# What AI Should Never Run in Your Business, and Why

*Guide — 2026-09-30 — by Mahmoud Zalt*

Six things should never be handed to AI: pricing, binding promises, moving money, people decisions, hard conversations, and licensed advice.

**Short answer.** **Six things should never be run by AI: setting prices, making binding promises, moving money out, decisions about people, relationship defining conversations, and anything requiring a licensed professional.** The test is not difficulty. It is whether the action binds you, whether it is hard to reverse, and whether somebody needs to be accountable by name. An AI Employee should prepare every one of these fully and then stop. Sistava is built to stop exactly there, from 50 per month.

Almost every article about AI at work tells you what to hand over. Far fewer tell you where to stop, which is strange, because the stopping line is the part that protects you. Handing over too little costs you some hours. Handing over the wrong thing can cost you a customer, a contract, or a lot of money.

So this piece is the other half. Six permanent boundaries, why each one holds, and what an AI Employee should do right up to the edge of it. These are not temporary limits waiting for a better model. They are the boundaries of ownership, and a business that gives them away is not more automated, it is unowned.

This matters practically because the boundaries have to be written down somewhere the system respects. In **Sistava** they live in the AI Employee's brief as an explicit stopping rule, so it prepares the work, escalates the decision, and waits. A boundary that only exists in your head is not a boundary.

## At a Glance

- **6** Permanent boundaries, none of them temporary
- **3 questions** Binds you, hard to reverse, needs a name behind it
- **Prepare** What AI does right up to every boundary
- **50/mo** Sistava entry plan with your first AI Employee

## How do you tell whether something crosses the line?

Ask three questions about the action. Does it bind the business to something? Would undoing it be expensive, slow, or embarrassing? Does somebody need to be answerable for it by name? One yes means keep it. Two or three means keep it and write the boundary down explicitly.

Notice that difficulty is not on the list. Plenty of hard work is safe to hand over, and plenty of easy actions are not. Typing a number into a quote is trivial. Being held to that number for six months is not. Sort by consequence, never by effort.

## Why should AI never set your prices?

Because a price is a strategic decision wearing the costume of a number. It carries your margin, your positioning, and what you can afford to deliver. An AI Employee has no way to know that you are deliberately underpricing one client to win a sector, or that this month's costs moved.

It can still do everything up to the decision. It can pull comparable jobs, calculate the margin at three price points, flag that the client asked for a discount twice before, and put a one screen summary in front of you. You spend thirty seconds instead of twenty minutes, and the number is still yours.

## Why should AI never make a binding promise?

Because you will be held to it. Delivery dates, refund exceptions, scope agreements, guarantees, and contract terms all create an obligation that survives the conversation. A message that sounds helpful in the moment can quietly commit you to work you cannot deliver at a cost you did not agree to.

The rule that works is simple: an AI Employee may repeat published commitments and may never create new ones. It can quote your standard delivery window because that is already public. It cannot agree to Friday for this one client, and it should recognize the request and hand it to you.

## Why should AI never move money out of the business?

Because an outgoing payment is close to irreversible and a mistake compounds quietly. Reading financial data is safe and useful. Initiating a transfer is a different category of action, and the convenience saved is measured in seconds while the exposure is measured in whatever the account holds.

So let it do the preparation and none of the execution. It can reconcile, spot the duplicate charge, group this week's supplier bills into one payment run, and show you the total before and after. You review a prepared list and press the button. That last click is not friction, it is the control.

## Why should AI never decide about people?

Because these decisions change someone's life and they need a human who can explain the reasoning, hear the response, and stand behind it. Hiring, promotion, performance, pay, and letting someone go all belong to a named person. Automated judgment about people is also where bias hides most easily and gets noticed least.

Around it, help is welcome. It can organize applications against criteria you wrote, prepare interview questions, handle onboarding paperwork, and make sure nobody's first day is a mess. What it must never produce is a verdict on a person, and no summary it writes should read like one.

## Why should AI never handle the conversation that defines a relationship?

Because those moments are precisely where a person matters most. A serious apology, telling a client bad news, a hard renegotiation, saying no to a long standing customer, or a complaint that has gone past frustration. In each one the other person is judging whether you can be trusted, and a polished automated reply reads as an insult.

The practical version is an escalation rule with teeth. Any message showing real anger, any mention of legal action, cancellation, or a serious failure on your side goes to a human immediately, with the history summarized so you can respond fast. Speed and a real person together are what saves the relationship.

## Why should AI never give licensed advice?

Because legal, medical, tax, and financial advice are regulated for a reason, and the liability lands on your business regardless of who typed it. Even a well meant answer about someone's tax position or symptoms can cause real harm, and the fact that it came from software is not a defence anyone accepts.

Write this boundary in the plainest language you have. It answers process questions, never professional ones, and it says so openly while pointing the person at the right human. A clinic that lets automated replies discuss symptoms has not saved time, it has created an exposure it cannot see.

## What happens when a boundary is missing?

Tobias runs a small IT support company with four engineers. He set up automated first replies for inbound work and, to be helpful, let them quote from his standard rate card. It went well for three weeks, which is exactly how these things go.

Then a request came in described as a simple email migration for twelve users. The reply quoted the standard fixed fee, and the client accepted it in writing that afternoon. The real job involved a fifteen year old server, two mailboxes nobody could open, and eleven days of work. He honoured the number because it was in writing, and the job cost him roughly three times what he charged.

The fix took one line in the brief. It may state published hourly rates and may never quote a fixed price for a project. Any request for a fixed number goes to him with a summary of what was described. Nothing else changed, replies still land in minutes, and the AI Employee still handles most of the intake.

That is the shape of almost every incident like this. Nothing malfunctioned. Nobody was careless. The boundary simply had not been written down, so the system did the most helpful thing available to it. Writing it down is the entire remedy.

## How do you put these boundaries in place?

Write them as a short list of never statements inside the brief, then check them against real output rather than assuming. Boundaries fail quietly, so the only reliable test is looking at what actually went out and asking whether anything in it commits you to something.

### Setting the boundaries so they hold

1. **Write your never list first** — Before any function goes live, write the specific things it must never do in your business. Six lines is usually enough and it takes ten minutes.
2. **Turn each never into a trigger** — Say what makes it stop, not just what it must avoid. "If anyone asks for a fixed price, a discount, or a date, send it to me" is a rule a system can actually follow.
3. **Make escalation fast and visible** — An escalation nobody sees for two days is worse than no escalation. Route it where you will actually notice it and keep the summary short.
4. **Audit real output monthly** — Read twenty real messages that went out. You are looking for one thing only: did anything commit the business to something you did not approve?
5. **Tighten after every near miss** — The first time something gets close to a line, add the rule that day. Near misses are free lessons and they never repeat once written down.

| Never hand over | Why the boundary holds | What AI does instead |
|---|---|---|
| Setting prices and discounts | It carries margin, positioning, and context AI cannot see | Prepares comparisons, margins, and client history for your call |
| Binding promises and contract terms | You are held to it long after the conversation ends | Repeats published terms only, escalates every new commitment |
| Moving money out | Close to irreversible, and mistakes compound quietly | Reconciles, spots errors, and prepares the payment run for approval |
| Decisions about people | It changes a life and needs a named human behind it | Organizes applications, prepares questions, handles onboarding admin |
| Relationship defining conversations | Trust is decided in those moments and a polished reply reads as an insult | Detects them instantly and hands you the history so you can respond fast |
| Legal, medical, tax, or financial advice | Regulated, and the liability lands on your business | Answers process questions and points the person to the right professional |

Held together, these boundaries are not a list of restrictions. They are a description of what you own. Everything above the line is repetition, and repetition is a cost. Everything below it is the judgment that makes the business yours, and there is no version of the future where handing that away leaves you better off.

## Frequently asked questions

## FAQ

### Will these boundaries move as AI gets better?

The first five will not, because they are about accountability rather than capability. Somebody has to be answerable for a price, a promise, a payment, and a person, and better technology does not create somebody to blame. The sixth is set by regulation, which moves slowly and not by your choice.

### What is the single most dangerous thing to hand over?

Anything that quotes a number or agrees a date. Those two create obligations instantly, they are almost always in writing, and they look completely harmless while they happen. Most real incidents start with a helpful reply containing a price or a deadline.

### How do I stop an AI Employee from over promising by accident?

Give it a rule about sources rather than a rule about tone. It may state facts that already exist on your published pages and nothing else, and anything requiring a new number, date, or exception goes to you. Tone instructions do not prevent commitments, source limits do.

### Is it risky to let AI read financial or customer data?

Reading is a different risk class from acting. Reading data to summarize, reconcile, or flag a problem is normal and useful, while initiating payments or sharing records outward is not. Keep the same access discipline you would apply to a new staff member and review who can see what.

### What should happen the moment a boundary is crossed?

Fix the commitment with the customer first, then fix the rule the same day, then check whether anything similar already went out. Honour what was promised if it is reasonable, because the reputational cost of arguing usually exceeds the loss. Then make sure it cannot happen twice.

### Do I need a written policy for a business with two people?

A page is plenty, but write it. Two people means two sets of assumptions, and the never list is what stops one of you loosening a rule the other relies on. It also makes onboarding your third person a five minute conversation instead of a discovery process.

**Tags:** ai-boundaries, ai-employee, business-risk, human-in-the-loop, ai-governance