By The Tool Trial Editorial Team
The question is not whether AI can write a follow-up email. It can. The question is what happens when it writes a good one containing a price nobody quoted, a slot nobody has open, or a promise nobody authorized — and it goes out because it read well.
This is a process for using an AI tool as a drafting assistant only. A person prepares the task, strips it to the minimum, asks for one draft, checks every sentence against verified facts, approves or rejects by name, sends manually, and records what happened afterwards.
Nothing here is tested. No result, timing, or failure rate is claimed, and this article does not say that AI follow-up is safe, proven, or effective. It describes a gate, not an outcome.
This is practical editorial guidance, not legal advice.
What the AI is for, and what it is not
It drafts. It does not decide. Every business decision in this process belongs to a named person, and the list further down is explicit about which ones.
It is not a source of facts. Anything it states that you cannot trace to your own records is unverified text, however confident it sounds.
It does not understand your business. It has your prompt and its training. It does not know your service area, your calendar, your prices, or who asked you to stop contacting them.
It sends nothing. No integration, no API, no automation, no scheduled dispatch. A person copies the approved text and sends it.
It guarantees nothing. Not accuracy, not tone, not consistency between one draft and the next.
No specific AI provider is required. The process assumes only a tool that accepts text and returns text. This article names none, recommends none, and describes no account, plan, or configuration.
The six-stage approval workflow
1. PREPARE — define the permitted task
Entry condition: a real inquiry exists and a person has decided a follow-up is appropriate.
Permitted input: the task itself — what kind of message, roughly how long, what it must not say.
Prohibited input: anything that presumes the decision to contact has been made by the tool.
AI role: none yet.
Human responsibility: decide the message is warranted, and that a permitted channel exists.
Required evidence: the inquiry record, and the channel the person actually used.
Pass condition: the task is written down in one sentence before anything is pasted.
Stop or reject: if you cannot state the task in one sentence, you are not ready to draft it.
2. MINIMIZE — remove unnecessary or sensitive information
Entry condition: a defined task from stage 1. Permitted input: the few verified facts the message needs — the service asked about, the area status, the agreed next step. Prohibited input: the full inquiry history, the customer’s name where a placeholder will do, account or payment details, health information, identification, third-party data, credentials. AI role: none yet. Human responsibility: strip the input to the minimum, replacing specifics with placeholders. Required evidence: a short fact list you can point to, line by line. Pass condition: every item in the input is necessary and verified. Stop or reject: if removing an item breaks the draft, the message may need a person to write it directly instead.
3. DRAFT — ask for one unapproved draft
Entry condition: a minimized, verified fact list. Permitted input: the fact list, the task, and an instruction to invent nothing. Prohibited input: an invitation to “fill in the gaps”, “make it compelling”, or add anything not supplied. AI role: produce one draft, explicitly unapproved. Human responsibility: ask once, and resist asking for variations to pick a favorite. Required evidence: the draft itself, kept as text, not sent anywhere. Pass condition: a single draft exists and nothing has been sent. Stop or reject: if the tool returns several options, treat that as one attempt and pick nothing yet.
4. VERIFY — compare every statement with verified facts
Entry condition: one unapproved draft. Permitted input: your records, the inquiry, and the fact list from stage 2. Prohibited input: the tool’s own assurance that the draft is accurate. AI role: at most, listing its own claims for a person to check — never confirming them. Human responsibility: read every sentence and ask “where does this come from?” Required evidence: each factual statement traced to a record, or struck out. Pass condition: no sentence survives that cannot be traced. Stop or reject: any invented price, date, availability, or commitment sends the draft back or into the bin. A single invented fact is a reject, not an edit, if it suggests the input was misread.
5. APPROVE — a named person approves, revises, or rejects
Entry condition: a verified draft with every statement traced. Permitted input: the exact final text. Prohibited input: an approval given on the gist, a summary, or an earlier version. AI role: none. The tool cannot approve its own output. Human responsibility: a person with authority records APPROVE, REVISE, or REJECT against that exact text, and their name. Required evidence: the decision, the name, and the date. Pass condition: an explicit APPROVE on the text that will actually be sent. Stop or reject: if the wording changes after approval, it needs approving again.
6. SEND AND RECORD — a person sends manually and updates the tracker
Entry condition: an approved exact text. Permitted input: the approved message and the recipient’s permitted channel. Prohibited input: any automation, scheduler, or bulk tool. AI role: none. Human responsibility: send it by hand, then record what was sent and when. Required evidence: the sent message and the updated record. Pass condition: the tracker reflects reality within the same day. Stop or reject: if the do-not-contact status changed while drafting, do not send — and the record is updated instead.
Decisions the AI does not make
None of these may be delegated, and none is answerable from a prompt. A person decides:
whether consent exists · whether the channel is valid · whether to make contact at all · how many times to contact · when to stop · whether someone is interested · whether an inquiry is legitimate · prices · discounts · availability · service area · commitments · outcomes · refunds · compensation · legal conclusions · whether a call is booked · whether a do-not-contact request can be set aside — it cannot · and whether a message is approved.
Each of these needs verified information and a person who is accountable for being wrong.
Privacy and data minimization
Send the least the task needs. As an editorial safeguard, we recommend giving an AI tool a short fact list with placeholders rather than a customer record.
Never put into an AI tool: passwords or credentials · financial or payment information · medical information · identity documents · a customer’s private information · full conversation histories · third-party data · confidential business information · personal details the message does not need.
Use placeholders in the input, not just the output. [First name], [Service requested], and [Approved next step] carry the shape of the message without carrying the person.
We make no claim about how any tool stores, retains, deletes, or protects what you send it. That depends on the provider, the plan, and the settings, and this article does not assess any of them. Check the terms of whatever you use before deciding what is acceptable to paste.
Five prompt patterns
Each pattern asks for one unapproved draft, forbids new facts, requires placeholders, and ends in human review. None permits sending.
P-01 — Create a draft from verified facts
Purpose: turn a short fact list into a first draft. Required inputs: the task, the verified fact list, the tone constraint. Prohibited inputs: anything unverified, and any instruction to embellish. Prompt skeleton: “Using only the facts below, write one short follow-up email. Do not add any fact that is not listed. Use placeholders where a detail is missing. Do not invent prices, dates, or availability. Return one draft.” Expected output: one short draft using only supplied facts. Human checks: every sentence traced; placeholders intact; no added specifics. Reject condition: any statement not in the input.
P-02 — Ask for missing information without inventing
Purpose: draft a message that requests one or two missing facts. Required inputs: what is known, and precisely what is missing. Prohibited inputs: guesses about why it is missing. Prompt skeleton: “Write one short email asking only for the two items listed as missing. Do not speculate about the reason. Do not assume anything about the sender. Return one draft.” Expected output: a brief request for exactly those items. Human checks: nothing sensitive is requested; the ask is minimal. Reject condition: it asks for anything beyond the listed items.
P-03 — Shorten or adjust tone without changing facts
Purpose: tighten a draft you already verified. Required inputs: the existing draft and the specific change. Prohibited inputs: permission to “improve” freely. Prompt skeleton: “Shorten the draft below without removing any factual statement and without adding one. Keep every placeholder exactly as written. Return one version.” Expected output: a shorter draft with identical factual content. Human checks: compare fact-for-fact against the version you approved into this step. Reject condition: a fact disappeared, changed, or appeared.
P-04 — Compare a draft against an approved fact list
Purpose: surface mismatches for a person to judge. Required inputs: the draft and the fact list. Prohibited inputs: the expectation of a verdict. Prompt skeleton: “List each factual statement in the draft below and mark whether it appears in the fact list. Do not correct anything. Do not conclude whether the draft is accurate.” Expected output: a list of statements with a match or no-match marker. Human checks: the person decides what the mismatches mean. The output is a prompt for attention, not a finding. Reject condition: it offers a conclusion, a fix, or an approval.
P-05 — Identify unsupported claims before human review
Purpose: flag likely invention before a person reads closely. Required inputs: the draft and the fact list. Prohibited inputs: authority to remove anything. Prompt skeleton: “Identify any statement in the draft that is not supported by the fact list, including implied availability, pricing, or commitments. List them. Change nothing.” Expected output: a list of candidate unsupported statements. Human checks: treat the list as incomplete — absence from it is not evidence of accuracy. Reject condition: it edits the draft or declares it clean.
The twelve-point approval checklist
Run before any message leaves.
| # | Check |
|---|---|
| 1 | Correct recipient and permitted channel |
| 2 | Verified reason for contact |
| 3 | Every factual statement supported |
| 4 | No invented action, promise, or outcome |
| 5 | No sensitive or unnecessary information |
| 6 | No unauthorized price, discount, or availability |
| 7 | Tone appropriate and non-pressuring |
| 8 | Placeholders fully and correctly replaced |
| 9 | Do-not-contact status checked |
| 10 | Next step authorized |
| 11 | No legal or compliance conclusion |
| 12 | Named human decision: APPROVE / REVISE / REJECT |
APPROVE here means one thing only: a person authorized that specific email draft to be sent by hand. It is not approval of anything else.
Three synthetic examples
All three are fictional. No real business, customer, inquiry, transaction, or tool output is involved, and none is evidence of how AI performs at this task.
Example 1 — approved after verification. A fictional two-person landscaping business has an inquiry with the service, the area status, and an agreed next step recorded. The owner writes the task, reduces it to four facts with placeholders, and asks P-01 for one draft. Every sentence traces to the list. The checklist runs clean, the owner records APPROVE with their name, sends it manually, and updates the record the same day.
Example 2 — rejected for invented availability. A fictional appliance repair business asks for a draft from three verified facts. The draft reads well and offers “a slot this Thursday” — which was never in the input and is not open. This is a reject, not an edit: the invented slot suggests the input was misread, so the draft is discarded and P-01 is run again with a tighter instruction. Nothing is sent.
Example 3 — stopped before drafting. A fictional cleaning business is preparing a follow-up when the record shows a do-not-contact request received the previous day. The process stops at stage 1. No prompt is written, no draft is generated, and no message is sent. The record is updated and the lead leaves the follow-up queue permanently.
Common mistakes
Asking for several drafts and picking the nicest. That selects for fluency, which is exactly the failure you are guarding against.
Letting the tool check its own work. P-04 and P-05 list candidates for a person. They do not verify, and a clean result means nothing.
Pasting the whole record “so it has context”. Context is what you deliberately chose to include.
Approving the gist. Approval attaches to the exact text that will be sent.
Treating a good draft as a verified draft. Fluency is not accuracy, and a confident sentence is not a sourced one.
Where evidence is still required
This article describes a design, not a tested system. The following remain open and are marked as such deliberately:
| Claim | Status |
|---|---|
| How often an AI tool invents a fact in this task | [EVIDENCE REQUIRED] |
| Which failure modes appear most, and under which prompts | [EVIDENCE REQUIRED] |
| Whether the checklist catches them, and at what rate | [EVIDENCE REQUIRED] |
| How long the six stages take in practice | [EVIDENCE REQUIRED] |
| Any comparison between tools | [EVIDENCE REQUIRED] |
No test has been run for this article, and no evidence from any other Tool Trial test transfers to it. Until a documented test exists with its own evidence log, the honest position is that this gate is reasoned, not measured.
Final takeaway
The value is not that AI writes faster. It is that a draft arrives with no authority attached — and the process keeps it that way until a person puts their name to the exact words. Minimize what you send it, ask once, trace every sentence, approve explicitly, and send it yourself.
Related resources, once they exist: the Simple Lead Follow-Up Tracker for Small Businesses; A Simple Lead Follow-Up Workflow: From New Inquiry to Booked Call; Lead Follow-Up Email Templates for Small Service Businesses; and our Methodology, Editorial Policy, and Contact pages — all planned, none published yet.
Sources
- National Institute of Standards and Technology — AI Risk Management Framework:
https://www.nist.gov/itl/ai-risk-management-framework - Federal Trade Commission — CAN-SPAM Act: A Compliance Guide for Business:
https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business
Both were opened and read on 2026-08-14.
A. What these sources support. NIST states that its framework “is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems” — that is, a federal framework for AI risk exists, it is voluntary, and it concerns trustworthiness considerations. The FTC page supports the commercial-email context referenced above: that an opt-out must be honored “within 10 business days”, that header information must not be false or misleading, and that “The subject line must accurately reflect the content of the message.”
Scope, stated plainly. The NIST page does not address lead follow-up, prompts, or approval gates — searched and not present. The FTC page says nothing about AI at all, so nothing here about AI is attributed to it; it is cited only for email. We looked for an official FTC page on AI claims and did not verify one for this article, so no such claim is made.
B. The Tool Trial editorial safeguards. The six stages, the eight fields per stage, the list of decisions reserved to people, the data-minimization rules, the five prompt patterns, the twelve-point checklist, and the requirement for a named human approval before manual sending. None of it comes from NIST or the FTC, and neither source is cited for any result, accuracy, timing, productivity, saving, or rate.
This article is not legal advice, and following this process does not establish compliance with CAN-SPAM or any other rule. It also makes no claim that AI-assisted follow-up is safe, proven, or effective.