Superintelligent Agent for Sales: Prospecting, Proposals, Follow-Up

Superintelligent Agent for Sales: Prospecting, Proposals, Follow-Up

A superintelligent agent for sales, in practical terms, is an AI agent that researches, drafts, organizes and remembers more consistently than an overloaded person, but within limits the company defines. Today it can prepare prospecting, build proposal drafts and run follow-up, while the salesperson keeps deciding price, exceptions, the relationship, and when to push or stop. The better the model behind it, the more tasks the agent can take on, but the design of limits and review is what makes the result reliable.

Key takeaways

• A useful sales agent today is an assistant with tools, rules and a decision log, not an autonomous salesperson.
• Prospecting, proposal drafting and follow-up are the three areas where time savings are most direct, because the work is repetitive and based on information.
• Better models improve reasoning and writing, but they do not replace permissions, human approval and ongoing evaluation.
• The salesperson still decides price, terms, exceptions, tone in delicate moments, and whether to give up on an opportunity.
• The safest path is to map the process, limit data and actions, pilot one real flow, and only then decide whether to expand.

What does the agent do in prospecting?

Prospecting is largely research and triage. Before any message goes out, someone has to understand who the company is, what it does, whether it fits the ideal customer profile, and what would be a legitimate reason to talk. A well-designed agent does this preparation at volume and hands the salesperson a short brief with sources.

In practice, the tasks that suit the agent well are:

• Enriching the lead record with public information and what already exists in the CRM.
• Scoring the lead against criteria written by the sales team, not criteria invented by the model.
• Suggesting a reason to reach out based on verifiable facts, showing where each fact came from.
• Drafting the first message in a tone that fits the company, for review.
• Detecting duplicates and contacts already being handled by another salesperson.

The point to watch is the channel. Outbound contact on WhatsApp requires prior permission and approved templates, and mass cold outreach tends to cause blocks and reputation damage. Read the guide on WhatsApp outbound messages, templates and opt-in before designing any outreach sequence. Data protection law also applies (LGPD in Brazil): personal data collected during research needs a clear purpose and legal basis, and the company should be able to explain where the contact came from.

How does the agent help build the proposal?

The proposal is where an agent saves the most time and where a mistake costs the most. Assembling the document (gathering scope, terms, timelines, product references, the introduction text) is mechanical. The decision about what to promise, and on what terms, is commercial and should stay with a person.

A sensible design separates three layers:

• Fixed data: catalog, standard scope, reference timelines and approved clauses, kept in a source the agent reads from and does not rewrite.
• Customer data: needs gathered in the conversation or meeting, recorded in a structured way.
• Text and assembly: the part where the model actually writes, combining the two previous layers.

In this setup the agent cites the origin of each piece of information, and any item without a source is marked as pending. This reduces the risk of the model filling a gap with something plausible but false. The technique of retrieving information from a knowledge base before writing (retrieval-augmented generation) exists for exactly this reason, and the company should treat it as part of the design, not a technical detail.

Price, discounts and special deadlines should not leave the agent as a commitment. It can calculate a range from an approved table and flag that the case falls outside the standard, but approval belongs to the sales lead. The same goes for promises of results: the agent should not write guarantees the company would not sign.

After meetings, the agent can also turn notes into pipeline fields. There is a step-by-step in the article on how to update your CRM with AI after a sales meeting, and anyone who records or transcribes calls must inform participants and have a legal basis for doing so.

Follow-up: where the agent is strong and where it becomes a nuisance

Follow-up is where many opportunities are lost, not for lack of interest but for lack of routine. A salesperson with many accounts forgets to reconnect, reconnects late, or reconnects without remembering what was said. The agent does not forget, and it reads the whole history before writing.

What it does well:

• Remembering agreed dates and proposing the follow-up at the right moment.
• Writing a message that picks up the specific point from the previous conversation, instead of a generic 'just checking in'.
• Answering objective questions about the proposal using only approved material.
• Logging every contact and reply in the CRM, so the salesperson sees the full picture.
• Stopping when the customer asks not to be contacted again, and recording that so no other sequence reactivates them.

What makes it a nuisance is persistence without judgment. An agent configured only to 'increase replies' tends to write more and more messages, and that wears down the relationship. So the design needs an attempt limit, a pause rule and a clear condition for handing off to a person. In WhatsApp conversations, the service window rules also change what the agent can send and when. The article on the WhatsApp 24 hour customer service window explains how the agent should act in each situation.

How does the design benefit from better models?

The word superintelligence tends to create inflated expectations. What a company runs today are AI agents with specific capabilities that keep improving, and it helps to understand the difference between that and the idea of general intelligence. The piece on the difference between AGI, ASI and AI agents helps separate what is a future debate from what is a working tool.

That said, better models change the design in concrete ways:

• Long context reading: the agent can consider more conversation history, more documents and more rules in the same task, with less loss of detail.
• Instruction following: rules such as 'never promise a deadline' or 'hand off if the customer mentions cancellation' are followed more consistently.
• Tool use: the agent chooses better when to query the CRM, the price table or the calendar, and fills in parameters more accurately.
• Writing: the text comes out more natural and better adapted to the customer profile, which reduces review of form, though not of content.
• Multi-step tasks: longer plans, such as researching, comparing, drafting and checking, fail less often midway.

This allows some parts of the system to be simplified. Flows that once needed many rigid rules can be replaced by a well-written instruction and a set of tools. On the other hand, what does not improve on its own is accountability. A more capable model can be wrong with more conviction, and the risk of prompt injection (someone hiding a command in an email or web page the agent reads) still exists. The list of risks for applications built on language models maintained by OWASP is a good reference for reviewing the design before giving the agent tools.

The guidance Anthropic has published on building agents recommends starting with the simplest design that solves the problem and adding complexity only when measurement justifies it. In sales, that means preferring a predictable flow with one step of model judgment over an agent with broad autonomy. When the model is swapped for a better one, the existing evaluation should show whether something improved or got worse, rather than the impression of someone who tested three conversations.

Where does the salesperson still decide?

There is a practical boundary between what you delegate and what you keep. A useful rule: the agent prepares and proposes, the person decides and takes responsibility. The decisions that stay with the salesperson or sales manager include:

• Price, discounts, payment terms and any exception to the standard.
• Prioritizing strategic accounts and deciding to invest time in an uncertain opportunity.
• Conversations involving emotional objections, distrust or a history of conflict.
• Contractual commitments, deadlines that depend on other teams, and promises outside the catalog.
• Deciding to walk away, or closing a negotiation without damaging the relationship.
• Any situation where the customer asks to speak to a person.

There is also a legal component. LGPD provides a right to review for decisions made solely through automated processing that affect the data subject, and this is worth considering when the agent scores or discards leads. The article on automated decisions and Article 20 of LGPD covers this in more detail. Keeping a person in the relevant decision and recording the criteria used is a simple safeguard.

Illustrative scenario: a B2B services company receives an after-hours message from a prospect asking for a quote. The agent replies within the channel rules, asks the qualification questions the company defined, logs the answers in the CRM and drafts a proposal based on the standard scope. On noticing that the request includes a term outside the price table, it does not respond on that point and alerts the salesperson, attaching the history and the draft. The next morning, the salesperson reviews, adjusts the term, approves and sends. The agent then schedules the follow-up and logs each step. This example is illustrative and does not represent any client's result.

How do you start without improvising?

The method we apply comes down to four verbs: map, limit, pilot and hand over. In sales, it looks like this:

• Map: describe the current process from first contact to close, note where time is lost, and set a baseline, such as response time and the number of leads with no follow-up.
• Limit: decide which data the agent can see, which tools it can use, which messages it can send on its own and which need approval. Give each tool only the permission it needs.
• Pilot: choose one small, real flow, for example post-proposal follow-up, with a decision log and a success metric agreed before starting.
• Hand over: at the end of the pilot, deliver the documentation, logs and evaluation criteria, and decide with the team whether to expand, adjust or stop.

For evaluation, build a set of anonymized real conversations and hard cases (an irritated customer, a request outside scope, an attempt to manipulate the agent) and run it with every model or instruction change. Anthropic's material on agent evaluations and NIST's risk management practices help structure this without reinventing the process. Writing the agent's instructions also deserves care, and the article on a sales agent prompt for WhatsApp AI shows how to organize role, limits and the moment of handoff.

A checklist of prerequisites before the pilot:

• A sales lead exists who approves the agent's rules.
• The pricing, scope and clause base is up to date and has an owner.
• Opt-in and the legal basis for contacts are documented.
• There is a clear path for the customer to reach a person.
• Every agent reply is logged, with the source used.
• There is a criterion for stopping the pilot if something goes off course.

Frequently asked questions

Can a sales agent close deals on its own? In general, it should not. It can lead the conversation, prepare the proposal and handle follow-up, but closing involves price, terms and trust, which depend on a responsible person. For simple, standardized sales, the company can allow the agent to confirm orders within closed rules, with logging and the possibility of review.

Do better models remove the need for human review of proposals? No. Better models make fewer mistakes in writing and interpretation, but they can still produce incorrect information or follow an instruction hidden in external text. Review can become lighter and more focused on price and commitments, but it remains necessary as long as the agent can send anything to the customer.

What does LGPD change for a prospecting agent? It changes how contact data is collected and used. The company needs a clear purpose, an adequate legal basis, transparency about use, and a channel for data subjects to exercise their rights. For details and official guidance, the ANPD website is the starting point, and the specific case should be assessed with legal support.

Where should a small sales team start? With the point that loses the most opportunities, which is usually follow-up after the proposal or replies outside business hours. Choose one flow, define what the agent can and cannot do, log everything, and evaluate against a metric agreed in advance. If the pilot does not show value, stopping is also a legitimate result.

Sources

• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents
• OWASP Top 10 for LLM Applications
• NIST AI Risk Management Framework