What should AI never be allowed to do in sales?

There probably isn't one universal list of things AI should never be allowed to do in sales.

Sending a routine meeting confirmation is very different from responding to a complaint.

Updating a customer's phone number is very different from marking a major opportunity as lost.

Retrieving an approved price is very different from deciding to offer a 25% discount.

AI might technically be capable of all of them.

That doesn't mean it should have the same authority over all of them.

A better question is:

WHAT HAPPENS IF THE AI GETS THIS WRONG?

The greater the consequence, the stronger the case for human control.

Capability isn't authority.

Capability isn't authority

This distinction runs through everything we build around agentic selling.

An AI system may be capable of:

• reading customer emails,
• writing responses,
• changing CRM records,
• sending messages,
• creating quotes,
• booking meetings,
• and updating opportunities.

But:

CAN DO

does not automatically mean:

MAY DO.

The business decides the second one.

Start by separating access from action

Imagine you connect AI to your CRM.

People sometimes describe that as:

"The AI has CRM access."

That's too vague.

There are two completely different questions.

Access and authority should be designed separately.

INFORMATION ≠ PERMISSION.

Four questions determine how much control AI should have

For every action, ask:

If this is wrong, how much does it matter?

Can we easily undo it?

Can the system reliably know when the action is appropriate?

Is this a decision the business actually wants software making?

Those four questions are more useful than a blanket rule that says:

"AI should never send emails."

Because sometimes sending an email is trivial.

Sometimes it absolutely isn't.

Illustrative only. Authority depends on your business and workflow.

Example: sending an email

Consider these two messages.

"Your meeting is confirmed for Tuesday at 10am. Here's the link."

Low ambiguity.

Easy to check.

Limited consequence.

Highly reversible.

Now:

"We're sorry you're unhappy. We've reviewed your complaint and agree to refund the full project fee."

That's customer communication too.

But it contains:

• a judgement,
• a financial commitment,
• and potentially wider consequences.

Same action:

SEND EMAIL.

Completely different authority decision.

So what should AI be particularly careful with?

There are several areas where consequences can increase quickly.

1. AI shouldn't invent prices

AI may be able to:

• retrieve approved prices,
• calculate from defined pricing rules,
• or prepare pricing information.

That's different from making up a price because one isn't available.

For example:

Approved installation price: £2,500.

AI retrieves:

£2,500

No approved installation price found.

AI estimates:

£2,250

because that seems reasonable.

If the price isn't known:

DON'T GUESS.

Pricing itself has several authority levels

Don't treat "pricing" as one permission.

More commercial judgement → more human control.

Those are very different jobs.

The further down the list you go, the stronger the case for human judgement.

2. AI shouldn't make commitments it isn't authorised to make

A customer asks:

"Can you guarantee this will be delivered by 1 November?"

The AI searches the available information.

It finds that normal delivery is approximately six weeks.

It replies:

"Yes, absolutely."

Problem.

An estimate is not necessarily a commitment.

A sensible workflow might instead say:

That's the system doing its job properly.

3. AI shouldn't turn uncertainty into certainty

This is one of the most important behaviours to control.

Imagine a prospect says:

"We'll probably move ahead once Finance has looked at it."

AI updates:

Verbal agreement

But that's not what the customer said.

They said:

probably.

and:

once Finance has looked at it.

A better interpretation might be:

Positive interest expressed.

Finance review still required.

Review required.

UNCERTAIN ≠ CONFIRMED.

Your AI system should be allowed to preserve uncertainty.

4. AI shouldn't silently make consequential CRM changes

Some CRM updates are low consequence.

Old → New.

Probably straightforward.

Others aren't.

£15,000 → £150,000.

Proposal → Lost.

Sarah → James.

Active customer → Closed.

These changes can affect:

• forecasting,
• reports,
• automation,
• follow-up,
• commission,
• and other systems.

The more consequential the field, the more useful review becomes.

A CRM permission model might look like this

The exact model will vary by business. That's the point. You design it.

5. AI shouldn't delete important information casually

Deletion is different because recovery may be difficult or impossible.

An AI system might identify:

• duplicate contacts,
• old tasks,
• irrelevant records,
• or information that appears redundant.

That doesn't necessarily mean:

DELETE IT.

A safer pattern may be:

• IDENTIFY
• RECOMMEND
• QUEUE FOR REVIEW
• DELETE IF APPROVED

And if the AI doesn't need deletion authority to do its job:

don't give it deletion authority.

6. AI shouldn't independently handle every complaint

A customer says:

"Thanks, can you resend the meeting link?"

Routine.

A customer says:

"Your team has caused us a serious financial loss and I'm considering legal action."

Not routine.

Both arrive through email.

Both are technically messages.

But one should probably take a completely different path.

A useful workflow should recognise situations that require escalation.

7. AI shouldn't negotiate beyond defined limits

Negotiation involves more than generating persuasive language.

A customer might ask:

"If we sign today, can you reduce the price by 20%?"

AI could easily write:

"Absolutely, we can offer 20% off if you confirm today."

The words are easy.

The authority isn't.

A safer workflow might know:

Discount up to 5% for approved product category.

Customer requested:

20%

Therefore:

ESCALATE.

The AI doesn't need to become a negotiator simply because it can produce negotiating language.

8. AI shouldn't invent product or service capabilities

This is particularly dangerous because generative AI can make invented capabilities sound completely plausible.

Customer asks:

"Does your system integrate with Platform X?"

The information source contains nothing confirming that.

A bad AI says:

"Yes, we offer seamless integration with Platform X."

A useful AI says:

No approved information confirming Platform X integration.

Next action: Technical review.

That's a success.

"I DON'T KNOW" CAN BE THE CORRECT ANSWER.

9. AI shouldn't decide a customer isn't worth pursuing without appropriate rules

AI can help qualify opportunities.

But imagine it decides:

"This lead is unlikely to convert."

and silently stops follow-up.

Why?

Perhaps:

• the company is small,
• the email was short,
• the budget wasn't mentioned,
• or some scoring model assigned a low number.

That's a consequential decision.

There is an important difference between:

PRIORITISE

and:

ABANDON.

AI might help identify which opportunities need attention.

But automatically removing opportunities deserves much more scrutiny.

10. AI shouldn't keep contacting someone indefinitely

An AI follow-up system needs stop conditions.

Without them, "persistent sales agent" can become:

software that annoys people very efficiently.

Define when it stops.

For example:

• customer replies,
• customer asks to wait,
• meeting booked,
• opportunity closed,
• opt-out received,
• complaint raised,
• maximum contact limit reached,
• or uncertainty appears.

A useful sales agent knows when not to continue.

Stop conditions are part of the design

Every agentic workflow should answer:

WHEN DOES THIS STOP?

Suppose the job is:

"Follow up this opportunity appropriately."

The stop conditions might include:

• Customer responds · STOP automated sequence.
• Customer asks to wait · Reschedule.
• Customer opts out · STOP.
• Customer raises complaint · ESCALATE.
• Commercial negotiation begins · HUMAN.
• Maximum follow-ups reached · STOP / REVIEW.

The goal isn't endless autonomy.

It's controlled continuation.

11. AI shouldn't hide what it has done

If AI can act in your sales systems, you should be able to understand what happened.

For example:

That's far more useful than:

Record updated successfully.

For important actions, visibility matters.

12. AI shouldn't quietly expand its own role

Suppose you build an AI workflow to:

Prepare meeting briefs.

It needs read access to relevant CRM records.

Later somebody decides:

"While it's in there, it could update the CRM too."

Then:

"And send the follow-up."

Then:

"And create the quote."

Then:

"And adjust the opportunity."

The system has moved from:

READ

to:

CHANGE + SEND + COMMIT

without anyone properly redesigning its authority.

Every new capability should be treated as a new permission decision.

MORE CAPABILITY SHOULD NOT QUIETLY BECOME MORE AUTHORITY.

The Authority Ladder

A useful way to design this is:

The objective isn't to push every process towards Level 5.

The objective is to find the appropriate level for each action.

One AI sales agent can have several authority levels

Imagine an enquiry-handling agent.

It can:

• Read enquiry · Automatic.
• Check CRM · Automatic.
• Identify requirement · Automatic.
• Prepare response · Automatic.
• Send routine response · Within limits.
• Quote standard fixed price · Within approved rules.
• Change price · Human approval.
• Offer discount · Human approval.
• Answer unconfirmed technical question · Escalate.
• Handle complaint · Escalate.

That's much more realistic than saying:

"Our AI agent has autonomy."

Autonomy over what?

That's the question.

Consequence matters more than whether something is customer-facing

It can be tempting to say:

"Internal actions can be automated. Customer-facing actions need approval."

That's too simple.

An internal action such as:

MARK £100,000 OPPORTUNITY LOST

may have greater consequence than sending:

"Your meeting is confirmed for Tuesday."

So don't divide the world into:

• internal = safe,
• external = risky.

Look at the action itself.

Reversibility matters too

Compare:

Wrong? Delete it. Easy.

Wrong? You can apologise, but you can't unsend what the customer read.

Often reversible.

Potentially difficult.

Much harder.

The harder something is to reverse, the stronger the case for control before it happens.

A simple consequence test

Before giving AI authority over an action, ask:

IF THIS IS WRONG...

• Will a customer notice?
• Will it cost us money?
• Will it create a commitment?
• Will it alter important data?
• Will another system act on it?
• Will it affect a relationship?
• Can we undo it?
• Would I want to know before it happened?

The answers tell you a lot about the right authority level.

What about high-value sales?

Value alone isn't the only consideration.

But it changes the consequence.

An automated follow-up on a £50 standard product enquiry may be perfectly reasonable.

An automated message during a £500,000 negotiation may deserve very different controls.

Same technology.

Different context.

That's why authority belongs in the workflow design rather than being a universal AI setting.

Approval isn't the same as checking everything

There's another trap.

You decide AI needs human oversight.

So every action produces:

Approve?
Approve?
Approve?
Approve?

All day.

Now the human has become a button-clicking component in the automation.

That's not necessarily good control.

Good approval should happen where judgement matters.

And when approval is needed, show enough information to make the decision.

A useful approval looks like this

AI prepares the decision. You make it.

The human doesn't have to investigate from scratch.

AI prepares the decision.

The person makes it.

Human in the loop only works if the human adds something

This phrase gets used constantly:

Human in the loop.

But putting a human somewhere in the process isn't automatically useful.

Ask:

WHAT JUDGEMENT IS THE HUMAN PROVIDING?

If the person is approving a completely routine action 200 times a day, perhaps the boundary is wrong.

If they're reviewing:

• commercial exceptions,
• uncertain information,
• sensitive communication,
• or important commitments,

their involvement has a purpose.

Put people where human judgement actually changes the quality of the decision.

What happens when AI is uncertain?

Give it a failure path.

For example:

Customer asks whether Product A supports Platform X.

No approved compatibility information found.

That is good agentic behaviour.

The AI doesn't have to finish every job

This is a really important mindset change.

An AI agent's job isn't necessarily:

Complete everything without human help.

It might be:

Move this process forward until either the job is complete or human judgement is required.

That's much safer and often much more useful.

A successful outcome can be:

DONE.

Or:

NEEDS YOU.

Both move the work forward.

Start with the least authority needed

Suppose you're introducing AI into sales follow-up.

Don't start with:

"Send whatever follow-up you think is appropriate."

Start with:

Identify what needs follow-up.

Then:

Tell us what should happen.

Then:

Draft it.

Then:

Person sends.

Once the process is understood and trusted, selected routine cases might move to:

ACT WITHIN LIMITS.

Autonomy can be earned.

It doesn't need to be granted on day one.

A practical AI Sales Control Check

Before giving AI a new sales action, ask:

If you can't answer those questions, the system probably isn't ready for more authority.

Don't use AI permissions as a substitute for process design

Suppose nobody in the business agrees:

• when a lead is qualified,
• who can approve discounts,
• when an opportunity is lost,
• what can be promised,
• or when a complaint should escalate.

You don't have an AI permissions problem.

You have a business-process problem.

AI simply makes it visible.

Define the process first.

Then encode the boundaries.

What should a small business do?

Keep it simple.

You don't need a 94-page AI governance manual before AI prepares a meeting brief.

Start by listing what the workflow can:

SEE

and what it can:

DO.

Then mark actions:

• Automatic · Low consequence, clear rules.
• Approval · Needs human confirmation.
• Human only · Important judgement or commitment.
• Never needed · The AI has no reason to have this permission.

That's already a useful control model.

More autonomy isn't the goal

Imagine two systems.

System A has more autonomy.

That tells us nothing about which system creates more business value.

AUTONOMY IS A DESIGN CHOICE.

VALUE IS THE OBJECTIVE.

So, what should AI never be allowed to do in sales?

There isn't a universal list.

But AI should not be given unchecked authority simply because it is technically capable of an action.

Be particularly careful where AI could:

• invent information,
• make commercial commitments,
• change important data,
• delete information,
• handle sensitive situations,
• negotiate outside defined limits,
• continue unwanted communication,
• or act when it doesn't have enough information.

For every action, ask:

WHAT'S THE CONSEQUENCE IF IT'S WRONG?

CAN WE UNDO IT?

DOES THE AI HAVE ENOUGH INFORMATION?

DOES THIS DECISION ACTUALLY BELONG TO SOFTWARE?

Then give it the least authority needed to do the job properly.

Because the question that matters isn't:

HOW AUTONOMOUS CAN WE MAKE THIS?

It's:

HOW MUCH AUTHORITY DOES THIS PARTICULAR JOB ACTUALLY NEED?

That's how you put AI into sales without quietly handing it the keys to the business.

Quick answers

Related reading: what is an AI sales agent, AI sales agent vs sales automation, can AI update my CRM, can AI write sales proposals, can AI follow up sales leads, can AI qualify sales leads, how a small business can use AI for sales, and build an agentic sales workflow.