This is the fill-in companion to Give a program a job it can't finish, and no way to say so, and it will find another way out. That article is the why. This is the page itself.
What you end up with: half a page, in your words, that says what your assistant may do on its own, what it may do only after you say yes, what it never touches, and what it says when it gets stuck. You hand that page to whoever installs it. Forty-five minutes, and twenty of those are one table.
What this is not. It is not a decision about whether to get an assistant at all. That question is arithmetic: after-hours calls, what a client is worth, hours given back. It is worked out, with a calculator attached, in the AI agent ROI guide. Different conversation, and it goes first. This one picks up the moment you have said yes.
Before you start, put five things on the table
Four of these five are not ours. They are the pre-work list from Monday's resource, the fifteen prompts for the calls nobody scripted, and if you ran that page you already gathered them. Go and get what you pasted in there rather than doing it twice.
- Your hours. The real ones, including whether anybody actually picks up on Saturday.
- Your service list with prices, however you keep it. If you publish none, that is an answer too, and it changes column one.
- What you turn away. The jobs you refer out, or cannot touch.
- How long you actually take to call somebody back. The truth, not the goal.
- The name of the person who can make a decision when you are not there. This one is ours. Monday does not ask for it, and everything in decision 1 and column two depends on it.
Decision 1: The Exit
The one sentence your assistant says when it cannot finish, and the person it hands to.
The words are already written. Go and get them.
This page does not ask you to write the sentence, because Monday's resource already writes it, in your own words, and writing it twice gets you two versions that disagree. Open the fifteen prompts for the calls nobody scripted and go straight to Block 4:
- Prompt 9, "When the script stops working" gives you up to four signals that a prepared answer has stopped working, each one in the caller's own words.
- Prompt 10, "The handoff" writes the line itself, in two versions: one for when the person who can decide is available, one for when they are not.
That is the wording. Give that page a proper sitting rather than a glance, then come back here holding it.
The part that is only true for a program
Here it is in one line, and it is the reason this guide exists.
A rule written for a person to say out loud is not inside the program.
Your team picks up judgment by being near you. They hear you turn a job down. They watch your face when somebody asks for a date you cannot promise. None of that travels. A program starts every single conversation knowing only what somebody typed into it, and if nobody typed in the exit, there is no exit.
So a sentence taped next to the phone is not the finished job. It has to become an instruction, in three parts:
| Part | The person version | The program version |
|---|---|---|
| When | They know when the call has gone sideways | You have to name the moment in words somebody could tick off a list |
| What | They say something close enough | The exact sentence, word for word, every time |
| Then | They walk down the hall | It stops, and the conversation goes to a named person |
Four checks that the words actually got in
The sentence is the easy half. The half that decides anything is whether it survived the trip from your page into the thing answering at nine on a Saturday night.
Four checks. The first one you run on paper, before anybody installs anything. The other three you run yourself, on the live assistant, in about five minutes total.
1. The underline pass. On paper, five minutes, before handover.
Read your handoff line from prompt 10 slowly, and underline every part of it that only works because the reader already knows your business. The name that means something. The "we'll get you on the schedule" that assumes a schedule everyone in your shop can picture. That underlined material is exactly what did not travel. Your team fills those gaps by standing near you. The program has nobody to stand near. So the underlining is not a criticism of your sentence, it is the list you hand over with it.
2. The cold start check. One minute.
Open a brand new conversation with your assistant, from nothing, the way a stranger would. Do not remind it of anything. Then trigger the exit. If the exit shows up in the chat where you just finished explaining it and does not show up in a fresh one, you now have a specific question to ask: where was that rule written down.
3. The stuck call test. Two minutes, and do it twice.
Message or call it yourself with something you know it cannot answer. Do it twice in a row, because once can be luck. What you are watching for, both times: it says your sentence, near enough word for word, and then it stops.
4. The push back test. One minute.
After it gives you the exit, push once, the way a real caller does. "Can't you just give me a rough idea?" Ask a second time, slightly friendlier. What you are watching for is whether the exit holds under the push or folds and produces a number.
What you see, and what to ask next
| What you see | What to ask whoever installed it |
|---|---|
| It answers anyway, helpfully, and never mentions a person | Where those signals are configured, and whether the assistant is allowed to stop answering at all |
| It says your sentence and then keeps talking | Whether the stop can be configured, so it waits for a person instead of carrying on |
| It says it in the chat where you explained it, not in a fresh one | Where that rule was written down, and whether it lives outside the conversation |
| It holds the first time and folds on the second ask | Whether that went in as a fixed instruction or as a preference that can be talked around |
| It says it, stops, and names your person, twice, from cold, under pressure | Nothing for now. Go to decision 2 |
Decision 2: The Three Columns
What it may do without asking you. This is the heart of the sheet, and it has to come from you. Your installer wires up whatever you decide. What they cannot do is decide it, because only you know what it costs your business when the wrong thing goes out without you.
Twenty minutes. Three columns. Be specific enough that somebody who does not work for you could take a new task and put it in the right column without asking.
The three columns, filled in for three businesses
Boat yard or marine repair
| On its own | Only with your yes | Never |
|---|---|---|
| Hours, location, gate and dock instructions | Any figure that is not on your published rate sheet | Take a deposit or any payment |
| Whether you work on that engine brand or that hull length | Any haul out or delivery date | Promise the boat is ready for a specific weekend |
| Take a name, a phone number, the make and length of the boat | Moving a job already on the schedule | Discuss another owner's boat |
| Book a walk through into a slot you already marked open | Anything about warranty or who pays for what | Talk to an owner who has already complained |
| Send the yard address and what to bring | Anything touching an insurance survey or claim | Anything about a job that ended badly |
Dental practice
| On its own | Only with your yes | Never |
|---|---|---|
| Hours, location, parking | Any quote for treatment | Anything that reads as advice about pain, symptoms or medication |
| Whether you are taking new patients | Whether a specific plan covers a specific procedure | Anything about a patient's records, history or treatment |
| The cost of a published, fixed-fee first visit | Moving or cancelling a booked appointment | Take payment |
| Take a name and a number | Anything about treatment already begun, here or elsewhere | Talk to a patient who has already complained |
| Book into an open new patient slot and send the forms | Anything sent under your name to more than one person at a time | Anything a licensed person should be saying |
Accounting practice
| On its own | Only with your yes | Never |
|---|---|---|
| Hours, and deadlines you already publish | Any fee quote | Anything that reads as advice on a specific tax position |
| Which services you offer and which you do not | Any statement about what somebody will owe or get back | Anything about a client's numbers or documents |
| Take a name and a number | Moving a filing appointment during season | Take payment |
| Book a first call into an open slot | Anything about an extension or a late filing | Talk to a client who has had a letter from the tax authority |
| Send the intake checklist | Anything sent under your name to a list | Talk to a client who has already complained |
Now yours
| On its own | Only with your yes | Never |
|---|---|---|
The middle column is where this gets decided
A middle column that runs on trust is not a middle column. An optional approval is a suggestion, and a suggestion is what gets skipped on a busy Tuesday. The version that holds is the one where the thing simply cannot go ahead until a person says yes.
Write it into the sheet in exactly those words, so your installer cannot read it as a preference:
For every item in column two, the assistant cannot proceed until a named person says yes. Not a reminder, not a note in the file. It stops and waits.
Be honest about column three
Half of what belongs in column three is decided by where your customer information actually lives and who else can reach it. If you have never worked that out for your own business, you rent the tool, you own the asset is the question underneath it.
And if you have already been through the seven-day lockdown guide, you have done more of this than you think. Start from what you wrote there rather than from an empty table. Its fourth layer asks four things: give every tool the minimum access; know what an app can read and do before you connect it; keep customer information out of public chats; and keep a person approving anything that sends, deletes or pays. Its fifth layer adds alerts, an incident plan, a periodic review, and an emergency contact.
What that guide gives you is the question, every quarter, across five layers. What this page gives you is the answer written down for one of them: the actual columns, in your words, on a page somebody else can be handed.
Decision 3: Who Looks, and on What Day
One name. One weekday. That is the whole decision, and it is the cheapest of the three.
The person: _______________________ The day: _______________________ The slot: _______________________ (put it in the calendar as a real appointment)
What that person actually does
Twenty conversations. Twenty minutes. Three questions, and nothing else:
- Did it answer anything it should have handed over?
- Did it promise anything you would not have promised?
- Did anybody hang up, or stop replying, mid conversation?
Then one line written down, even if the line is "nothing this week".
Two things that decide whether this survives past the first month
If the person is you, protect the slot. A habit without a calendar entry lasts about two weeks, and this one is quiet enough that nobody will remind you.
If the person is on your team, tell them out loud that finding a bad answer is the win. Otherwise they will read twenty conversations, report that everything looks fine, and mean it. Say the sentence plainly: "I am asking you to find the ones that went wrong. If you find none for two weeks running, tell me that too, because then something is off with what we are reading."
One thing to check on the very first pass: did the assistant ever hand a conversation to a person? If it never once did, that is not a clean sheet. It means the exit from decision 1 is not reaching it.
The Finished Sheet
This is the half page. Fill in the blanks, delete the guidance, and you are done.
Permissions sheet - [your business name] - [today's date]
1. The exit When: [your signals, from prompt 9] It says, word for word: "[your line, version A, from prompt 10]" Then: it stops, and the conversation goes to [name]. If [name] is not available: [your line, version B, from prompt 10]. Checked from a cold start, twice, and once under a push back: [date].
2. What it may do On its own: [your column one] Only after [name] says yes: [your column two]. For every item here, the assistant cannot proceed until a person says yes. It stops and waits. Never, under any circumstances: [your column three]
3. Who reads what it did [Name] reads twenty conversations every [weekday], and answers three questions: did it answer something it should have handed over, did it promise something we would not have, did anybody drop out mid conversation.
4. When this page changes Whenever we add a service, change our hours, change who answers, or find something in column two that went ahead without a yes. Next review: [date].
Handing It Over
Copy this into a message to whoever installs or manages your assistant, with the sheet underneath it.
Hi [name],
Before we switch this on, here is the page that says what it is allowed to do. I have written it rather than left it to the setup screen, so there is one version we both work from.
Three things I need in place, and I would rather hear now if any of them is hard:
- The exit sentence, stored word for word, with the conversation stopping after it and going to a named person.
- Column two held so the assistant cannot proceed without a yes from a person.
- A way for me to read a week of conversations without asking you for them.
The sheet is below. If anything on it is impossible or expensive to do the way I have written it, tell me which line and why, and I will decide rather than guess.
Thanks, [your name]
Four questions for whoever installs it
Ask these once, early, and in this order.
- "Show me where that sentence lives." You want to see the actual words, in the actual place. If nobody can point at them, they are not there.
- "After it says the sentence, does the conversation stop?" Stopping is the part that gets skipped. A handoff that keeps talking is not a handoff.
- "What stops it doing a column two item without a yes?" If the answer is some version of "we told it not to", ask again. Being told not to is what your sheet says. You are asking what holds it.
- "Where do I read what it said, and can I get a week of it without calling you?" Decision 3 is only real if the answer is yes.
