A sales AI agent that asks, filters, and routes opportunities should not be designed as a chat window with a long prompt. It should be designed as a component of the sales process: it understands what information is missing, applies qualification criteria, prepares a useful summary, and knows when to hand the conversation off to a human.
The goal is not for the agent to talk more. The goal is to create less sales noise: fewer repeated questions, fewer leads without context, fewer incomplete records, and fewer lost opportunities due to poor handoff.
In summary
A well-designed sales AI agent starts with the business flow, not the model. First, you define what opportunity it should qualify, what data it needs, what rules it must apply, what tools it can use, what decisions it can’t make, and what it should deliver to the human team.
The minimal architecture combines inputs, rules, knowledge base, tools, validations, structured output, human handoff, and measurement. OpenAI, Anthropic, Meta Llama, and n8n document patterns for tools and structured calls; ReAct provides the principle of alternating reasoning and action; HubSpot shows why context must end up in traceable CRM objects and activities.
The problem the architecture must solve
Before choosing a model, tool, or automation, you need to describe the operational problem.
In many sales processes, the bottleneck is not a lack of leads. It’s the lack of useful context to decide what to do with them.
- Forms, chats, or emails arrive with ambiguous messages.
- The sales team always asks the same questions.
- There are no clear criteria to separate opportunities, curiosity, support, or disqualification.
- The CRM receives incomplete data or notes that are hard to reuse.
- A person has to copy, interpret, summarize, and route manually.
- Good opportunities wait too long before getting a next step.
Designing the agent means turning that manual sequence into a controlled system. For a deeper dive into the full framework, the article AI-powered sales automation: guide for companies and agencies explains how this type of agent fits into the sales process.
Core principle: it doesn’t start with the prompt
A sales AI agent doesn’t start with an instruction like “you are a sales expert.” It starts with a more uncomfortable question:
What sales decision should be better prepared after this interaction?
From there, you define the architecture.
A sales AI agent that asks, filters, and routes opportunities is defined as a system that gathers context, applies business rules, uses tools when appropriate, produces a structured output, and delivers the case to a human with enough information to decide the next step.
This definition separates the agent from a generic chatbot. A chatbot can respond. A sales agent must prepare for action: qualify, prioritize, log, route, and measure.
System components
The architecture should be simple to understand, but explicit. If a component isn’t defined, the agent improvises.
| Component | Design question | Practical decision |
|---|---|---|
| Agent objective | What should it resolve or prepare? | Qualify, summarize, route, disqualify, or request context. |
| Information inputs | Where does the opportunity appear? | Form, chat, email, landing page, CRM, or transcribed call. |
| Knowledge base | What must it know to ask the right questions? | Offer, industries, use cases, boundaries, indicative pricing, fit criteria. |
| Business rules | What criteria separate priority, review, or disqualification? | Fit, urgency, budget, size, need, channel, country, or customer type. |
| Conversational flow | What does it ask first and what should it avoid asking? | Progressive questions, not long questionnaires up front. |
| Tools | What can it consult or trigger? | CRM, calendar, email, n8n, databases, forms, or APIs. |
| Integrations | Where should the result be logged? | Contact, company, lead, deal, activity, note, task, or event. |
| Validations | What should it never execute without control? | Sensitive data, CRM changes, external sends, permanent disqualifications. |
| Human handoff | When should it handoff to a person? | When there’s fit, ambiguity, risk, urgency, or a sales decision needed. |
| Metrics | How will you know if it works? | Qualified leads, disqualifications, meetings, response time, quality of the brief. |
OpenAI and Anthropic treat tools as capabilities with schema, name, description, and parameters. In sales practice, this means you must write tools as contracts: what they do, when to use them, what data they need, and what result they return.
Step-by-step flow
The flow shouldn’t try to close the entire sale. It should better prepare the first stage.
- Lead input: the user arrives via form, chat, email, or an integration.
- Intent identification: the agent separates sales inquiry, support, partnership, job application, spam, or ambiguous case.
- Minimal questions: it only asks for the context needed to decide the next step.
- Rule application: it compares the information to qualification criteria.
- Classification: it assigns status, priority, reason, and opportunity type.
- Structured summary: it generates a brief with problem, context, signals, and gaps.
- Logging: it creates or updates objects, properties, notes, activities, or tasks.
- handoff: it delivers the case to a person, team, or follow-up flow.
- Measurement: it saves events and results to adjust rules.
The ReAct pattern is useful as a conceptual reference because it combines reasoning and action: the system doesn’t just produce text, it can also decide when it needs information, when to use a tool, and how to update the plan based on the observation. This doesn’t mean exposing internal reasoning to the user, but designing a flow that enables traceable action.
Key design decisions to lock in
Real design happens in small decisions. These are the ones that most affect agent quality.
| Decision | Best practice | Risk if undefined |
|---|---|---|
| What to ask | Ask only what’s needed to determine fit and next step. | The agent turns into a long interrogation. |
| What not to ask | Avoid sensitive, irrelevant, or premature data. | Friction, drop-off, or poor experience. |
| When to insist | Only insist if a necessary data point is missing for classification. | Repetitive or intrusive conversations. |
| When to handoff | handoff with a summary when there’s fit, urgency, or relevant ambiguity. | Good leads waiting or over-delegated decisions. |
| When to stop | Close or pause if there’s no fit, consent, or enough context. | Uncontrolled automation. |
| How to summarize | Use structured fields: need, signals, doubts, priority, and next step. | Long, vague, or hard-to-action notes. |
| How to log | Map the result to CRM, task, email, or internal tool. | Loss of traceability. |
This logic connects with lead qualification: the article How to automate lead qualification with AI explains how to define fit, intent, urgency, and priority without relying on manual review.
How to use tools without losing control
Tools expand what the agent can do, but also increase risk. That’s why you shouldn’t give it generic access to “the whole system.”
OpenAI documents function calls with JSON arguments and recommends strict mode to adhere to the schema. Anthropic emphasizes the importance of clear descriptions: a tool must explain what it does, when to use it, how it behaves, and what data it expects. Meta Llama shows tool calling formats where the model produces a structured call and another external system executes the action. n8n allows you to operate this logic with nodes, tools, parsers, intermediate steps, traceability metadata, and human review for sensitive actions.
In a sales agent, tools should be designed like this:
classify_opportunity: receives context and returns status, priority, and reason.create_sales_summary: returns a brief with predictable fields.find_crm_contact: checks if the contact or company already exists.log_lead: creates or updates the corresponding record.create_human_task: assigns follow-up with summary and next step.request_review: pauses sensitive actions until human approval.
The rule is simple: the agent can prepare, consult, and propose. Actions that modify critical records, send external messages, or disqualify important opportunities should have validations or human approval.
Where the context should end up
An agent that classifies well but doesn’t leave a useful record still creates manual work.
HubSpot structures processes around objects, properties, records, associations, and activities. For a sales AI agent, this means the result shouldn’t just stay in a conversation. It should be logged as operational information.
| Agent result | Where it should go | What it’s for |
|---|---|---|
| Contact data | Contact | Identify the person and avoid duplicates. |
| Company or domain | Company | Link the opportunity to an account. |
| Need and context | Note or custom property | Prepare the call or proposal. |
| Qualification status | Lead or deal | Know if it’s open, qualified, or closed. |
| Next step | Task, activity, or meeting | Prevent lost manual follow-up. |
| Disqualification reason | Property or note | Improve rules and reporting. |
| Summarized conversation | Associated activity | Give context to the human team. |
If the CRM doesn’t reflect the agent’s result, the team is still operating blind.
Common technical mistakes
The most common failures aren’t model-related. They’re architectural.
- Prompt too open: the agent responds well once, but doesn’t maintain consistent criteria.
- No explicit rules: it doesn’t know when to qualify, pause, handoff, or disqualify.
- No scoped tools: it has instructions, but can’t log, consult, or trigger actions.
- No structured output: it generates text that’s useful to read but hard to process.
- No human handoff: the team receives full conversations instead of actionable summaries.
- No logs or metrics: you can’t tell which questions work, which leads are lost, or which rules fail.
- No boundaries: it tries to close sales decisions that should stay in human hands.
The article Chatbot vs sales AI agent: real differences explains this distinction: the value isn’t in conversation, but in connecting conversation, rules, tools, and action.
Minimum viable version
You don’t need to design the perfect agent on day one. But you do need to design a first measurable version.
If you need to break down that first version by phases, the Implementation playbook: sales AI agent in 30 days explains how to go from diagnosis to validation without trying to automate everything at once.
| Layer | Recommended MVP | Leave for later |
|---|---|---|
| Input | One main form or chat. | All channels at once. |
| Questions | 5 to 8 critical questions based on intent. | Long conversations with multiple branches. |
| Rules | Simple fit, priority, and disqualification criteria. | Complex predictive scoring. |
| Tools | CRM logging, internal task, and summary. | Advanced multichannel automations. |
| handoff | handoff to a person with a brief. | Complex automatic assignment by territory or portfolio. |
| Measurement | Qualified leads, disqualifications, meetings, and response time. | Full attribution and forecasting models. |
The first version should prove one thing: that the team receives better opportunities with less manual work.
Technical-sales checklist
Before launching the agent, I would review these questions:
- Is the agent’s objective written in a concrete sentence?
- Is it clear what minimum information is needed to qualify?
- Are there criteria for priority, disqualification, and human review?
- Do the tools have clear names, descriptions, and schemas?
- Is the agent’s output structured and easy to log?
- Does the human team receive a summary, signals, doubts, and next step?
- Are there logs or traceability of decisions and actions?
- Do sensitive actions require validation or approval?
- Does the CRM receive useful data, not just a transcript?
- Are qualified leads, meetings, disqualifications, and response time measured?
If several answers are “no,” you don’t have an architecture yet. You have a demo.
How Nicolás Torres would approach it
The right approach wouldn’t be to start with “which model do we use.” I would start by mapping the current sales process:
- Where the opportunities come from.
- What information is almost always missing.
- What questions the team repeats.
- What criteria separate a useful opportunity from a non-fit inquiry.
- What tool should receive the summary.
- What decisions must remain in human hands.
- What metrics will prove the agent improves the flow.
Then, yes, you design the agent: instructions, knowledge base, tools, rules, outputs, handoff, and measurement. That order keeps the right positioning: AI-powered sales automation architecture, not chatbot installation.
Frequently asked questions
What does it mean for an AI agent to ask, filter, and route opportunities?
It means the agent gathers context, applies business rules, classifies the opportunity, and delivers an actionable summary to the human team when appropriate.
Where do you start when designing a sales AI agent?
You start with the sales process: objective, inputs, qualification criteria, required data, boundaries, tools, and human intervention.
What’s the difference between a prompt and an agent architecture?
A prompt only guides a response. An architecture defines rules, context, tools, validations, structured outputs, records, metrics, and handoff.
When should the agent handoff to a human?
It should handoff when the opportunity is qualified, there is relevant ambiguity, commercial judgment is missing, there is risk, or a human decision is required.
What should you build first?
A minimal version with one input, key questions, simple rules, structured summary, CRM or internal tool logging, and basic measurement.
Let’s design your sales AI agent
If your company or agency receives inquiries that require repeated questions, manual classification, or unclear handoff, the first step is not choosing a tool. It’s designing the flow architecture: what it should ask, how it should filter, when it should handoff, and where it should be logged.
Analyze my sales AI agent architecture
Frequently Asked Questions
- What does it mean for an AI agent to ask, filter, and route opportunities?
- It means the agent gathers context, applies business rules, classifies the opportunity, and delivers an actionable summary to the human team when appropriate.
- Where do you start when designing a sales AI agent?
- You start with the sales process: objective, inputs, qualification criteria, required data, boundaries, tools, and human intervention.
- What's the difference between a prompt and an agent architecture?
- A prompt only guides a response. An architecture defines rules, context, tools, validations, structured outputs, records, metrics, and handoff.
- When should the agent handoff to a human?
- It should handoff when the opportunity is qualified, there is relevant ambiguity, commercial judgment is missing, there is risk, or a human decision is required.
- What should you build first?
- A minimal version with one input, key questions, simple rules, structured summary, CRM or internal tool logging, and basic measurement.