Do you actually need an AI sales agent?

Maybe.

But probably not for everything you're currently thinking about giving it.

AI sales agents can:

• understand enquiries,
• qualify leads,
• research prospects,
• prepare responses,
• follow up opportunities,
• update CRMs,
• prepare meetings,
• help create proposals,
• and coordinate work across systems.

That's useful.

It also doesn't mean:

YOU NEED AN AI SALES AGENT.

Sometimes you need better automation.

Sometimes you need AI to help with one part of the process.

Sometimes your existing software can already do the job.

And sometimes the sales process itself needs fixing.

So before choosing the technology, start somewhere much less exciting.

WHAT IS THE ACTUAL PROBLEM?

Start with the work

Suppose somebody says:

"We need an AI sales agent."

Ask:

WHY?

Perhaps the answer is:

"We're losing leads."

Good.

Why?

"People forget to follow them up."

Why?

"There's no consistent next action in the CRM."

Now we have a problem worth solving.

But we haven't established that it needs AI.

Perhaps:

• Proposal sent
• Automatically create follow-up task

would solve most of it.

That's automation.

Not an AI agent.

And that's completely fine.

The objective isn't to use AI

This sounds obvious.

But AI projects can easily reverse the normal order of technology decisions.

Instead of:

We have a problem. What should solve it?

we start with:

We have AI. Where can we put it?

Then perfectly good processes get rebuilt because someone wants an agent involved.

The objective should be:

MAKE THE SALES PROCESS BETTER.

AI earns its place if it helps do that.

Use this decision process

Before adding an AI sales agent, ask six questions.

• 1. DOES THIS WORK NEED TO EXIST?If not: remove it.
• 2. CAN WE MAKE IT SIMPLER?If yes: simplify it.
• 3. IS THE NEXT ACTION PREDICTABLE?If yes: automate it.
• 4. DOES SOMETHING NEED INTERPRETING?If yes: AI may help.
• 5. DOES THE WORK NEED TO CONTINUE ACROSS STEPS?If yes: agentic behaviour may help.
• 6. WHAT AUTHORITY DOES IT ACTUALLY NEED?Give it only what's required.

That's a much better route to an AI agent than starting with:

"Which agent platform should we buy?"

1. Does the work need to exist?

This is the question AI vendors aren't particularly incentivised to ask you.

Imagine your sales team manually creates a weekly spreadsheet containing:

• opportunities,
• stages,
• next actions,
• values,
• and owners.

All of that information already exists in the CRM.

The spreadsheet exists because:

"We've always done it."

You could build an AI agent that creates the spreadsheet automatically.

Or:

STOP CREATING THE SPREADSHEET.

The cheapest automation is deleting unnecessary work.

Remove before you automate

Look at the sales process and ask:

• Who uses this?
• What decision does it support?
• What happens if we stop doing it?
• Does this information already exist somewhere else?
• Are we doing this because the system is badly configured?

AI makes it easier to automate unnecessary work.

That doesn't make the work necessary.

2. Can you simplify it?

Perhaps the work genuinely needs doing.

But the current process has become complicated over time.

For example:

• Website enquiry arrives.
• Someone copies it into a spreadsheet.
• Someone checks the spreadsheet.
• Someone creates the CRM record.
• Someone emails the salesperson.
• Salesperson replies.
• Someone updates the spreadsheet.

Why?

Perhaps the simpler process is:

• Website
• CRM
• Owner
• Next action

Before asking AI to navigate complexity:

REMOVE SOME COMPLEXITY.

AI shouldn't become glue for a process nobody understands

If the workflow requires:

• 17 exceptions,
• three unofficial spreadsheets,
• two inboxes,
• a CRM nobody trusts,
• and Dave remembering which customers get different pricing,

you don't necessarily have an AI problem.

You have a process problem.

Fix enough of that first.

Then AI has something sensible to work with.

3. Is the next action predictable?

Now we're getting into technology.

Suppose:

When a website enquiry arrives, create a CRM record.

Predictable.

Automation.

When a proposal is sent, create a follow-up task five days later.

Predictable.

Automation.

When a meeting is booked, send the standard confirmation.

Predictable.

Automation.

You don't need generative AI to perform a rule reliably.

IF X → DO Y

is what ordinary automation is very good at.

Boring automation is underrated

There is a strange temptation at the moment to replace reliable automation with AI because AI feels more advanced.

But:

PREDICTABLE WORK SHOULD STAY PREDICTABLE.

If a rule can perform the task:

• consistently,
• cheaply,
• quickly,
• and transparently,

that's an advantage.

Not a technological embarrassment.

Don't hire an AI agent to press a button

Imagine the requirement:

"Every new website lead should be assigned to Sarah."

You don't need:

• an agent,
• reasoning,
• memory,
• tool use,
• multi-step planning,
• or a tiny animated AI employee called Steve.

You need:

A RULE.

Use the rule.

4. Does something need interpreting?

Now AI becomes more interesting.

Suppose the website enquiry says:

"We're replacing the website for our two property offices in London and Bristol. We're currently on WordPress and need to connect to our property management system. Ideally we'd like everything live before January."

Now somebody needs to understand:

• What does the customer want?
• Which service is relevant?
• What information have they provided?
• What's missing?
• Does the requirement fit?
• What should happen next?

That's no longer simply:

IF X → DO Y.

The information needs interpreting.

AI may be useful.

This is where AI assistance can be enough

But even now:

YOU STILL MAY NOT NEED AN AGENT.

Perhaps AI:

• reads the enquiry,
• extracts the important information,
• prepares a summary,
• and drafts the response.

Then a salesperson:

• checks it,
• sends it,
• updates the CRM,
• and manages what happens next.

That's AI assistance.

It might solve the problem perfectly well.

An AI assistant helps with the work

Think:

• YOU COORDINATE
• Ask AI.
• Receive output.
• Review.
• Move information.
• Take action.
• Continue.

AI is helping.

The person still owns the workflow.

For many sales tasks, that's enough.

So when does an agent become useful?

When the job doesn't end after producing an answer.

Suppose the requirement is:

Make sure every genuine website enquiry receives the appropriate next action.

Now the system may need to:

• notice the enquiry,
• understand it,
• check existing customer information,
• retrieve relevant business information,
• identify missing details,
• prepare a response,
• update the CRM,
• create the next action,
• monitor what happens,
• interpret the customer's reply,
• and continue until the process reaches a person or an appropriate endpoint.

That's different.

THE AI ISN'T JUST HELPING WITH A TASK.

IT'S HELPING MOVE THE WORK FORWARD.

That's where agentic behaviour becomes interesting.

Task versus responsibility

This is probably the easiest way to understand the distinction.

The second requires the system to keep track of:

• state,
• context,
• actions,
• results,
• and what happens next.

Ask who coordinates the work

This is another useful test.

AI ASSISTANT

YOU coordinate the work.

AI AGENT

THE SYSTEM helps coordinate the work.

The difference isn't simply that the agent writes better emails.

It's that it can take responsibility for defined movement through a process.

5. Does the work need to continue across steps?

This is where many good agent use cases appear.

Consider follow-up.

A normal automation:

• Proposal sent
• Wait five days.
• Send email.

Done.

An AI-assisted workflow:

• Proposal sent
• Five days later AI prepares contextual email.
• Human sends it.

Done.

An agentic workflow:

• Proposal sent
• Record next action.
• Wait.
• Check whether customer replied.
• Understand what happened.
• Decide whether follow-up is appropriate.
• Prepare or send within authority.
• Record result.
• Create next action.
• Continue or escalate.

That's a continuing job.

Agents are useful around loops

Look for work that contains:

• OBSERVEWhat happened?
• UNDERSTANDWhat does it mean?
• DECIDEWhat should happen?
• ACTDo the permitted thing.
• CHECKDid it work?
• CONTINUEWhat happens next?

Those loops are much more naturally agentic than:

"Write me an email."

But multi-step doesn't automatically mean AI agent

This distinction matters.

A traditional automation can also contain:

• branches,
• multiple systems,
• loops,
• conditions,
• and several actions.

A workflow being complicated doesn't automatically make it agentic.

Ask:

DOES THE SYSTEM NEED TO INTERPRET VARIABLE INFORMATION AND CHOOSE BETWEEN POSSIBLE NEXT ACTIONS?

If not, ordinary automation may still be the better component.

A practical comparison

The problem determines the technology.

Not the other way around.

6. How much authority does it need?

Even if you've identified a genuinely agentic job, there's another question:

WHAT SHOULD THE AGENT ACTUALLY BE ALLOWED TO DO?

An agent doesn't have to be autonomous.

It can:

• READUnderstand what is happening.
• RECOMMENDSuggest the next action.
• PREPAREPrepare the work.
• ACT WITH APPROVALA person confirms.
• ACT WITHIN LIMITSRoutine actions happen independently.

And whenever necessary:

ESCALATE.

An agent that needs approval can still be useful

This is important.

Imagine an enquiry agent that:

• reads every enquiry,
• checks CRM,
• identifies the requirement,
• retrieves relevant information,
• prepares the response,
• suggests the next action,
• and presents it to a person.

The person clicks:

APPROVE.

The system then:

• sends,
• updates CRM,
• creates the next action,
• and continues monitoring.

That's still considerably different from manually coordinating the whole workflow.

Human approval doesn't make the agent pointless.

More autonomy isn't automatically better

Suppose:

Version A saves 80% of the admin while a person approves customer-facing actions.

Version B saves 85% but can independently:

• send pricing,
• make commitments,
• change opportunity stages,
• and handle unusual requests.

Version B isn't automatically the better system.

The extra 5% needs to justify the extra risk and complexity.

AUTONOMY SHOULD EARN ITS PLACE TOO.

Use consequence to decide authority

Ask:

What happens if this is wrong?

If the answer is:

"We correct a meeting note."

low consequence.

If the answer is:

"We quoted the customer £12,000 instead of £21,000."

rather different.

Authority should reflect:

• consequence,
• reversibility,
• customer impact,
• financial impact,
• and uncertainty.

Four possible answers

Once you've worked through the questions, you will usually land somewhere around one of these.

Sometimes one sales process needs all four

This is probably the most realistic answer.

Take inbound enquiries.

• FORM SUBMITTED AUTOMATION
• UNDERSTAND REQUIREMENT AI
• CREATE / FIND CRM RECORD AUTOMATION
• ASSESS QUALIFICATION AI + RULES
• UNUSUAL COMMERCIAL REQUEST HUMAN
• PREPARE RESPONSE AI
• SEND HUMAN APPROVAL / AGENT WITHIN LIMITS
• RECORD AUTOMATION
• MONITOR NEXT ACTION AGENT
• SALES CONVERSATION HUMAN

• FIX THE PROCESSDefine what counts as an enquiry.
• AUTOMATIONCreate CRM record.
• AIInterpret free-text requirements.
• AUTOMATIONRetrieve known customer record.
• AIAssess against defined qualification criteria.
• HUMANReview unusual commercial request.
• AIPrepare response.
• HUMAN OR AGENTApprove/send depending on authority.
• AUTOMATIONRecord activity.
• AGENTMonitor next action.
• HUMANHandle the sales conversation.

The objective isn't to choose:

HUMAN

or:

AUTOMATION

or:

AI.

It's to design the workflow using the right component at each point.

Don't turn your business into an AI org chart

Another sign that things have gone too far:

Meet Alex, our AI SDR.

Meet Sophie, our AI Account Manager.

Meet Oliver, our AI Researcher.

Meet Chloe, our AI Proposal Specialist.

You now apparently employ four fictional people.

But what work are they doing?

Start with:

THE JOB.

Not the imaginary employee.

Perhaps one workflow can:

• research,
• qualify,
• prepare,
• update,
• and follow up.

It doesn't need five names.

Agents don't need names

They don't need:

• profile pictures,
• job titles,
• personalities,
• birthdays,
• or a spot on the organisation chart.

They need:

• A CLEAR JOB.
• THE RIGHT INFORMATION.
• APPROPRIATE TOOLS.
• DEFINED AUTHORITY.
• STOP CONDITIONS.
• A HUMAN ESCALATION PATH.

Much less fun for the stock-photo industry.

Much more useful for the business.

Don't buy an AI sales agent because competitors have one

This is another bad reason.

Perhaps a competitor announces:

"We've deployed 15 AI agents across sales."

You have no idea:

• what those agents do,
• whether they work,
• whether anybody uses them,
• whether they created value,
• or whether somebody counted 15 workflows because it sounded more impressive than three.

Technology strategy shouldn't be based on:

AI AGENT ENVY.

Find your own work.

Don't buy one because the demo looked clever

The demo:

• New enquiry.
• AI researches customer.
• AI responds.
• AI books meeting.
• AI updates CRM.
• Everything turns green.

Very satisfying.

Now ask:

• What happens if the CRM contains two contacts?
• What happens if pricing is missing?
• What happens if the customer is angry?
• What happens if the AI isn't sure?
• What happens if the API fails?
• What happens if the customer asks for something outside policy?
• What happens if a person needs to take over?

If those answers aren't clear, you haven't seen the important part of the demo yet.

The best first agent is usually boring

Not:

Autonomously run our global sales operation.

Try:

Make sure every website enquiry is reviewed and receives a next action.

Or:

Make sure every proposal gets appropriate follow-up.

Or:

Prepare one useful brief before every sales meeting.

Or:

Keep routine CRM information current after sales conversations.

These jobs are:

• contained,
• frequent,
• understandable,
• checkable,
• and measurable.

That's useful.

Use the first project to learn

Your first agent doesn't have to transform the company.

It should help you learn:

• which information matters,
• where the process breaks,
• what AI does well,
• where it struggles,
• which actions are safe,
• what people actually trust,
• and where the value comes from.

Then expand based on evidence.

Not excitement.

What if your business is small?

You may actually have some particularly good opportunities.

Not because you need:

an AI sales department.

But because one person may currently be:

• salesperson,
• researcher,
• sales admin,
• CRM updater,
• proposal writer,
• follow-up system,
• and occasional printer technician.

AI doesn't magically turn that person into a 50-person sales organisation.

But it may remove some of the surrounding work that doesn't require them.

That's meaningful.

You may need capability, not headcount replacement

This is another way to think about it.

Perhaps nobody currently:

• researches every prospect properly,
• prepares every meeting,
• checks every opportunity,
• or follows up every quote consistently.

Not because the business doesn't value those things.

There simply isn't enough time.

AI can sometimes create:

CAPABILITY THE BUSINESS DIDN'T PREVIOUSLY HAVE.

That's different from:

"Replace somebody."

And often more interesting.

What if you already have lots of AI tools?

Then I'd be even more cautious about buying another one.

Audit them.

For each tool:

• What job does it do?
• Who uses it?
• What business problem does it solve?
• What information does it access?
• What actions can it take?
• Does another tool already do the same thing?
• Would we notice if we cancelled it?

If the answer to the last question is:

"Probably not."

you've learned something.

The AI sales agent test

Before buying or building one, answer these:

1. What exact job are we giving it?If the answer is "sales", stop. Too broad.
2. What starts the job?An enquiry? Meeting? Proposal? CRM change?
3. What information does it need?Not everything. The information required for this job.
4. What needs interpreting?If nothing, consider automation.
5. What tools does it need?CRM? Email? Calendar? Documents?
6. What actions should it take?Be specific.
7. What needs approval?Define it.
8. When should it stop?Uncertainty? Missing information? Commercial exception?
9. Who takes over?Name the role.
10. How will we know it worked?Define the business measure.

If you can't answer those yet, don't shop for agents.

Map the work first.

A decision tree

Start here:

• DOES THE WORK NEED TO EXIST?NO → Remove it.YES
• CAN WE SIMPLIFY IT?YES → Simplify it.
• IS THE NEXT ACTION PREDICTABLE?YES → Automate it.NO
• DOES SOMETHING NEED INTERPRETING?YES → Use AI.
• DOES A PERSON STILL COORDINATE THE WORK EFFECTIVELY?YES → AI assistance may be enough.NO
• DOES THE WORK NEED TO CONTINUE ACROSS STEPS, SYSTEMS OR TIME?YES → Consider an AI agent.
• WHAT AUTHORITY DOES IT NEED?Start with the least required.
• WHAT HAPPENS WHEN IT CAN'T CONTINUE?Stop → Explain → Escalate.

That's it.

No robot required.

So, do you actually need an AI sales agent?

Maybe.

If your problem is:

an unnecessary process,

fix the process.

If the work is:

predictable,

use automation.

If the work needs:

interpretation,

use AI.

If AI can help but a person can comfortably coordinate what happens next:

use an AI assistant.

And if the work requires a system to:

• understand,
• retrieve,
• decide,
• act,
• check,
• remember state,
• and continue across multiple steps until a goal is reached or a person is needed:

AN AI AGENT MAY BE THE RIGHT COMPONENT.

Notice the wording.

COMPONENT.

Not strategy.

Not employee.

Not magic salesperson.

A component inside a properly designed sales process.

That's what agentic selling is really about.

Not putting agents everywhere.

It's understanding:

• WHERE AI SHOULD HELP.
• WHERE IT SHOULD ACT.
• WHERE AUTOMATION IS ENOUGH.
• AND WHERE A PERSON SHOULD TAKE OVER.

The goal isn't to build an AI sales agent.

The goal is to build a sales process that works better.

If an agent helps you do that:

use one.

If it doesn't:

don't.

Quick answers

Where to go next

If you've reached the conclusion:

YES, THERE'S A REAL JOB HERE FOR AN AGENT

don't start with the software.

Start with the workflow.

If you're still working out what an agent actually is:

And if the problem is really missed leads: