones, which is precisely the failure you were hired to prevent. 39. Twelve documents for deciding whether to do this: the self-assessment built around whether you will genuinely read transcripts every week, the model comparison grid, the four-rules reference card, the scope worksheet carrying the handoff destination and the must-not-attempt list, the build and review cycle card, the platform standard planner keyed on transcript access and refusal behaviour, the refusal list, the screening card whose first question outranks the five variables, the one-page plan with its reading ceiling, and a practice plan that starts by breaking your own bot. 40. Fourteen documents for setting up: the subscription inventory separating your accounts from clients', the review capacity planner keyed on measured minutes per bot per week, the client agreement checklist whose load-bearing clause is the handoff specification as a precondition to launch, the contracts and liability question set led by responsibility for a wrong statement, the onboarding record requiring the handoff tested during hours, outside hours and at a weekend, the rate card with no per-question pricing, the build standards, the customer conversation and consent register, the integration register whose failure column must read "refuses and hands off", and the support agreement specification. 41. Twelve working documents: the question audit worksheet whose answers-from-scratch count sets the price, the permission checklist, the knowledge base register with sources, owners and approval dates, the conversation design worksheet, the handoff specification with its three test dates, the adversarial test set including false premises and compound questions, the pre-launch checklist, the bot log with weekly review as its own category, the handover pack, the transcript review question set that records the minutes taken, the incident sheet asking how it was found, and the monthly close routine. 42. Twelve documents for the money: the working-hour cost worksheet that keeps weekly review out of overhead and in delivery, cost per bot-month from measured review minutes, revenue per bot read against it, the break-even count tested against the reading ceiling, the rate change notice, the no-guarantee checklist led by never claiming it will not make things up, the wrong answer and correction policy requiring a count rather than an instance, payment terms stating that the review does not pause silently, the profitability sheet, the concentration table, the SOP template, and the monthly quality checklist with its thirty-second refusal spot-check. 43. Twelve documents for finding and keeping work: ten targets qualified on the handoff question, the outreach script that opens with the same eight questions rather than the technology, the second builder agreement and its explicit boundary, the day-thirty survey whose fifth question finds stale entries before a customer does, confidentiality rules treating a gap log of verbatim customer questions as personal data, the six client conversations, the case study permission request, support plans with no tier below weekly reading, correction templates that give the count and name the cause, the incident record tracking found-by-review against found-by-customer, and the income tracker watching usage cost as a share of support income. 44. The KPI scorecard led by the grounded answer rate and the count of supported bots unread in the last week, the second builder onboarding checklist, the thirty-day launch plan that forbids launching a client bot, the ninety-day growth plan ending in a measured reading ceiling, the one-year review worksheet whose first number is weeks with a bot unread, and the complete library of all 148 prompts. 45. The four rules restated, the two numbers that run the business, the instruction to read one week of transcripts and time yourself because that measured number sets everything, and the closing note that the platforms have automated the easy half and have no incentive to sell the reading. Every worksheet, register and checklist is included, and every financial figure in the book is left blank on purpose — you fill them in from your own measured costs rather than someone else's guesses. Instant download. Yours to keep and print as often as you like.