Before you build an AI sales agent, fix these 5 things

Before you automate

You want an AI sales agent.

Great.

Before you choose the model, buy the software, connect the CRM or give anything permission to email customers, there are five things worth checking.

1. 1. PROCESS Does everybody know how the work should happen?
2. 2. INFORMATION Does the system have reliable information to work from?
3. 3. SYSTEMS Can the tools involved actually work together?
4. 4. AUTHORITY Have you decided what AI may and may not do?
5. 5. MEASUREMENT Will you know whether any of this made the sales process better?

If those five things aren't reasonably clear, building the agent probably isn't your first job.

Because AI can automate work.

It can also automate confusion.

1. Fix the process

Start here.

Before asking:

How can AI do this?

ask:

How does this actually work now?

Not how the process diagram says it works.

Not how it worked when somebody designed the CRM three years ago.

What actually happens?

Take website enquiries

You might think the process is:

• Enquiry
• Qualify
• Respond
• Follow up

Simple.

Then you watch what really happens.

A website enquiry arrives.

Someone sees it in the inbox.

Unless they're busy.

They read it.

Sometimes they ask Dave whether it's a good lead.

Sometimes it goes into the CRM.

Sometimes it doesn't.

If the enquiry looks interesting, somebody replies.

If information is missing, they ask for it.

Unless they forget.

Someone may create a follow-up task.

Or remember it.

Eventually.

That's the real process.

And that's the process AI will encounter.

Map what actually happens

Use this:

• TriggerWhat starts the work?
• InformationWhat does someone need?
• DecisionWhat needs deciding?
• ActionWhat happens?
• CheckHow do we know it worked?
• Next actionWhat happens afterwards?

For an enquiry:

• Trigger Website form submitted.
• Information Customer details, request, services, locations, previous relationship.
• Decision Can we help? What do we need to know? Who should handle it?
• Action Respond, route or request information.
• Check Did the customer receive the right response?
• Next action Meeting, follow-up, quote, human review or close.

Now we have something we can design around.

Don't automate the workaround

This is a common trap.

Imagine the sales team:

• copies CRM information into a spreadsheet,
• because the CRM report isn't useful,
• then emails the spreadsheet every Friday.

You could build AI to:

• read the CRM,
• create the spreadsheet,
• summarise it,
• and email it.

Wonderful.

You've automated a workaround.

The better question might be:

Why does the spreadsheet exist?

Perhaps the CRM view needs fixing.

Before automating a step, ask:

Should this step exist at all?

Use this order

Before adding AI:

This one sequence can save a remarkable amount of unnecessary AI.

Example: quote follow-up

Problem:

"We forget to follow up quotes."

You could build an agent that monitors every quote, reads all customer communication, determines follow-up strategy and sends messages autonomously.

Or you could discover:

nobody records a follow-up date.

Fix that first.

Perhaps:

• Quote sent
• Create next action

solves 70% of the problem.

Then AI can help with the part that actually needs AI:

What should the follow-up say given what has happened since?

That's much cleaner.

Your process doesn't need to be perfect

Don't misunderstand this.

You don't need six months of process consultancy before trying AI.

Real businesses are messy.

The process just needs to be clear enough that you can answer:

• what starts the work,
• what information matters,
• what decisions happen,
• what actions are expected,
• what exceptions exist,
• and what a successful outcome looks like.

If you can't explain the job to a person, you'll struggle to give it to an agent.

2. Fix the information

Once the process is clear, ask:

What does the AI need to know to do the job?

This is where a lot of AI projects become data projects without anyone noticing.

Imagine an enquiry agent

Its job is:

Understand new enquiries and prepare the appropriate response.

Sounds straightforward.

Until it needs to know:

• What services do you offer?
• Where do you operate?
• What don't you offer?
• What information is needed for a quote?
• What are the current prices?
• Which technical claims are approved?
• What are normal delivery times?
• Which customer already exists?
• What did they ask previously?
• What exceptions need a person?
• Where does that information live?

"It's on the website" isn't always the answer

Perhaps some information is on the website.

Some is in:

• PDFs,
• CRM notes,
• emails,
• spreadsheets,
• product systems,
• old proposals,
• shared drives,
• and people's heads.

And sometimes those sources disagree.

The website says:

Delivery: 5-7 working days.

The pricing document says:

7-10 working days.

Dave says:

"It's usually about two weeks at the moment."

Which one should AI tell the customer?

That's not primarily an AI problem.

That's an information problem.

AI needs an information hierarchy

Decide which sources are authoritative.

For example:

• Pricing Approved pricing system.
• Delivery Current operations data.
• Product specification Maintained technical source.
• Customer history CRM plus relevant communication.
• Terms Current approved terms.

Then when sources conflict, the workflow knows:

which source wins,

or:

Escalate.

Current matters

AI can have access to perfectly accurate information that is six months out of date.

That's still a problem.

Ask:

• Who owns this information?
• Who updates it?
• How often?
• What happens when it changes?
• Can old versions still be retrieved accidentally?
• Can the system tell which version is current?

Agentic workflows make information maintenance more important, not less.

Because now the information can drive actions.

Give it what it needs, not everything you have

There can be a temptation to connect:

• the whole CRM,
• all email,
• every document,
• all customer data,
• every shared folder,

because:

"More context will make the AI better."

Not necessarily.

More irrelevant information can make the job harder.

It can also create unnecessary access.

Start with:

Work → Context.

What does this specific job need to know?

Give it that.

Information isn't authority

Another important distinction.

An agent may need to see pricing to prepare a quote.

That does not mean it may:

change the price.

It may need customer email to understand context.

That does not mean it may:

send email.

It may need CRM data to prepare a meeting brief.

That does not mean it may:

edit the CRM.

We'll come back to that under authority.

3. Fix the systems

Now look at where the work actually happens.

Your sales process may involve:

• website,
• CRM,
• email,
• calendar,
• forms,
• proposal software,
• documents,
• pricing,
• project management,
• accounting,
• and other systems.

An agentic workflow often has to move between them.

The human may currently be the integration

For example:

For example:

• Website enquiry
• Dave reads it.
• Dave checks CRM.
• Dave searches email.
• Dave finds product information.
• Dave writes response.
• Dave updates CRM.
• Dave creates calendar task.
• Dave remembers follow-up.

The systems aren't connected.

Dave is the API.

That's often where the opportunity sits.

Not necessarily in replacing any of the individual tools.

But in reducing the manual coordination between them.

Don't replace software that already works

If your CRM reliably:

• stores customer information,
• creates tasks,
• runs reports,
• and tracks opportunities,

keep it.

If your calendar reliably books meetings:

keep it.

If ordinary automation reliably:

creates a CRM record after a form submission,

keep it.

AI doesn't need to replace reliable software.

Use the right component for the job.

Predictable work should stay predictable

For example:

• Form submitted → Create CRM contact. Automation.
• Meeting booked → Send confirmation. Automation.
• Customer email arrives → Understand whether it changes the opportunity. AI may help.
• Customer asks for 20% discount → Human.

This is the architecture:

• Predictable → Automation
• Interpretive → AI
• Consequential → Human

Don't turn every box green because you're building an AI project.

Check whether the systems can actually connect

Before promising an agent can:

• read CRM,
• send email,
• book meetings,
• prepare proposals,
• or update records,

check whether the systems involved support what you need.

Questions include:

• Is there an API?
• Are webhooks available?
• What permissions exist?
• Can actions be limited?
• Can changes be traced?
• What happens if the integration fails?
• Are there rate or usage limits?
• Can the workflow recover?

A beautiful architecture diagram doesn't make unsupported integrations appear.

Think about failure between systems

Suppose:

AI prepares the response correctly.

Then:

the CRM API fails.

What happens?

• Does the email still send?
• Does the next action get created?
• Does the salesperson know the CRM wasn't updated?
• Or does the workflow quietly continue as though everything succeeded?

This is why:

Check

belongs in the process.

Not:

AI → Action → Done.

But:

Action → Check → Continue or escalate.

4. Fix the authority

Now we get to the part that becomes particularly important when AI can act.

You need to decide:

What can it see?

and separately:

What can it do?

"Give the AI access to the CRM" isn't a permission model

Inside a CRM, AI might potentially:

• read a contact,
• create a contact,
• add a note,
• create a task,
• change a field,
• change opportunity stage,
• change deal value,
• mark an opportunity won,
• mark it lost,
• or delete it.

Those aren't equivalent actions.

So don't grant:

CRM access.

Design:

CRM authority.

Use the Authority Ladder

Start here:

And alongside all of those:

Not every workflow needs to reach Level 5.

Start with the least authority needed

Suppose the job is:

Help us follow up quotes properly.

You don't need to begin with:

AI may send whatever it thinks appropriate to every customer.

Start:

Watch

Which quotes need attention?

Then:

Recommend

What should happen?

Then:

Prepare

Draft the follow-up.

Then perhaps:

Act with approval.

Only after you understand the workflow might selected routine cases move to:

Act within limits.

Autonomy can be earned.

Define limits explicitly

For example:

This is much clearer than:

"Human in the loop."

Decide what happens when AI isn't sure

Every agentic workflow needs a failure path.

• StopDon't take the uncertain action.
• ExplainShow what is unclear.
• EscalateBring in the right person.

For example:

That's a successful outcome.

Authority should depend on consequence

Ask:

• What happens if this is wrong?
• Can we undo it?
• Will a customer notice?
• Does it cost money?
• Does it create a commitment?
• Does another system act on it?
• Would I want to know first?

The answers help determine the right authority level.

5. Fix the measurement

Finally:

How will you know this worked?

This gets skipped surprisingly often.

The project launches.

Everyone watches the agent doing things.

There are dashboards.

Lots of little green ticks.

Someone says:

"It processed 847 actions this month."

Great.

Did the sales process improve?

Activity isn't value

Don't measure:

• AI emails sent.
• AI tasks completed.
• AI summaries created.
• AI actions performed.

Those may be useful operational metrics.

But they don't tell you whether the business got better.

Measure the problem you started with.

If the problem was slow enquiry response

Measure:

time to useful response,

not:

number of AI responses.

If the problem was missed follow-up

Measure:

• opportunities without next actions,
• overdue follow-up,
• and appropriate follow-up completion.

Not:

number of emails AI sent.

If the problem was sales admin

Measure:

• human admin time,
• CRM completeness,
• manual copying,
• and corrections.

Not:

number of records AI touched.

If the problem was meeting preparation

Measure:

• preparation time,
• missing context,
• forgotten commitments,
• and usefulness of meeting briefs.

Not:

number of briefs generated.

Measure before you automate

This is important.

If you don't know how the process performs now, it becomes difficult to know whether AI improved it.

Before changing the workflow, capture a baseline.

For example:

• Current Average useful enquiry response time.
• Current Percentage of quotes with a next action.
• Current Time spent preparing proposals.
• Current CRM records missing next action.
• Current Human time spent updating records.

Then compare after implementation.

Otherwise:

"It feels faster."

may be all you have.

Include the cost of checking AI

Suppose AI saves:

10 minutes.

But the person spends:

8 minutes checking and correcting it.

Net improvement:

2 minutes.

That's still an improvement.

But it's not 10.

Human review belongs in the measurement.

So do:

• corrections,
• escalations,
• errors,
• and rework.

Measure quality as well as speed

A workflow that replies to every enquiry in 11 seconds may look excellent on a dashboard.

Unless the replies are rubbish.

Measure:

Speed

and:

Quality.

Likewise:

more booked meetings

isn't automatically better if they are poor-fit meetings.

More follow-up isn't automatically better if customers are annoyed.

More proposals aren't automatically better if scope is wrong.

Don't optimise the easy number and accidentally damage the useful one.

Use the value ladder

Think:

• Task valueDid one task improve?
• Workflow valueDid the overall sales process improve?
• Business valueDid capacity, cost, customer experience or sales improve?
• New capabilityCan the business now do something it couldn't realistically do before?

This helps stop AI ROI becoming:

"The model produced 2,000 words in 14 seconds."

And failure runs through all five

Failure handling isn't a separate sixth box.

It belongs everywhere.

• Process failure What happens when the situation doesn't match the normal process?
• Information failure What happens when information is missing or contradictory?
• System failure What happens when an integration breaks?
• Authority failure What happens when the required action is outside AI's permissions?
• Measurement failure How will you know the system is producing bad outcomes?

Every stage needs an answer.

A good agent doesn't just know what to do

It also needs to know:

When it can't continue.

That's one of the biggest differences between a demo and a business system.

The demo assumes:

everything works.

The business system assumes:

eventually, something won't.

Example: the enquiry agent

Let's put all five together.

The job:

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

• Process Enquiry → Understand → Qualify → Respond → Record → Next action.
• Information Services, locations, qualification criteria, customer history, approved product information.
• Systems Website → AI → CRM → Email → Calendar.
• Authority Read automatically. Prepare automatically. Routine responses with approval initially. Commercial exceptions escalate.
• Measurement Useful response time. Enquiries with next action. Manual handling time. Corrections. Missed enquiries.

Now:

We have something worth building.

Compare that with the vague version

"We want an AI sales agent that handles our leads."

Okay.

• Which leads?
• From where?
• What does "handle" mean?
• What information can it use?
• How does it know whether a lead is good?
• What can it send?
• Can it quote?
• Can it update CRM?
• What happens when it doesn't know?
• Who takes over?
• How will you know it worked?

Until those questions have answers, you're not really specifying an agent.

You're describing an ambition.

Don't buy the software to force the conversation

Sometimes businesses buy a tool hoping implementation will make them work this out.

Occasionally it does.

More often:

• the unclear process gets configured badly,
• the data gets connected inconsistently,
• permissions are decided during setup,
• and measurement arrives after launch.

Reverse it.

Define the work.

Then choose the technology.

You don't need to fix the entire company first

This is important too.

Suppose your CRM is messy.

That doesn't mean:

No AI until every CRM record is perfect.

Pick one workflow.

Identify the information that that workflow needs.

Fix enough around that job to make it reliable.

For example:

"For new website enquiries, we need accurate service information, location rules, qualification criteria and CRM contact matching."

Fix those.

Build that workflow.

Learn.

Then expand.

Start narrow enough to understand what happens

A good first project is usually:

• Frequent Happens often enough to matter.
• Understood You know what good looks like.
• Information-ready The necessary context exists.
• Easy to check A person can judge the output.
• Recoverable Mistakes can be corrected.
• Measurable You can tell whether it improved.

That's a much stronger first agent than:

Autonomously run our entire sales process.

Your AI Sales Readiness Check

Before building, ask:

If you can answer those questions, you're in a much better position to build something useful.

So, what should you fix before building an AI sales agent?

Five things:

• Process Make the work clear.
• Information Make the context reliable.
• Systems Make the components work together.
• Authority Decide what AI can see and do.
• Measurement Know what success actually means.

And across all five:

Design for failure.

• Don't assume every customer will behave normally.
• Don't assume every piece of information will exist.
• Don't assume every API will work.
• Don't assume AI will always be certain.
• Don't assume activity equals value.

Because the objective isn't:

Build an AI sales agent.

It's:

Build a sales process that works better because AI is in it.

Sometimes fixing these five things will show you exactly where an agent belongs.

And sometimes you'll discover that the biggest improvement had nothing to do with AI at all.

That's useful too.

Quick answers