Build an AI sales agent

Build an AI Sales Agent for Your Business

Give AI a defined job inside your sales process.

You have a sales process. Somewhere inside it, there is probably work that is repeated, delayed, copied, checked, chased or forgotten. An AI sales agent can take responsibility for a clearly defined part of that work.

It might handle incoming enquiries. Prepare salespeople for meetings. Keep CRM information moving. Watch for opportunities that need attention. Prepare follow-up. Gather information from several systems. Or carry out a completely different job specific to your business.

We design and build AI sales agents for UK businesses, with clear responsibilities, controlled access, human approval where it matters and escalation when the agent reaches its limits.

A fair question

Before we build an agent, we ask one question.

Do you actually need one?

You may arrive here thinking: "We need an AI sales agent." You might. But perhaps the problem can be solved with: A CRM feature you already have. A conventional automation. A better integration. A simpler workflow. AI assisting a salesperson. Or a change to the underlying process.

If that solves the problem properly, building an agent would add unnecessary complexity. So we do not start by asking: "What agent should we build?" We start with: "What work are you trying to improve?"

Definition

What is an AI sales agent?

AI responsible for a defined piece of sales work.

An AI sales agent can understand information, work towards a defined outcome and use permitted tools or systems to help move work forward. For example: A new enquiry arrives. The agent reads it. Identifies what the person needs. Checks permitted CRM context. Applies your agreed qualification or routing logic. Determines the appropriate next step. Prepares the action. Carries it out if authorised. Or asks a person to approve it. If something falls outside its rules, it escalates.

That is very different from simply asking AI to: "Write a sales email."

The job description

Start with the job description.

"Help us sell more" is not a job.

Before building an agent, we define what it is actually responsible for. A useful agent job might be: "Make sure every genuine website enquiry reaches the appropriate salesperson with the information they need." Or: "Prepare an account brief before every scheduled sales meeting." Or: "Find active opportunities without a clear next action and surface them to the owner." Or: "Turn agreed meeting information into prepared CRM updates and follow-up actions."

Now we have something that can be designed. We know what triggers the job. What information is needed. What outcome is expected. Where the boundaries are. And when the agent should stop.

An agent should have a job description too.

Agent patterns

What AI sales agents can we build?

The right agent depends on your process.

New enquiry agent

Understands incoming enquiries, gathers context and prepares the appropriate next step.

Lead qualification agent

Applies criteria defined by your business and identifies missing information or situations needing review.

Lead routing agent

Determines where an enquiry should go and provides useful context with the handover.

Meeting preparation agent

Gathers permitted information and prepares a concise briefing before a sales conversation.

CRM agent

Prepares or performs agreed CRM updates using information generated during the sales process.

Follow-up agent

Watches for agreed follow-up situations, gathers context and prepares or performs the appropriate next action.

Sales attention agent

Identifies opportunities, enquiries or actions that need somebody's attention.

Handover agent

Prepares useful context when an opportunity moves between people or teams.

Sales knowledge agent

Helps retrieve approved information your sales team needs during the sales process.

These are patterns. Your agent does not need to fit neatly into one of them. It needs to fit your process. See AI sales agent examples for how these play out in practice.

The wider workflow

We build the workflow around the agent too.

An agent rarely works alone.

A useful AI agent may need to interact with:

Your CRMEmailWebsite formsCalendarsMeeting informationDocumentsInternal systemsDatabasesAutomation platformsExisting business softwarePeople

The agent is one participant inside a wider workflow. For example:

Website form
Automation
AI Agent
CRM
Human Approval
Email
Follow-Up Workflow

Some steps need AI. Some do not. That is why we design the complete workflow rather than trying to make every step agentic.

Context

What does the agent need to know?

Context matters.

Imagine asking a salesperson to handle an enquiry while giving them only the customer's first sentence. They would probably need more information. Agents are similar. Depending on the job, useful context might include:

The original enquiryCustomer or company informationPrevious conversationsCRM historyApproved service informationCurrent opportunity detailsRelevant documentsBusiness rulesQualification criteriaPrevious actions

But more information is not automatically better. The agent should receive the context necessary for the job. Not everything your business happens to know.

Systems and access

What systems does the agent need to access?

Connect what is necessary.

Once the job is clear, we look at the systems involved. Perhaps the agent needs to: Read website enquiries. Search CRM records. Prepare an email. Create a task. Retrieve approved documents. Check calendar information. Update an opportunity.

Each capability creates a permission decision. Does it need read access? Write access? Permission to send? Permission to create? Permission to change? Or should it only prepare the action for somebody else? We define those boundaries before handing over unnecessary access.

Capability vs authority

Decide what the agent can do.

Capability and authority are different things.

An AI agent may technically be capable of sending an email. That does not mean it should be allowed to send every email. It might technically be capable of updating your CRM. That does not mean it should be allowed to change every field.

We separate: What the agent can do from: What the agent is allowed to do. That distinction is fundamental to how we build, and it sits at the heart of good agent governance.

The authority ladder

Choose the level of authority.

Not everything needs to become autonomous.

01ReadThe agent can understand agreed information.
02RecommendIt can suggest the next action.
03PrepareIt can prepare the work for somebody.
04Act with approvalIt can carry out an action once a person approves it.
05Act within limitsIt can independently carry out specific agreed actions under defined conditions.
06EscalateIt stops and passes the situation to a person when it reaches its boundaries.

Different actions inside the same agent can have different levels. Autonomy should earn its place.

Escalation

Build the escalation path.

The agent needs somewhere to go when it cannot continue.

An agent may encounter:

Missing informationConflicting recordsAn unusual requestA sensitive conversationAn important commercial decisionA system failureA situation outside the workflowSomething it cannot interpret reliably

The wrong design is: "Try your best." The better design may be: Stop. Gather the relevant context. Explain what is unclear. Pass it to the right person.

Escalation is part of the workflow, not a failure of it.

The build

Then we build.

Around your actual systems and process.

Once the workflow is understood, we can design the technical implementation. That may involve:

AI modelsAPIsAutomationCRM integrationsWebhooksDatabasesStructured business rulesRetrieval from approved information sourcesHuman approval interfacesNotificationsLoggingExisting softwareCustom components

The exact stack depends on the job. We do not need to rebuild software that already solves the commodity parts well. Buy the commodity. Build the difference.

Testing

Then we try to break it.

The perfect example is not the test.

A demonstration often looks wonderful because everything behaves exactly as expected. Real businesses are messier. What if:

We test the awkward cases too. Because those are the situations that tell us where the workflow's boundaries really are.

Controlled authority

Start with controlled authority.

You do not need to trust the agent blindly on day one.

An agent can begin by:

Begin by
Reading
The agent observes real situations before it changes anything.
Begin by
Recommending
It suggests what should happen next, and a person decides.
Begin by
Preparing
It prepares the next action, ready for review.
Begin by
Showing its work
It shows its work to a person, so you can see how it behaves.

That gives you an opportunity to see how it behaves with real situations. Where does the team agree with it? Where do they change its recommendation? Which situations cause problems? Which actions appear sufficiently predictable?

Specific authority can then increase where there is a reason. Or decrease where more oversight is needed. Authority should not only move in one direction.

Measurement

Measure whether it improved the process.

Not how busy the agent became.

An agent completing 5,000 actions does not automatically mean it created value. We care about what happened to the underlying work. Depending on the workflow, useful questions might include:

Activity is not value.

Cost

What does an AI sales agent cost?

It depends on the job.

We do not pretend every AI agent is the same project. A focused internal agent working with one source of information is very different from a workflow connecting several systems, making decisions, taking actions and handling human approvals. The scope depends on things such as:

The process being improvedThe systems involvedThe integrations availableThe information requiredThe number and complexity of actionsThe level of authorityApproval requirementsException handlingTesting requirementsOngoing operation

That is why we understand the workflow before defining the build.

Tell us what you want to improve

Existing software

Do you need to replace your existing software?

Usually, no.

Your business may already have perfectly good:

CRM softwareEmailCalendarFormsAutomationMeeting softwareDocument systemsSales tools

We are interested in how the work moves between them. Often, the useful part is connecting existing systems with the logic specific to your business. How you handle an enquiry. What information matters. What counts as qualified. When somebody should be involved. What should happen next. Where the process needs to stop. That is the difference worth designing.

Start small

Can we build one agent first?

Yes. We would usually prefer that.

Start with one defined job. Build it properly. Test it. Use it. See what happens. Then decide whether there is another part of the process worth changing.

You do not need an army of AI agents. You need something useful.

When automation is enough

What if the first answer is automation?

Then we use automation.

If the problem is: "Whenever this form is submitted, create this task." we do not need to involve AI just to make the project sound more advanced. If your existing CRM can already solve the problem cleanly, use the CRM. If an integration removes the manual work, use the integration.

AI earns its place where interpretation, context, preparation or decision making adds something useful. The technology follows the problem. If you are weighing the two, see AI agent vs automation.

How we work

What working with us looks like

A clear route from the problem to something useful. See more about our method.

01

Bring us the problem

Tell us what part of the sales process is frustrating, slow or unreliable.

02

We map the job

We understand what happens today, who is involved and where the friction sits.

03

We decide whether an agent makes sense

Sometimes it does. Sometimes something simpler is better.

04

We design the workflow

Triggers, information, decisions, actions, permissions, approvals and escalation.

05

We connect the systems

Using your existing software where practical.

06

We build and test

Including the situations that do not follow the perfect path.

07

We introduce it carefully

Starting with an appropriate level of authority.

08

We improve from evidence

Based on what happens when the workflow meets real work.

Discuss your AI sales agent

The first conversation

What should you bring to the first conversation?

You do not need a technical specification.

Bring us something like:

That is enough to start. We can work backwards from the problem.

Frequently asked questions

The questions UK businesses ask us first.

Can you build an AI sales agent for my business?
We design and build AI agents and workflows around defined sales tasks. The first step is understanding the job, systems involved and whether an agent is actually the appropriate solution.
How long does it take to build an AI sales agent?
That depends on the complexity of the workflow, integrations, information sources, actions, approval requirements and testing involved. We scope the work after understanding the process rather than giving every agent the same artificial build time.
How much does an AI sales agent cost?
The cost depends on the scope of the project and ongoing systems required. A small focused workflow and a multi-system agent with several actions, approvals and integrations are very different builds.
Can an AI agent work with our existing CRM?
Potentially, yes. It depends on the CRM, available integrations, permissions and what the agent needs to do. We assess the existing system before recommending changes. See AI CRM automation.
Can an AI sales agent send emails?
Technically, an agent can be given permission to send email where the systems support it. Whether it should send automatically is a separate design decision. External communications can remain subject to approval or be limited to tightly defined situations.
Can we approve actions before the agent does them?
Yes. An agent can prepare actions for human approval rather than carrying them out independently. Different actions can also have different approval requirements.
Can the agent be made more autonomous later?
Yes. Specific authority can increase after testing and real use where there is evidence that doing so is appropriate. It can also be reduced if more oversight is needed. This is a core part of agent governance.
What happens if the AI does not know what to do?
A properly designed workflow should have escalation behaviour. Rather than forcing the agent to make a decision outside its boundaries, it can stop, gather the relevant context and involve a person.
Do we need several agents?
Probably not to start. One agent with one useful, clearly defined job is often a much better starting point than building several interconnected agents immediately.
What if we do not actually need AI?
Then we should not build AI into that part of the process. A simpler solution is still a successful outcome if it solves the problem properly.
Build the job. Not the demo.

Build the job. Not the demo.

An impressive demonstration is easy. A useful agent has to work when: The information is incomplete. The customer says something unexpected. The CRM is messy. A system does not respond. The normal rule does not apply. Somebody needs to approve an action. The AI should not continue.

That is the agent we are interested in building. One with a job. Boundaries. Access it actually needs. A route back to a person. And a reason to exist.