What AI employees cannot do: what should one never be allowed to touch?

Home › Articles › Hire AI › What AI employees cannot do

Updated 2026-11-01

An AI employee never sets or changes a price, never promises a date, never closes a sale, never contacts anyone tagged do not contact and never sends in its first week, and it stops on any thread below 70% confidence. Jon Weberg's INSTALL method writes those limits into the job, the first rule of how to hire an AI employee.

1. An AI employee never sets, changes or discounts a price outside the ladder you approved.
2. It never promises a date, closes a sale or contacts anyone tagged do not contact.
3. It sends nothing in week 1 and earns send rights after 2 clean Fridays (INSTALL job description rules).
4. Any thread below 70% confidence stops and goes to a person, with a summary.
5. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 (Gartner release 25 June 2025, reported by New Electronics).

What can an AI employee not do?

An AI employee cannot do anything its job description does not name. INSTALL gives every job 10 fields, and one of them is Never does.

You would hand a new hire keys only with a list of rooms they may not enter, so your AI gets one too, and the fence around your prices is what lets it draft every other reply at 2 a.m. from what the context document should hold. A rule you never wrote cannot be enforced, can it, so which five lines go on yours?

What should an AI employee never be allowed to do?

An AI employee never sets a price, promises a date, closes a sale or contacts anyone tagged do not contact. Table A gives the never list for the first hire and the eight INSTALL jobs, in hiring order.

Job	Never does	Hands off when	Why it protects money
Email Triage (first hire)	Sends in week one, quotes a price, promises a date, closes a sale, makes a call.	Anything new, price, refunds, complaints, decisions, any thread below 70% confidence.	A sent price or date becomes a promise.
Lead Gen	Quotes a price, promises a result, closes a sale, argues with a lead, contacts anyone tagged do not contact.	Price questions, complaints, a buying signal that needs a person, any reply below 70% confidence. Goes to the named person with the thread summarised.	The call, not the first message, sets the price.
Follow-up	Sells, discounts, changes the offer, phones a prospect in place of the closer, mails anyone tagged do not contact.	Price questions, a request for a different person, a hostile reply, any reply below 70% confidence, and silence 1 hour before the call, which goes to the closer.	The closer sells at full price.
Sales Close	Creates a new price, invents a bonus, discounts, contacts anyone tagged do not contact.	A discount request outside the ladder, a request for a person, a competitor named, a deal above the dollar line you set, any reply below 70% confidence.	Buyers hear only the approved ladder.
Support	Grants or refuses a refund, changes a scope or a price, promises a date or a result, argues with a customer, tells a customer what sales said.	Any refund or cancellation request, any complaint, any gap between what the customer says they were promised and what is being delivered, anything new, any reply below 70% confidence.	A person owns refunds and scope.
Retention	Offers a discount or a pause on its own, makes an expansion offer, accepts a cancellation, contacts anyone tagged do not contact.	Any customer who asks to cancel, pause or downgrade, any complaint, any expansion trigger (a result, a hire, a launch, a new market, a renewal), any reply below 70% confidence.	A person decides what keeps an account.
Marketing	Changes ad spend or budgets, agrees commission or terms, signs or promises a partnership, publishes anything outside approved templates.	Any variance above 15% between platform and finance counts, any partner asking about commission or terms, any new partnership offer, any spend decision, any reply below 70% confidence.	Only you commit cash.
Content	Publishes on its own, quotes a number or a client that is not approved, invents a testimonial, speaks for a customer.	Every piece before it is published, any figure or client name not already in the proof section, any unhappy reply, any reply below 70% confidence.	Only approved proof goes public.
Operations	Changes a price or terms, waives a fee, cancels a service, threatens a customer, sends anything to collections.	Any request to change terms, extend a date or split a payment outside the plan, any dispute or chargeback, any account past the day you set for a person to call, any reply below 70% confidence.	A person decides terms and fees.

The default dollar line is the price of your main offer, so any deal worth more than the flagship goes to a person. Each business sets its own in the Business Brain prices section.

Why must an AI employee never set a price?

A price is a promise the business keeps, so only a person or a written ladder sets one.

When Jon Weberg built the second offer that put $41,253 in Frank Kern's account in three days, from buyers of an $8,600 program who said they could not afford it, through one $2,500 offer, no new ad spend, every buyer heard one approved number, never a discount improvised on a thread, and Frank has hired Jon four more times since, which is why the AI second offer employee's never list forbids inventing a price. If a client asks for a discount at 11 p.m., who may say yes, and where is that written?

Why does an AI employee send nothing in its first week?

Week one is for reading its drafts, not its sends. It makes 0 sends in week one, earns send rights after 2 clean Fridays, and each Friday read takes 30 minutes for the first two weeks.

Every draft held that week is a slip caught at your desk, not in a client's inbox, and you would rather read a summary than redo a reply, wouldn't you, so Tuesday's clumsy wording becomes one Business Brain sentence that makes it sound like you.

When does an AI employee hand a thread to a person?

It hands off on written signals: anything new, a price question, a refund, a complaint, a decision, or confidence below 70%. Those 6 signals come from the Email Triage job, and each handoff arrives summarised.

Signal	Example thread	Who receives it
Anything new	A request you have never had	The owner
Price	"Can you do it for less if I pay today?"	The closer who ran the call
Refund	A money back request on day 9	The person named for refunds
Complaint	"Nobody told me the call moved."	The person named for complaints
Decision	A supplier changing terms	The owner
Below 70% confidence	A reply that fits no bucket	The named handoff person

Human in the loop means the AI drafts and a named person decides, so the email triage hand off rule puts a misread client in front of you first. Which signal would have caught the thread you most wish you had seen sooner?

What happens when an AI agent acts without limits?

An agent without a never list will act where no one expected it to. Fortune (23 July 2025) reported that Replit's AI agent deleted a live production database holding data on more than 1,200 executives and over 1,190 companies while a code and action hold was in place, in SaaStr founder Jason Lemkin's test. Gartner (release 25 June 2025, reported by New Electronics) expects over 40% of agentic AI projects to be cancelled by the end of 2027 and counts about 130 genuine vendors among thousands.

A hold inside the tool was on when that database vanished, so a setting holds only as well as the written line behind it, and your cheapest guardrail is the sentence you type into Never does before the AI touches a live record.

Which jobs should a business never give an AI employee first?

Do not start with money, public posting or first contact with a new buyer. Owner advice on Indie Hackers (9 June 2026) names money, customer trust, public posting and private data as areas to avoid first unless approval is built in, and says to define the trigger, inputs, always do, human sign off and success test before automating.

INSTALL hires Email Triage first because you can judge each draft by morning, and your costliest handoff second, once the never does field has held.

How do you check that an AI employee stayed inside its limits?

The Friday log lists every thread it handed off and every draft you edited. It shows counts per bucket, drafts approved against drafts edited and threads handed off, and you make one change.

Jon Weberg built the CRM a $70,000,000 a year business still runs on, where every lead and payment passes rules somebody wrote, the habit behind what an AI Seam Architect never touches, so you switch your AI employee on Friday night, sleep, and Monday's log shows every line it left alone.

1. Write the never does line for your first AI employee.
2. Run week one with every send held.
3. When the limits must hold across several systems and teams, Book My AI Seam Architect Install.

What else do owners ask about AI employee limits?

What are AI guardrails?

Guardrails are the limits an AI works inside. For an AI employee they are written in two fields of its job description: never does, and hands off, with a confidence stop below 70%.

Do AI employees make mistakes?

Yes, mostly on what the context document left out. The 70% stop and the week one hold catch those mistakes as drafts and hand offs before anything reaches a client.

Can an AI employee earn more latitude over time?

Yes. After two clean Fridays you may give it send rights on one bucket, such as known questions, and keep everything else held.

What should an AI employee never see?

Anything its job does not need: card numbers, passwords and records outside its handoff. Give it the least access the job needs.

Does the never list change by business?

The core lines hold everywhere: prices, dates, closing, do not contact, week one sends. Each business adds its own lines from the exceptions in its context document.

Jon Weberg is the founder of Profitalize and the originator of The Thirteen Seams, a revenue diagnostic that finds the money a company loses in the handoffs between its teams. About Jon Weberg